Preprint
Article

This version is not peer-reviewed.

An Artificial Neural Network Surrogate Model for Predicting Equilibrium Gas Yields During Supercritical Water Gasification of Açaí Residual Biomass

A peer-reviewed version of this preprint was published in:
Processes 2026, 14(20), 3380. https://doi.org/10.3390/pr14203380

Submitted:

26 September 2026

Posted:

29 September 2026

You are already at the latest version

Abstract
Thermodynamic equilibrium modeling is a useful approach for predicting gas production during supercritical water gasification (SCWG) of wet biomass; however, Gibbs energy minimization can be computationally demanding for repeated process simulations. This study developed an artificial neural network (ANN) surrogate model to predict the equilibrium yields of H₂, CO, CO₂, and CH₄ during SCWG of açaí residual biomass from Maranhão, Brazil. Four biomass samples collected under different geographical and seasonal conditions were characterized by ultimate analysis and used to generate a thermodynamic database with 1,000 simulations per sample using the TeS-Thermo PyPI Python library. The ANN comprised 10 input variables, two hidden layers with 12 and 6 neurons, and 4 output neurons, and was trained using the Levenberg–Marquardt algo-rithm. The selected ANN achieved R² values of 0.99989, 0.99553, 0.99957, and 0.99963 for H₂, CO, CO₂, and CH₄, respectively. External validation using an independent biomass composition provided evidence of the ANN's generalization capability within the in-vestigated compositional domain, with R² values above 0.98 for all four species. Sensi-tivity analyses further demonstrated that the ANN reproduced the effects of temperature (700–1100 K), water-to-biomass molar ratio (3–6 mol mol⁻¹), and pressure (230–260 bar), with low deviations from TeS-Thermo predictions. These results demonstrate that the ANN provides an accurate and computationally efficient surrogate model for rapid thermo-dynamic screening, sensitivity analysis and process optimization of SCWG biomass processes.
Keywords: 
;  ;  ;  ;  ;  ;  ;  

1. Introduction

The increasing demand for renewable and low-carbon energy carriers has intensified the development of thermochemical routes for hydrogen (H₂) production from renewable sources, such as biomass. Among these technologies, supercritical water gasification (SCWG) is particularly attractive for wet and water-rich feedstocks because it enables the conversion of organic matter in a high-pressure aqueous environment, reducing or eliminating the need for energy-intensive feedstock drying. Above the critical point of water (around 647 K and 220 bar), changes in water properties promote the conversion of biomass-derived compounds through reactions such as reforming and water-gas shift (WGS), resulting in a gas mixture mainly composed of H₂, carbon dioxide (CO₂), carbon monoxide (CO), and methane (CH₄) [1,2]. The distribution of these products is strongly influenced by operating conditions, such as temperature, pressure, and water content, as well as by the composition of the biomass used as feedstock. Due to the multicomponent and the possibility of multiphase conditions in such systems, their description using kinetic or thermodynamic models can become complex and computationally demanding. In this context, the development of reliable and robust predictive models is essential for process monitoring, analysis, and optimization [3].
Among biomass resources available for thermochemical conversion, açaí seed (Euterpe oleracea Mart.) stands out as a high-potential regional feedstock due to the large quantity of residues generated during pulp processing. In 2024, Brazilian açaí production reached approximately 1.99 million tones, with about 1.78 million tons produced in the state of Pará, corresponding to 89.5% of national production [4]. Considering that seeds account for approximately 80.0% of the fruit mass, this production represents a potential generation of approximately 1.59 million tons of açaí residues in Brazil. Previous studies have demonstrated the potential of açaí seed for thermochemical conversion through gasification [5], hydrothermal processing [6], torrefaction and pyrolysis [7]. These findings highlight the energy potential of this residue and the importance of accounting for its physicochemical characteristics and variability when modeling its thermochemical conversion.
Biomass compositional variability poses a challenge for predictive modeling, as moisture, elemental composition, volatile matter, ash, and fixed carbon can vary with feedstock characteristics and processing conditions, affecting equilibrium gas yields. Gibbs energy minimization (MinG) models rigorously account for these effects while satisfying thermodynamic constraints, but their repeated solution can become computationally demanding when evaluating multiple scenarios [7,8,9,10].
This computational limitation is particularly relevant in engineering applications involving process optimization, sensitivity and uncertainty analyses, feedstock selection, scale-up, techno-economic assessment, and process integration. Such applications may require thousands or millions of simulations, making repeated thermodynamic equilibrium calculations computationally demanding across the multiple required scenarios [10]. Consequently, machine learning (ML) techniques and artificial neural networks (ANNs) have recently been investigated as alternatives for predicting H2 production and syngas composition during SCWG [11,12,13,14]. These studies demonstrate that ML models can capture nonlinear relationships between biomass properties and operating conditions while requiring substantially less computational effort than repeated thermodynamic simulations. Surrogate models have also been proposed for the techno-economic analysis of SCWG, enabling the rapid evaluation of different feedstocks, plant capacities, and process characterization [14].
The use of computationally efficient surrogate models aligns with the transition toward advanced digital process architectures, including soft sensors, model predictive control, optimization frameworks, and digital twins. Recent studies in biomass gasification have integrated process models, data-driven models, real-time measurements, and optimization tools to support process monitoring and decision-making [15,16]. In such architectures, surrogate models can replace computationally demanding thermodynamic calculations while retaining the response of the underlying model, enabling rapid evaluation of operating scenarios. This capability is particularly relevant to SCWG because the simultaneous variation of feedstock composition, temperature, pressure, and water content creates a multidimensional operating space that is computationally demanding to explore using conventional equilibrium calculations.
Despite these advances, an important gap remains between rigorous thermodynamic modeling and the computational requirements of large-scale and digitally enabled SCWG applications. Most existing ML approaches have been developed primarily from experimental databases or have focused on specific outputs, such as H₂ yield, gasification efficiency, or economic indicators [14,15,16]. Comparatively limited attention has been given to multi-output surrogate models trained directly on rigorous thermodynamic calculations, such as MinG coupled with equations of state, capable of simultaneously reproducing the equilibrium yields of H₂, CO, CO₂, and CH₄ while accounting for biomass composition and operating conditions. Moreover, the generalization of such models to independent biomass compositions not used during model development remains insufficiently investigated. Addressing these gaps is essential for developing accurate predictive tools capable of rapidly evaluating diverse feedstocks and operating scenarios.
Therefore, this study aimed the validation of a thermodynamically grounded artificial neural network (ANN)-based surrogate model for predicting H₂, CO, CO₂, and CH₄ yields during SCWG of açaí seeds residual biomass. Specifically, the objectives are to:
  • develop a comprehensive thermodynamic database using the TeS-Thermo Python Package Index (PyPi) library and the TeS-MinG Gibbs free energy minimization framework, incorporating experimentally characterized variations in açaí seed composition, temperature, pressure, and feed-water content;
  • develop and evaluate an ANN capable of simultaneously predicting the equilibrium yields of H₂, CO, CO₂, and CH₄;
  • assess the ability of the ANN to reproduce the nonlinear response profiles of the underlying TeS-MinG model as a function of the main operating variables;
  • evaluate the generalization capability of the ANN using an independent açaí seed composition reserved exclusively for external validation and not used during model development;
  • quantify the computational efficiency of the ANN relative to the original TeS-MinG calculations; and
  • assess its potential as a computationally efficient surrogate model for large-scale scenario exploration, process optimization, techno-economic assessment, and digital-process applications.

2. Materials and Methods

2.1. Biomass Preparation and Characterization

Açaí fruits (Euterpe oleracea Mart.) were collected from local processing facilities in the state of Maranhão (MA), Brazil, within the eastern Amazon region. After pulp extraction, the residual seeds were collected directly from the processing facilities and subjected to size reduction using a hammer mill (TE-330, Tecnal, Brazil) for 2 minutes. The resulting material was subsequently processed in a ball mill for 10 minutes to improve sample homogenization and obtain a finer and more uniform particle size suitable for subsequent physicochemical and thermal characterization. The milled material was then subjected to particle-size classification using an electromagnetic sieve shaker (Solotest, Brazil) for 15 min. The fraction retained between 20 and 35 mesh was selected for subsequent analyses to ensure greater granulometric uniformity and improve analytical reproducibility. The selected biomass fraction was transferred to sealed polyethylene bags, properly labeled, and stored until analysis.

2.1.1. Proximate Analysis

The proximate composition of the açaí seed biomass was determined in terms of moisture content, ash content, volatile matter, and fixed carbon, following the applicable ASTM standard procedures: ASTM D3173-11 [17], D3174-12 [18], D3175-20 [19], and D3172-13 [20], respectively. The analyses were performed with four açaí seed biomass samples collected from different regions of Maranhão state and corresponding to distinct seasonal conditions, to evaluate potential variations in proximate composition associated with geographical and seasonal variability. All measurements were performed in triplicate using previously dried and weighed porcelain crucibles. Prior to each analysis, the crucibles were cooled in a desiccator to minimize moisture uptake and ensure measurement accuracy.

2.1.2. Ultimate Analysis

The elemental composition of the açaí seed biomass was determined using a CHNS elemental analyzer (Vario Macro Cube). Carbon (C), hydrogen (H), nitrogen (N), and sulfur (S) contents were determined by complete combustion of the samples, followed by quantitative elemental analysis. The oxygen (O) content was estimated by difference according to Equation (1).
O = 100 − C + H + N + S + A s h
where C, H, N, S, and Ash are expressed in wt.%. The analyses were conducted for the same four biomass samples considered in the proximate analysis, allowing the elemental composition to be compared across the different geographical and seasonal conditions.

2.2. Thermodynamic Model—TeS-Thermo Python library

The experimental characterization data of the four açaí seed biomass samples were used as inputs for thermodynamic equilibrium simulations using the TeS-Thermo PyPI Python library, following the methodology of dos Santos-Júnior et al. [21]. The elemental composition of each sample was incorporated into thermodynamic MinG model as a pseudocomponent to represent biomass composition. Simulations were performed at 700–1100 K, 230–260 bar, and water-to-biomass molar ratios of 3.0–6.0, generating equilibrium yields of H₂, CO, CO₂, and CH₄. The resulting dataset served as the reference database for developing and validating the ANN surrogate model and assessing the effects of the operating variables. The integrated workflow used to construct the data, and the ANN surrogate model is illustrated in Figure 1.

2.2.1. Gibbs Energy Minimization Model (MinG)

The coupled chemical and phase equilibrium problem can be formulated as the minimization of the total Gibbs energy of the system at constant temperature (T) and pressure (P). For a system comprising multiple phases and chemical components, this formulation is expressed by Equation (2) [22].
m i n G = ∑ i = 1 N C ∑ k = 1 N F n i k μ i k
where G is the total Gibbs energy of the system, n i k is the number of moles of chemical species i in phase k , and μ i k is the chemical potential of species i in phase k . N C denotes the total number of chemical species considered, whereas N F represents the number of phases.
At the minimum Gibbs free energy condition, the system must also satisfy two fundamental constraints: the non-negativity of the number of moles (Equation (3)) and the conservation of atoms (Equation (4)). These constraints ensure that the resulting equilibrium state is both physically feasible and thermodynamically consistent with the process conditions.
n i k ≥ 0 ;       i = 1 , … ,   N C ; k = 1 , … ,   N F
∑ i = 1 N C a m i ∑ i = 1 N F n i k = ∑ i = 1 N C a m i n i 0 ;       m = 1 , … , N E
Where the variable a m i is the number of atoms of element m present in chemical species i, n i 0 is the initial number of moles of species i, and NE is the total number of chemical elements considered in the system.
Initially, the model considers solid carbon (C(s)) as the only solid-phase component, with this phase assumed to behave as an ideal solid. Under this assumption, Equation (2) can be simplified and expressed in the form presented in Equation (5).
m i n G = ∑ i = 1 N C n i g μ i g + R T l n P + l n y i + l n ϕ ^ i + n C ( s ) μ C ( s )
where n i g is the number of moles of gaseous species i, μ i g is the standard chemical potential of species i in the gas phase, R is the universal gas constant, yi is the mole fraction of gaseous species i, and ϕ ^ i is the fugacity coefficient of species i. The term n C ( s ) represents the number of moles of solid carbon, while μ C ( s ) is its chemical potential.
An important assumption adopted in this study is that the non-ideal behavior of all systems was described using the φ–φ approach, with fugacity coefficients calculated from the Peng–Robinson equation of state (PR EoS) [23]. The chemical potential of each species was determined from standard-state formation properties, with its temperature dependence evaluated from the corresponding enthalpy and heat-capacity relationships [21]. The heat capacities were calculated using the polynomial correlations reported by Smith et al. [24].
The TeS-Thermo PyPi library supports different models for describing thermodynamic non-idealities, including the Peng–Robinson (PR) and Soave–Redlich–Kwong (SRK) [25] equations of state (EoS) and the second-order truncated Virial equation [26]. The Peng–Robinson (PR) equation of state was selected in this study because it provides a suitable balance between computational efficiency and thermodynamic accuracy. Moreover, the PR EoS has been widely employed for the thermodynamic modeling of biomass supercritical water gasification (SCWG) systems. The combination of the thermodynamic model, equation of state, and pseudocomponent representation adopted in this work is consistent with approaches reported in previous studies, which have demonstrated satisfactory performance in representing a range of thermochemical processes [27,28,29,30]. Details about the thermodynamic formulation and Strategies for Solving Nonlinear Gibbs energy minimization Problems can be found in TeS-Thermo PyPi library can be found in dos Santos Júnior et al., [21] and in Santos Júnior et al. [31].
A total of 12 potential gas-phase components were considered in the SCWG equilibrium modeling, representing the main chemical species expected to form during biomass SCWG, based on consolidated evidence from the literature [8,9]. The considered species were H₂O, H₂, CH₄, CO₂, CO, O₂, N₂, CH₃OH, C₂H₆, C₃H₈, NH₃, and C₂H₄, while pure solid carbon (C(s)) was considered as the unique solid-phase component as presented in Equation (5).
A thermodynamic database comprising 3,000 simulations was generated from three independently characterized biomass compositions (S1, S2 and S3), with 1,000 simulations assigned to each biomass. For each composition, temperature (700–1100 K), pressure (230–260 bar), and initial H₂O/biomass ratio (3–6) were systematically varied using ten equally spaced levels for each variable, resulting in a regular Cartesian grid of 10 x 10 x 10 unique combinations. This strategy provided uniform coverage of the investigated operating domain and generated the dataset used for ANN development.
The thermodynamic framework implemented in TeS-Thermo has been validated against experimental and literature-simulated SCWG data for methanol and ethanol, achieving RMSE values below 0.23 [21]. Gibbs energy minimization models have also been widely applied to biomass gasification, generally showing satisfactory agreement with experimental data for major gaseous products and overall process behavior [27,28,29,30,31]. Although deviations may occur due to kinetic and non-equilibrium effects, these results support the reliability of the MinG framework as the thermodynamic basis for generating the simulation database used in this study.

2.3. Surrogate ANN Model

The ANN models were developed using MATLAB R2025b® and the Neural Network Toolbox. The dataset generated from the thermodynamic equilibrium simulations was composed of the operating conditions and biomass elemental composition as input variables, while the equilibrium yields of H₂, CO, CO₂, and CH₄ were considered as output variables. The input dataset comprised 10 variables, including temperature, pressure, water-to-biomass molar ratio, and the elemental composition of the biomass.
The dataset was divided into training (70%), validation (15%), and testing (15%) subsets. The training subset was used to adjust the network parameters, the validation subset was employed to monitor the training process and reduce the risk of overfitting, and the testing subset was reserved for the independent assessment of the model predictive performance, following a strategy recently applied in ANN modeling of biomass gasification [32].
To identify a suitable ANN configuration for the surrogate model, a systematic model-selection procedure was performed, since the network architecture and training algorithm can substantially affect the representation of nonlinear relationships, predictive accuracy, convergence behavior, and computational cost. Therefore, evaluating multiple ANN configurations was necessary to determine whether increasing network complexity would provide meaningful improvements in prediction accuracy or, alternatively, lead to unnecessary computational demand. The overall ANN architecture-selection strategy is illustrated in Figure 2.
Initially, networks with one hidden layer were tested by varying the number of neurons from 5 to 20, considering the configurations of 5, 8, 10, 12, 15 and 20. Subsequently, networks with two hidden layers were evaluated using the configurations 8–4, 12–6, 16–8, 20–10, and 24–12 neurons, where each pair represents the number of neurons in the first and second hidden layers, respectively. These configurations were selected to investigate the influence of network complexity on predictive performance and computational cost of the surrogate model.
For each architecture, four training algorithms available in the MATLAB Neural Network Toolbox were evaluated: Levenberg–Marquardt (trainlm), Bayesian regularization (trainbr), scaled conjugate gradient (trainscg), and BFGS quasi-Newton (trainbfg). Thus, the ANN architecture and training algorithm were treated as independent model-selection parameters, allowing the predictive accuracy and computational efficiency of different configurations to be compared.
The hidden layers employed hyperbolic tangent sigmoid activation functions, whereas a linear activation function was adopted in the output layer [33]. Model performance was evaluated separately for H₂, CO, CO₂, and CH₄ using the coefficient of determination (R2), root mean square error (RMSE) and mean absolute error (MAE). The coefficient of determination was calculated according to Equation (6).
R 2 = 1 − ∑ i = 1 N y i − y ^ i 2 ∑ i = 1 N y i − y ¯ i 2
The RMSE, which quantifies the magnitude of the prediction errors while giving greater weight to larger deviations, was calculated using Equation (7).
R M S E = 1 N ∑ i = 1 N y i − y ^ i 2
The MAE, which represents the average magnitude of the absolute prediction errors, was calculated by Equation (8).
M A E = 1 N ∑ i = 1 N y i − y ^ i
where y i is the reference value obtained from the TeS-MinG model, y ^ is the corresponding value predicted by the ANN, y ¯ i is the mean of the reference values, and N is the number of observations.
In addition to predictive accuracy, training and prediction times were considered in the selection of the final surrogate model, seeking a suitable balance between accuracy and computational efficiency. The final ANN architecture was then selected based on its overall predictive performance, computational characteristics and was subsequently used to reproduce the equilibrium trends obtained from the TeS-MinG simulations under the investigated SCWG operating conditions.

2.4. External Validation Against TeS-MinG Simulated Data Using an Independent Biomass Dataset

To assess the generalization capability of the developed ANN surrogate model, external validation was performed using an independent açaí seed biomass Sample (S4), which was completely excluded from ANN development, training, architecture selection, and optimization. The other three biomass Samples (S1–S3) were used to construct the thermodynamic equilibrium database and develop the surrogate model. This strategy provides a more stringent assessment of generalization to an unseen biomass composition and is consistent with external-validation approaches reported for machine-learning models of biomass conversion [34].
For S4, its elemental composition was incorporated into the TeS-MinG model to calculate equilibrium yields of H₂, CO, CO₂, and CH₄ under the investigated operating conditions. The same conditions and biomass composition were then provided to the trained ANN, and its predictions were compared with the corresponding TeS-MinG results using R², RMSE, and MAE. The analysis also evaluated whether the ANN reproduced the thermodynamic trends across the operating conditions.
Successful prediction of S4 provides evidence that the ANN learned the relationships between biomass composition, operating conditions, and equilibrium gas yields rather than simply reproducing the cases contained in the training database, supporting its use as a computationally efficient surrogate for different açaí seed biomass compositions.

3. Results and Discussion

3.1. Experimental Characterization of Açaí Seed Biomass

Table 1 presents the proximate and elemental compositions, atomic ratios, and higher heating values (HHVs) of the three biomass samples used for ANN training (Samples 1, 2 and 3) and Sample 4 used for external validation. The samples exhibited similar proximate compositions, characterized by high volatile matter and low ash contents, whereas greater variability was observed in elemental composition, particularly nitrogen.
Sulfur contents were consistently low (0.073–0.090 wt.%), which is favorable for thermochemical conversion because sulfur-containing species, particularly H₂S, may contribute to catalyst poisoning, corrosion, and additional gas-cleaning requirements [35,36]. Accordingly, sulfur-containing species were excluded from the equilibrium calculations, considering the focus on the major gas products (H₂, CO, CO₂, and CH₄). This simplification reduced the chemical and computational complexity while retaining the main C–H–N–O chemistry. The HHVs ranged from 12.068 to 12.673 MJ/kg for the training samples, while Sample 4 presented 12.168 MJ/kg, further indicating its compositional consistency with the training domain. HHVs were calculated using the Channiwala and Parikh correlation [37].
The data in Table 1 indicate that Sample 4 presents small but meaningful compositional differences relative to the samples used for ANN training, particularly in nitrogen content. At the same time, its overall proximate and elemental characteristics remain within the ranges represented by the training dataset. This provides an appropriate basis for evaluating the external predictive performance and generalization capability of the proposed ANN surrogate model, since the model is tested against an independent sample while maintaining a chemically comparable, yet not identical, biomass composition.
Compared with other lignocellulosic feedstocks commonly investigated for thermochemical conversion, including sugarcane bagasse [38], corn cob [39], wheat straw [40], peanut shell [41], woody residues [42], and fruit-processing residues [43], açaí seed exhibits a distinctive physicochemical profile, characterized particularly by its low ash content, relatively high volatile matter, and high moisture content in the as-received condition, which directly influence its thermochemical conversion behavior.
A key aspect of the present study is that the açaí seed biomass was evaluated in its as-received condition, exactly as generated after pulp processing, without any additional drying step, either under natural conditions or through forced drying in laboratory equipment. Consequently, the relatively high moisture content measured in the samples (approximately 28%) contributes directly to the lower HHV observed in the present study (12.32 MJ kg⁻¹) compared with values commonly reported for dried açaí seeds, which are typically close to 18–21 MJ kg⁻¹ [44]. Thus, the lower HHV obtained here should not be interpreted as an intrinsic limitation of the açaí seed, but rather as a direct reflection of its actual moisture-rich condition immediately after pulp processing. This approach provides a more realistic representation of the feedstock available for thermochemical conversion and avoids artificially increasing its apparent energy density through pre-drying.
More recently, Carvalho et al. [7] reported an HHV of 19.43 MJ kg⁻¹ for raw açaí seeds and demonstrated that progressive torrefaction and pyrolysis increased the HHV to 32.97 MJ kg⁻¹, accompanied by a substantial increase in fixed carbon and a decrease in the H/C and O/C ratios. A relevant comparison can be made with pequi stone (Caryocar brasiliense), another Amazonian fruit-processing residue investigated for thermochemical hydrogen production. Monteles et al. [45] reported 6.74% moisture and 91.31% volatile matter for pequi stone, together with H/C and O/C molar ratios of 1.63 and 0.51, respectively. These values indicate a substantially lower moisture content and higher volatile matter than those observed for the açaí seed investigated in the present study, while the lower O/C ratio suggests a less oxygen-rich carbonaceous feedstock.
The relatively low ash content of açaí seed is nevertheless advantageous for thermochemical processing, as a lower mineral fraction can reduce ash-related operational problems during pyrolysis, gasification, and combustion. Importantly, Monteles et al. [45] demonstrated the suitability of pequi stone for SCWG, with thermodynamic equilibrium simulations using MinG and entropy maximization (MaxS) models, predicting H₂ molar fractions exceeding 50% under favorable operating conditions. Similarly, previous investigations have demonstrated the potential of açaí seed for thermochemical conversion and hydrogen-rich gas production, including pyrolysis [44], hydrothermal processing [46,47] and other reforming/gasification routes [4,5].

3.2. ANN Architecture and Training Algorithm Selection

The ANN configuration was selected by comparing the predictive performance and computational requirements of networks with different architectures and training algorithms. The results obtained for the networks with one hidden layer are presented in Table 2, while those obtained for the two-hidden-layer configurations are presented in Table 3. This comparison was used to assess whether increasing the network complexity resulted in a meaningful improvement in the prediction of the equilibrium yields of H₂, CO, CO₂, and CH₄. Although all evaluated configurations showed high predictive performance, increasing the number of neurons did not consistently improve the statistical indicators, indicating that greater network complexity did not necessarily provide a substantial advantage.
Networks with a single hidden layer generally achieved high predictive performance, although increasing the number of neurons did not consistently improve accuracy and, in several cases, increased the computational cost. The two-hidden-layer configurations provided greater flexibility for representing the nonlinear relationships in the dataset, with the 12–6 topology offering the most favorable balance between predictive performance and computational requirements. Among the evaluated training algorithms, trainlm was selected based on the combined assessment of R², RMSE, MAE, training time, and prediction time. The selected 10–12–6–4 ANN achieved a mean R² of 0.99976, with a training time of 17.608 s, while larger architectures provided only marginal accuracy gains at substantially higher computational costs. For example, the 24–12/trainbr network achieved a mean R² of 0.99979 but required 265.68 s for training. Thus, the 10–12–6–4/trainlm configuration was considered the most appropriate compromise between predictive accuracy and computational efficiency for surrogate modeling of the TeS-MinG calculations. The final ANN architecture is presented in Figure 3.
The selected ANN was trained using the 3,000 simulated data generated from the TeS-MinG simulations, using TeS-Thermo PyPI library. The model required 11.438 s for training and 0.151 s for prediction, demonstrating that the selected configuration provides a computationally efficient representation of the thermodynamic model. The predictive performance and detailed behavior of this final ANN are discussed in the following section.

3.3. Performance of the Selected ANN

The predictive performance of the selected ANN was evaluated using the complete dataset generated from the thermodynamic equilibrium simulations. The model exhibited a high degree of agreement between the predicted and reference values for all four gaseous species considered (H₂, CO, CO₂, and CH₄). The regression plots for the training, validation, testing, and overall datasets (Figure 4) show that the predicted values were closely distributed around the ideal Y = X line, indicating a consistent response of the network across the different data subsets.
The overall correlation coefficients were R = 0.99959, 0.99948, 0.99955, and 0.99956 for the training, validation, testing, and overall datasets, respectively. The small differences among these values indicate that the predictive performance was maintained across the different subsets, with no evident deterioration in the validation or testing stages. This behavior suggests that the selected network was able to capture the nonlinear relationships between biomass composition and operating conditions and the equilibrium gas yields without a substantial loss of predictive capability outside the training subset.
Considering the individual gaseous species, the ANN achieved R2 values of 0.99989, 0.99553, 0.99957, and 0.99963 for H₂, CO, CO₂, and CH₄, respectively. The corresponding RMSE values were 0.015453, 0.058502, 0.1514, and 0.05892 g kg-1, while the MAE values were 0.012308, 0.037926, 0.12725, and 0.04728 g kg-1, respectively. The H₂ and CO₂ predictions showed the highest coefficients of determination, whereas CO presented the lowest R2 among the four species.
The comparatively lower R2 obtained for CO can be associated with the very low equilibrium CO yields over a substantial portion of the investigated operating domain, particularly under conditions of higher water availability. Under SCWG conditions, CO is not only a primary gasification product but also an important intermediate in subsequent equilibrium reactions. In particular, the water-gas shift reaction (WGS, Equation (9)), directly consumes CO while increasing H₂ and CO₂ formation.
line representing ideal agreement and the fitted lines indicating the regression relationship between predicted and target values.
Consequently, increasing water availability favors the conversion of CO toward these products, resulting in very low CO equilibrium yields. CO can also be consumed through methanation reaction (Equation 10), further contributing to its depletion under favorable equilibrium conditions, mainly at high operational temperatures [48]. Previous thermodynamic studies of biomass SCWG have identified water-gas shift and CO methanation as key reactions governing the equilibrium distribution of H₂, CO₂, CO, and CH₄, with CO often occurring at relatively low concentrations compared with H₂ and CO₂ [48,49,50,51].
C O + H 2 O ⇋ C O 2 + H 2
C O + 3 H 2 ⇋ C H 4 + H 2 O
Therefore, the lower R2 for CO should be interpreted in the context of its narrow and low-value response range rather than as evidence of poor predictive capability. Indeed, the RMSE and MAE remained relatively small (0.058502 and 0.037926 g kg⁻¹, respectively), indicating that the absolute deviations between the ANN and TeS-MinG predictions were limited. Similar behavior has been reported in SCWG studies, where CO concentrations decrease as the WGS reaction becomes increasingly favored, particularly at conditions promoting greater interaction between the biomass-derived species and water [48].
The residual distribution further supports the predictive performance of the network. As shown in Figure 5, for all 3,000-simulation data used in ANN training, the standardized residuals were predominantly concentrated around zero and remained largely within the ±2 range, with only a small fraction of observations extending beyond the ± 3 limits, corresponding to outlier fractions of 0.57%, 0.33%, 0.07%, and 0.30% for H₂, CO, CO₂, and CH₄, respectively. The absence of a pronounced systematic displacement of the residuals from zero indicates that the ANN did not exhibit a substantial global prediction bias. Together with the high R2 values and low RMSE and MAE values, these results demonstrate that the selected ANN provides an accurate surrogate representation of the TeS-MinG calculations for the equilibrium yields of H₂, CO, CO₂, and CH₄.
From a computational perspective, the ANN required only 17.067 s for training and 0.151 s to predict the complete dataset. This substantial reduction in computational demand is particularly relevant for applications involving repeated thermodynamic calculations, such as process screening, sensitivity analysis, and optimization. Thus, beyond its predictive accuracy, the selected ANN provides a computationally efficient surrogate of the thermodynamic model, while the independent biomass dataset discussed in the following section provides a more stringent assessment of its generalization capability.

3.4. Validation of ANN Predictions Under Different Operating Conditions

The predictive performance of the selected ANN was further assessed by comparing its responses with those obtained from the TeS-MinG model under systematic variations of the three main operational variables: water-to-biomass molar ratio, pressure, and temperature. This analysis was performed to verify whether the ANN was able to reproduce not only the overall statistical behavior of the reference model, but also the changes in gas yield associated with variations in the operating conditions. The comparisons are presented in Figure 6 (a–c), where the black lines represent the TeS-MinG results and the red markers correspond to the ANN predictions. All gas yields were expressed as g kg⁻¹ of dry biomass.
For the water-to-biomass ratio (Figure 6(a)), the ANN closely followed the trends predicted by TeS-MinG for all four gas species. Increasing the water content resulted in an increase in H₂ yield, whereas CO and CH₄ decreased progressively. CO₂ showed an overall increasing tendency, reaching a relatively stable region at higher water-to-biomass ratios. The ANN reproduced these changes with small deviations from the reference model, with RMSE values of 0.0123, 0.0889, 0.1430, and 0.0498 g kg⁻¹ for H₂, CO, CO₂, and CH₄, respectively. The close agreement across the entire range is particularly relevant because it demonstrates that the ANN captured the nonlinear response of the gas composition to changes in the amount of water supplied to the system.
The effect of pressure is shown in Figure. 6(b). In this case, the ANN also reproduced the main behavior predicted by TeS-MinG. H₂ exhibited a slight decrease with increasing pressure, while CO and CO₂ presented comparatively smaller variations over the investigated range. CH₄ remained at a relatively low yield, with a slight increase toward higher pressures. The deviations between the ANN and TeS-MinG predictions remained small, with RMSE values of 0.0154, 0.0201, 0.1366, and 0.0577 g kg⁻¹ for H₂, CO, CO₂, and CH₄, respectively. Although CO presents a more evident offset between the two models than the other species, the ANN preserved the overall pressure-dependent behavior, indicating that the trained network could represent the influence of pressure on the equilibrium gas composition.
The ANN response to temperature variation is presented in Figure 6(c). Temperature produced the most pronounced changes in gas composition among the investigated operational variables. H₂ increased strongly with temperature before approaching a plateau at the upper end of the range, whereas CH₄ decreased continuously and approached negligible values at high temperatures. CO increased progressively, while CO₂ increased rapidly at lower temperatures and subsequently approached a maximum, followed by a slight decrease at the highest temperature. The ANN predictions closely overlapped the TeS-MinG results throughout these trajectories, reproducing both the increasing and decreasing trends as well as the changes in slope observed in the reference model. For H₂, CO₂, and CH₄, the corresponding RMSE values were 0.0154, 0.1366, and 0.0577 g kg⁻¹, respectively, confirming the low prediction error relative to the magnitude of the predicted yields.
The results in Figure 6 demonstrate that the selected ANN was not limited to reproducing the statistical distribution of the training dataset but was also able to reproduce the physical response trajectories generated by the TeS-MinG model under independent variations of the main operating variables. The agreement observed for water-to-biomass ratio, pressure, and temperature supports the use of the ANN as a surrogate of the thermodynamic model, providing a substantially faster means of estimating gas yields while retaining the characteristic response of the underlying equilibrium calculations.

3.5. External Validation Using an Independent Biomass Dataset

To assess the generalization capability of the developed ANN, an independent validation was performed using a biomass composition not included in model development or training. The comparison with the corresponding TeS-MinG predictions is presented in Figure 7. Overall, the parity plots show strong agreement between both approaches over the investigated gas-yield range, with predictions generally close to the ideal 1:1 line. For H₂, the ANN achieved R² = 0.99398, RMSE = 0.1191 g kg⁻¹, and MAE = 0.0974 g kg⁻¹, with only small deviations, mainly between 850 and 1000 K, where a slight overprediction is observed. CO presented the largest deviations, with R² = 0.98117, RMSE = 0.1256 g kg⁻¹, and MAE = 0.0788 g kg⁻¹. The discrepancies were concentrated mainly between 950 and 1100 K, where the ANN increasingly underpredicted the TeS-MinG values. Despite being more evident than for the other species, these deviations remained relatively small in absolute terms. CO₂ showed excellent agreement, with R² = 0.99555, RMSE = 0.5039 g kg⁻¹, and MAE = 0.4244 g kg⁻¹. Although its RMSE was higher than those of H₂, CO, and CH₄, this reflects the substantially larger CO₂ yields, which reached approximately 69 g kg⁻¹; the relative deviations remained small, particularly above 950 K. CH₄ exhibited the closest agreement, with R² = 0.99905, RMSE = 0.0980 g kg⁻¹, and MAE = 0.0775 g kg⁻¹, with predictions closely distributed around the 1:1 line from approximately 700 to 1100 K and no pronounced temperature-dependent deviation.
The temperature-dependent distribution therefore indicates that the ANN maintains high predictive consistency across the validation range, with the most evident deviations occurring for CO at high temperatures, followed to a lesser extent by H₂ and CO₂, whereas CH₄ remains highly consistent throughout. Because the validation biomass was independent of the training dataset, these results demonstrate that the ANN can transfer the learned nonlinear relationships to an independent biomass composition within the compositional domain represented by the training data. The high R² values and relatively low absolute errors confirm the robustness of the ANN as a surrogate for the TeS-MinG model under different thermochemical conditions.
The responses of the ANN and TeS-MinG models to variations in temperature, feed water content, and pressure are presented in Figure 8(a–c). Overall, the ANN accurately reproduces the trends of the four gaseous species, preserving the nonlinear dependence of gas yields on the operating variables. Temperature produced the most pronounced effects (Figure 8a). H₂ increased continuously from approximately 1.0 g kg⁻¹ at 700 K to 5.1 g kg⁻¹ at 1100 K, with the strongest increase between approximately 750 and 950 K, followed by stabilization at higher temperatures. The ANN captured this nonlinear increase and subsequent plateau with only minor deviations. CO remained very low below approximately 780 K and then increased markedly, reaching approximately 2.3–2.5 g kg⁻¹ at 1100 K; however, the ANN progressively underpredicted the TeS-MinG response above approximately 950 K, consistent with the external-validation results.
The predicted increase in H₂ with temperature agrees with independent experimental observations. Barros et al. [27] reported an increase in H₂ concentration during SCWG of cassava starch at 240 bar and 600–800 K, reaching 65.7 mol% at 800 K, compared with 66.6 mol% from thermodynamic prediction. Similarly, experiments with cellulose, hemicellulose, and lignin identified temperature as the most significant factor affecting H₂ production, with maximum yields at 500 K of 1.95 mmol g⁻¹ for cellulose and 2.26 mmol g⁻¹ for xylose under optimized conditions [54]. Continuous SCWG of biodiesel-industry wastewater also showed increased H₂ production at higher temperatures, reaching 73 mol% H₂ and 357 NmL g⁻¹ of effluent feed at 700 K [55]. Rezende et al. [28] further demonstrated, through experimental investigation combined with thermodynamic simulation and process design, the strong dependence of H₂ production on operating conditions and the applicability of equilibrium-based models to describe biomass gasification. These studies support the physical consistency of the temperature–H₂ relationship predicted by TeS-MinG and successfully reproduced by the ANN.
For CO₂, the yield increased strongly up to approximately 950–1000 K, followed by a near-plateau and slight decrease at the highest temperature. The ANN closely reproduced this behavior, with only minor deviations at high temperatures. In contrast, CH₄ decreased sharply from approximately 8.8 g kg⁻¹ at 700 K to values approaching zero at 1100 K, and the ANN almost completely overlapped the TeS-MinG response. Thus, the model accurately represents both increasing and decreasing nonlinear responses across the investigated temperature range without a systematic deterioration in predictive performance.
The effect of feed water content is shown in Figure 8b. Increasing the water-to-biomass molar ratio from approximately 3 to 6 increased H₂ yield from about 4.65 to 5.2 g kg⁻¹, while CH₄ decreased from approximately 1.0 to 0.1 g kg⁻¹. The ANN closely reproduced both trends, with only small deviations at the highest water contents. CO decreased progressively with increasing water content, with a more noticeable change around a ratio of 4.5–5.0; the ANN captured the overall behavior, although some deviations occurred in this transition region. In contrast, CO₂ increased from approximately 64.5 to 68.5 g kg⁻¹, and the ANN closely followed the TeS-MinG profile. Overall, increasing water availability favored H₂ and CO₂ formation while suppressing CH₄ and, to a lesser extent, CO.
The influence of pressure is presented in Figure 8c and was considerably weaker than that of temperature and feed water. Between approximately 230 and 260 bar, H₂ remained near 5.0–5.2 g kg⁻¹, with only a slight decrease, while CO and CO₂ also exhibited relatively small decreases. The ANN satisfactorily reproduced these weak variations, with minor differences near changes in the slope of the TeS-MinG profiles. CH₄ showed the largest discrepancy: although the ANN correctly reproduced its overall increase with pressure, it predicted a smoother, approximately linear response, whereas TeS-MinG exhibited a step-like increase in the higher-pressure region followed by a gradual rise.
Overall, the response profiles in Figure 8(a–c) indicate that temperature and feed water content exert stronger effects on gas yields than pressure under the investigated conditions, consistent with previous experimental and thermodynamic studies of hydrothermal and supercritical water gasification. Experimental SCWG of Psidium guajava waste demonstrated a strong dependence of H₂, CO₂, CO, and CH₄ formation on temperature and biomass-to-water ratio, with H₂ production reaching 7.743 mol kg⁻¹ [56]. Similar temperature-dependent behavior has been reported for other biomass feedstocks, showing that increasing temperature generally favors H₂ formation while altering the relative contributions of reactions governing CH₄ and CO₂ production [3]. Thermodynamic studies have likewise identified temperature and feed concentration as dominant operating variables, whereas pressure has a comparatively smaller effect within the typical SCWG range of approximately 20–35 MPa [48]. The importance of the biomass-to-water ratio is also associated with its influence on reactions such as water-gas shift and steam reforming [8]. The limited pressure effect observed here is further consistent with previous SCWG studies reporting smaller changes in gas composition due to pressure than those associated with temperature and feed concentration [29].
Thus, the trends predicted by TeS-MinG are consistent with experimentally observed SCWG behavior, providing physical support for the model predictions. The close agreement between ANN and TeS-MinG profiles demonstrates that the surrogate preserves these physically meaningful nonlinear relationships across the investigated operating conditions. Although localized deviations occur, particularly for CO at high temperatures and CH₄ at higher pressures, they do not alter the overall response patterns, confirming the ANN as an adequate surrogate for the underlying thermodynamic model.

3.6. ANN-Based Analysis of the Coupled Effects of Temperature and H₂O/Biomass Ratio on Gas Yields

The previous analysis demonstrated that temperature and H₂O/biomass ratio exert substantially stronger effects on the predicted gas yields than pressure within the investigated operating range. Based on this observation, pressure was fixed at 230 bar for the subsequent response analysis. This condition was selected to reduce the dimensionality of the analysis while retaining the two operating variables that most strongly affect the predicted gas distribution.
The objective of this analysis was not to further validate the ANN, but to demonstrate its capability as a surrogate model for exploring the coupled effects of temperature and water availability on gas production during SCWG of açaí seed biomass. Accordingly, temperature was varied from 700 to 1100 K and the H₂O/biomass molar ratio from 3 to 6, while the elemental composition of Sample 4 and pressure were kept constant. A total of 10,000 operating points were evaluated using a 100 × 100 computational grids, and the ANN was used to predict H₂, CO, CO₂, and CH₄ yields at each point. The resulting predictions were then used to construct the two-dimensional response maps presented in Figure 9. Importantly, all operating points were generated within the ANN training domain, avoiding extrapolation beyond the conditions represented in the model development dataset.
As shown in Figure 9a, the H₂ yield exhibits a strong and predominantly positive dependence on temperature. At 700 K, the predicted H₂ yield is approximately 0.7–1.2 g kg⁻¹, depending on the H₂O/biomass molar ratio. As the temperature increases, the H₂ yield rises continuously, reaching values of approximately 4.5–4.7 g kg⁻¹ at 1100 K. The effect of the H₂O/biomass ratio is comparatively less pronounced than that of temperature, although increasing water availability generally favors H₂ production [47,49,51,55]. For example, at intermediate temperatures around 850–900 K, increasing the H₂O/biomass ratio from 3 to 6 results in an increase in H₂ yield of approximately 0.5–1.0 g kg⁻¹, depending on the operational temperature. The maximum H₂ production is therefore observed in the high-temperature region, particularly at elevated H₂O/biomass ratios, in agreement with previous studies reporting enhanced H₂ production at higher temperatures and water-to-biomass ratios. This behavior is consistent with the greater extent of steam-assisted reactions at elevated temperatures, including reforming and water–gas-shift-related pathways, which favor the formation of hydrogen-rich gaseous products [57].
The response surface for CO presents a different behavior (Figure 9b). At temperatures below approximately 800 K, the CO yield remains close to zero, generally below 0.3–0.5 g kg⁻¹, over most of the investigated H₂O/biomass range. A marked transition occurs between approximately 800 and 900 K, after which the CO yield progressively increases with temperature. At 1100 K, the predicted CO yield reaches approximately 3.0–3.2 g kg⁻¹ at the lower H₂O/biomass ratios, whereas substantially lower values are obtained at higher water availability. Thus, at the highest temperature, changing the H₂O/biomass ratio from approximately 3 to 6 can reduce the CO yield by roughly 1.0–1.5 g kg⁻¹. This behavior indicates that the availability of steam does not affect all gaseous products in the same manner. Although additional water can promote reforming reactions, it can also favor pathways that consume CO, particularly through the water–gas shift reaction, thereby producing competing effects on the final equilibrium CO yield [58].
The CO₂ map (Figure 9c) also shows a strong dependence on temperature. At 700 K, the predicted CO₂ yield is approximately 48–52 g kg⁻¹, depending on the H₂O/biomass ratio. With increasing temperature, the CO₂ yield increases substantially, reaching approximately 60–65 g kg⁻¹ over a broad region of the high-temperature domain. The total variation associated with temperature is therefore on the order of 10–15 g kg⁻¹. Compared with H₂, the CO₂ surface exhibits more pronounced local variations, particularly between approximately 750 and 850 K. These local features correspond to changes in the slope of the response surface and indicate that the effect of temperature is not strictly uniform throughout the operating domain. At high temperatures, the CO₂ yield remains relatively high, although the H₂O/biomass ratio continues to influence the magnitude of the predicted yield [56]. The difference between the low- and high-water regions becomes particularly evident at intermediate and high temperatures, where variations of several g kg⁻¹ can be observed across the CO2 surface.
In contrast, CH₄ exhibits an overall decreasing trend with increasing temperature (Figure 9d). At 700 K, the predicted CH₄ yield reaches approximately 7.0–8.0 g kg⁻¹, with the highest values occurring at the lower H₂O/biomass ratios. As temperature increases, the CH₄ yield progressively decreases, falling below approximately 1.0 g kg⁻¹ at temperatures above 1000 K and approaching 0.0 g kg⁻¹ in the highest-temperature region. Consequently, increasing the temperature from 700 to 1100 K can reduce the CH₄ yield by approximately 7.0–8.0 g kg⁻¹. The effect of the H₂O/biomass ratio is also evident, particularly at lower and intermediate temperatures, where increasing water availability contributes to a further reduction in methane formation. This behavior is consistent with the thermodynamic tendency of methane to become less favorable at elevated temperatures, where steam reforming and other reactions promote the conversion of CH₄ into lighter gaseous products, particularly H₂ and CO [59]. Thus, the CH₄ surface provides a complementary response to the H₂ surface, highlighting the redistribution of carbon and hydrogen among the equilibrium gas species as operating conditions change.
The results further demonstrate the advantage of using the ANN as a surrogate model for TeS-MinG. Instead of performing independent thermodynamic equilibrium calculations for approximately 10,000 operating conditions, the trained ANN rapidly evaluated the entire computational grid, enabling the generation of continuous response maps while preserving the nonlinear relationships learned from the TeS-MinG simulations. This approach facilitates the rapid identification of favorable operating regions and can support subsequent optimization, sensitivity analysis, and multi-objective process design without requiring a new equilibrium calculation for each condition.
The response maps should, however, be interpreted as surrogate representations of the thermodynamic behavior predicted by TeS-MinG rather than as independent experimental measurements. Their agreement with the underlying model supports the use of the ANN for parametric exploration within the training domain. Attention should be given to regions with steep gradients or local changes in the response surfaces, especially for CO and CO₂, where small variations in temperature or water availability may produce appreciable changes in the equilibrium gas distribution.

3.7. Computational Efficiency of the ANN Surrogate

Beyond its predictive performance, the computational efficiency of the developed ANN is a key aspect of its application as a surrogate model for the TeS-MinG framework. The thermodynamic calculations require the iterative solution of a Gibbs energy minimization problem for each combination of biomass composition and operating conditions. Consequently, the computational demand can increase substantially when many simulations is required, particularly for parametric studies, sensitivity analyses, process optimization, uncertainty assessment, and process-level simulations [60,61]. In contrast, once trained, the ANN predicts the gas yields through direct forward propagation of the input variables through the network, without requiring the iterative solution of the underlying thermodynamic equilibrium problem.
The developed ANN was trained using feed water content, temperature, pressure, and biomass elemental composition as input variables, while H₂, CO, CO₂, and CH₄ yields were predicted simultaneously. This multi-output configuration allows the four gaseous species to be obtained from a single network evaluation, further contributing to the computational efficiency of the surrogate approach. Although ANN development involves an initial computational cost associated with generating the TeS-MinG database and optimizing the network parameters, this cost is incurred only once during model development. After training and validation, ANN evaluation requires only forward propagation through the trained network, resulting in a substantially lower computational cost for repeated evaluations of different operating conditions.
The computational advantage of the ANN becomes increasingly relevant as the number of required model evaluations increases. As summarized in Table 4, TeS-MinG required approximately 30.6 s for 1,000 simulations and 8.5 h for 1,000,000 simulations, whereas the trained ANN required only 0.069 and 0.2317 s, respectively. These results demonstrate the substantially lower computational requirements of ANN evaluation for extensive exploration of the operating space.
The computational times were determined using the same MacBook Air M2 computer with 16 GB of RAM. The TeS-MinG simulations were performed using the publicly available TeS-Thermo Python library distributed through PyPI, whereas the ANN evaluations were performed using the trained surrogate model. The difference in computational requirements reflects the fundamentally different numerical procedures involved: TeS-MinG requires the iterative solution of a Gibbs free-energy minimization problem for each operating condition, whereas the trained ANN performs a direct numerical mapping from the input variables to the predicted gas yields. Therefore, once trained and validated, the ANN provides a computationally efficient surrogate for repeated evaluations of the thermodynamic model.
The computational benefit of the ANN is particularly relevant when considered together with its predictive performance. External validation yielded R² values above 0.98 for all four gaseous species, while the ANN also reproduced the principal response trends associated with temperature, feed water content, and pressure. Thus, the reduction in computational cost was achieved while preserving the main thermodynamic response patterns predicted by TeS-MinG. The ANN can therefore be used to rapidly explore operating conditions after the thermodynamic database has been generated and the surrogate has been appropriately trained and validated.
Importantly, the proposed approach is not intended to replace TeS-MinG as the thermodynamic basis of the predictions. Rather, TeS-MinG serves as the thermodynamically rigorous reference model used to generate the database and establish the input–output relationships, whereas the ANN provides a computationally efficient representation of these relationships for subsequent evaluations. This complementary strategy combines the thermodynamic foundation of Gibbs energy minimization with the computational efficiency of machine learning, making the proposed surrogate particularly suitable for large-scale parametric analyses, process optimization, sensitivity and uncertainty analyses, and integration into higher-level process simulations.

5. Conclusions

This study developed and externally validated an artificial neural network (ANN) surrogate using an independent, unseen simulated dataset to predict the equilibrium yields of H₂, CO, CO₂, and CH₄ during the supercritical water gasification (SCWG) of açaí seed biomass. The main conclusions are as follows:
  • The TeS-MinG Gibbs energy minimization framework was successfully used to generate a thermodynamic database covering variations in biomass elemental composition, temperature, pressure, and water-to-biomass ratio. The use of experimentally characterized biomass samples allowed compositional variability to be incorporated into the equilibrium calculations.
  • The selected 10–12–6–4 ANN, trained using the Levenberg–Marquardt algorithm, accurately predicted the equilibrium yields of H₂, CO, CO₂, and CH₄. The model achieved high predictive performance, with R2 values of 0.99989, 0.99553, 0.99957, and 0.99963, respectively, demonstrating its ability to simultaneously represent the four gas species during SCWG process.
  • The ANN successfully reproduced the nonlinear response profiles generated by TeS-MinG as a function of water-to-biomass ratio, pressure, and temperature, including the increasing, decreasing, plateau, and slope-change behaviors observed for the different gas species in response to operational variables changes.
  • The ANN maintained strong agreement with TeS-MinG when evaluated using an independent açaí seed biomass simulated composition, that was not used during model development. All four species presented R2 > 0.98, confirming that the network was able to generalize the learned relationships beyond the biomass compositions used for training. The largest deviations occurred for CO at higher temperatures, although the overall predictive agreement remained satisfactory.
  • The ANN substantially reduced the computational effort required for repeated evaluations. For one million input points, the ANN required only 0.2317 s, whereas the corresponding TeS-MinG simulations were estimated to require approximately 8.5 h. Although these values were obtained under different benchmarking conditions and should not be interpreted as a controlled speed comparison, they demonstrate the substantially lower computational cost of ANN inference.
  • The results demonstrate that the proposed ANN can function as a computationally efficient surrogate model of TeS-MinG, combining the thermodynamic foundation of Gibbs energy minimization with the rapid evaluation capability of machine learning methods. The approach is therefore suitable for large-scale parametric exploration and provides a basis for process optimization, sensitivity and uncertainty analyses, and integration into higher-level process simulations.
This study demonstrates that a thermodynamically grounded ANN surrogate can accurately reproduce the equilibrium behavior of SCWG of real açaí seed biomass, generalize to an independent biomass composition within the investigated compositional domain, and substantially reduce the computational burden associated with repeated equilibrium calculations.

Author Contributions

Conceptualization, A.C.D.Freitas and A.D.S.Vidotti; methodology, H.S.Rocha, L.P.Reis, M.S.Monteles, J.M.Santos Júnior., R.Guirardello and A.C.D.Freitas; software, A.C.D.Freitas, J.M.Santos Júnior. and R.Guirardello; validation, H.S.Rocha, L.P.Reis, A.D.S.Vidotti and A.C.D.Freitas.; formal analysis, H.S.Rocha, L.P.Reis, M.S.Montelles and A.C.D.Freitas; investigation, H.S.Rocha, L.P.Reis and M.S.Monteles; resources, A.D.S.Vidotti, T.V.Barros, L.Cardozo-Filho. and A.C.D.Freitas; data curation, H.S.Rocha, L.P.Reis and M.S.Montelles; writing—original draft preparation, H.S.Rocha and A.C.D.Freitas; writing—review and editing, A.D.S.Vidotti, T.V.Barros, and A.C.D.Freitas; visualization, H.S.Rocha and L.P.Reis; supervision, A.D.S.Vidotti, R.Guirardello and A.C.D.Freitas; project administration, A.C.D.Freitas; All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Data Availability Statement

The data generated and analyzed during the current study, including the thermodynamic simulation results and datasets used for the development, training, and validation of the artificial neural network model, are available from the corresponding author upon reasonable request.

Acknowledgments

Financial support from the Human Resources Training Program of the National Agency of Petroleum, Natural Gas and Biofuels (PRH 54.1-ANP), managed by the São Paulo Research Foundation (FAPESP), Brazil, under grant numbers 2025/02550-5 and 2025/17586-5.

Conflicts of Interest

The authors declare no conflicts of interest.

Abbreviations

The following abbreviations are used in this manuscript:
ANN Artificial Neural Network
ANNs Artificial Neural Networks
ASTM American Society for Testing and Materials
CHNS Carbon, Hydrogen, Nitrogen and Sulfur
EoS Equation of State
HHV Higher Heating Value
MAE Mean Absolute Error
ML Machine Learning
MinG Gibbs Energy Minimization
PR Peng Robinson
RMSE Root Mean Square Error
R2 Coefficient of Determination
SCWG Supercritical Water Gasification
SRK Soave-Redlich-Kwong
TeS-MinG Gibbs Energy Minimization using PyPI Library
TeS-Thermo Thermodynamic Equilibrium Simulation PyPI Library
WGS Water Gas Shift

References

  1. Rodriguez Correa, C.; Kruse, A. Supercritical Water Gasification of Biomass for Hydrogen Production—Review. J. Supercrit. Fluids 2018, 133, 573–590. [Google Scholar] [CrossRef]
  2. Kruse, A. Supercritical Water Gasification. Biofuels Bioprod. Biorefin. 2008, 2, 415–437. [Google Scholar] [CrossRef]
  3. Itai, Y.; Santos, R.; Branquinho, M.; Malico, I.; Ghesti, G.F.; Brasil, A.M. Numerical and Experimental Assessment of a Downdraft Gasifier for Electric Power in Amazon Using Açaí Seed (Euterpe oleracea Mart.) as a Fuel. Renew. Energy 2014, 66, 662–669. [Google Scholar] [CrossRef]
  4. Instituto Brasileiro de Geografia e Estatística (IBGE). Produção da Extração Vegetal e da Silvicultura 2024; IBGE: Rio de Janeiro, Brazil, 2025. [Google Scholar]
  5. Musial, C.M.; Staudt, J.; Reichert, A.V.; da Silva, E.A.; Ribeiro, C.; Lima, F.V.; Borba, C.E. Modeling and Simulation of the Gasification of Euterpe oleracea Waste: Optimization of Hydrogen Production and Energy Potential through Operability Analysis. ACS Omega 2026, 11, 34034–34046. [Google Scholar] [CrossRef] [PubMed]
  6. da Silva, C.M.S.; de Castro, D.A.R.; Santos, M.C.; Almeida, H.S.; Schultze, M.; Lüder, U.; Hoffmann, T.; Machado, N.T. Process Analysis of Main Organic Compounds Dissolved in Aqueous Phase by Hydrothermal Processing of Açaí (Euterpe oleraceae, Mart.) Seeds: Influence of Process Temperature, Biomass-to-Water Ratio, and Production Scales. Energies 2021, 14, 5608. [Google Scholar] [CrossRef]
  7. de Carvalho, A.L.; Silveira, E.A.; Protásio, T. de P.; de Oliveira, T.J.P. Açaí Seed Valorization through Torrefaction and Pyrolysis: Pathways toward High-Energy and Low-Carbon Biocoal. Biomass Bioenergy 2026, 211, 109192. [Google Scholar] [CrossRef]
  8. Freitas, A.C.D.; Guirardello, R. Comparison of several glycerol reforming methods for hydrogen and syngas production using Gibbs energy minimization. Int. J. Hydrogen Energy 2014, 39, 17969–17984. [Google Scholar] [CrossRef]
  9. Silva, I.P.; Lima, R.M.A.; Silva, G.F.; Ruzene, D.S.; Silva, D.P. Thermodynamic Equilibrium Model Based on Stoichiometric Method for Biomass Gasification: A Review of Model Modifications. Renew. Sustain. Energy Rev. 2019, 114, 109305. [Google Scholar] [CrossRef]
  10. Roos, W.A.; Zietsman, J.H. Accelerating Complex Chemical Equilibrium Calculations—A Review. Calphad 2022, 77, 102380. [Google Scholar] [CrossRef]
  11. Zhao, S.; Li, J.; Chen, C.; Yan, B.; Tao, J.; Chen, G. Interpretable Machine Learning for Predicting and Evaluating Hydrogen Production via Supercritical Water Gasification of Biomass. J. Clean. Prod. 2021, 316, 128244. [Google Scholar] [CrossRef]
  12. Shenbagaraj, S.; Sharma, P.K.; Sharma, A.K.; Raghav, G.; Kota, K.B.; Ashokkumar, V. Gasification of food Waste in Supercritical Water: An Innovative Synthesis Gas Composition Prediction Model Based on Artificial Neural Networks. Int. J. Hydrogen Energy 2021, 46, 12739–12757. [Google Scholar] [CrossRef]
  13. Deepankumar, S.; Senthil Kumar, K.L. Hydrothermal liquefaction and gasification of industrial waste algae: Experimental and AI-assisted optimization for biofuel and hydrogen production. Biomass Conv. Bioref. 2026, 16, 125. [Google Scholar] [CrossRef]
  14. Rodgers, S.; Bowler, A.; Wells, L.; Lee, C.S.; Hayes, M.; Poulston, S.; Lester, E.; Meng, F.; McKechnie, J.; Conradie, A. A Surrogate Model for the Economic Evaluation of Renewable Hydrogen Production from Biomass Feedstocks via Supercritical Water Gasification. Int. J. Hydrogen Energy 2024, 49, 277–294. [Google Scholar] [CrossRef]
  15. Jankovic, S.; Stanger, L.; Bartik, A.; Hammerschmid, M.; Benedikt, F.; Mittermayr, M.; Binder, M.; Kozek, M.; Müller, S. Design and Implementation of a Digital Twin for a Biomass-to-Gas Plant. Appl. Energy 2025, 402, 126861. [Google Scholar] [CrossRef]
  16. Akhator, P.; Oboirien, B. Digitilising the Energy Sector: A Comprehensive Digital Twin Framework for Biomass Gasification Power Plant with CO₂ Capture. Clean. Energy Syst. 2025, 10, 100175. [Google Scholar] [CrossRef]
  17. ASTM INTERNATIONAL. ASTM D3173-11: Standard Test Method for Moisture in the Analysis Sample of Coal and Coke; ASTM International. West Conshohocken, 2011. [CrossRef]
  18. ASTM INTERNATIONAL. ASTM D3174-12: Standard Test Method for Ash in the Analysis Sample of Coal and Coke from Coal; ASTM International. West Conshohocken, 2012. [CrossRef]
  19. ASTM INTERNATIONAL. ASTM D3175-20: Standard Test Method for Volatile Matter in the Analysis Sample of Coal and Coke; ASTM International. West Conshohocken, 2020. [CrossRef]
  20. ASTM INTERNATIONAL. ASTM D3172-13: Standard Practice for Proximate Analysis of Coal and Coke; ASTM International. West Conshohocken, 2013. [CrossRef]
  21. dos Santos Junior, J.M.; Daltro de Freitas, A.C.; Mariano, A.P. A Python-Based Thermodynamic Equilibrium Library for Gibbs Energy Minimization: A Case Study on Supercritical Water Gasification of Ethanol and Methanol. Eng 2025, 6, 208. [Google Scholar] [CrossRef]
  22. Castillo, J.; Grossmann, I.E. Computation of phase and chemical equilibria. Comput. Chem. Eng. 1981, 5, 99–108. [Google Scholar] [CrossRef]
  23. Robinson, D.B.; Peng, D.-Y.; Ng, H.-J. Applications of the Peng-Robinson Equation of State; American Chemical Society: Washington, DC, USA, 1977; pp. 200–220. [Google Scholar] [CrossRef]
  24. Smith, J.M.; Van Ness, H.C.; Abbott, M.M.; Swihart, M.T. Introduction to Chemical Engineering Thermodynamics; McGraw-Hill: Singapore, 2018. [Google Scholar]
  25. Soave, G. Equilibrium constants from a modified Redlich-Kwong equation of state. Chem. Eng. Sci. 1972, 27, 1197–1203. [Google Scholar] [CrossRef]
  26. Tsonopoulos, C. An empirical correlation of second virial coefficients. AIChE J. 1974, 20, 263–272. [Google Scholar] [CrossRef]
  27. Barros, T. V.; de Freitas, A. C. D.; Aranha, A. C. R.; Guirardello, R.; Ferreira-Pinto, L.; Cardozo-Filho, L. Supercritical Water Gasification of Cassava Starch as a Model Component of Cassava Wastewater: Thermodynamic and Experimental Analysis. Starch-Stärke 2026, 78(no. 6), e70250. [Google Scholar] [CrossRef]
  28. Rezende, R.F.; Souza, G.B.M.; Mourão, L.C.; Dias, I.M.; Vidotti, A.D.S.; Freitas, A.C.D.; Pavão, L.V.; García-Jarana, M.B.; Abelleira-Pereira, J.M.; Miguelez, J.R.P. Experimental, simulation and process design of hydrogen production from biomass-based products co-gasification in supercritical water. Fuel 2026, 404, 136316. [Google Scholar] [CrossRef]
  29. Barros, T.V.; Carregosa, J.d.C.; Wisniewski, A., Jr.; Freitas, A.C.D.; Guirardello, R.; Ferreira-Pinto, L.; Bonfim-Rocha, L.; Jegatheesan, V.; Cardozo-Filho, L. Assessment of black liquor hydrothermal treatment under sub- and supercritical conditions: Products distribution and economic perspectives. Chemosphere 2022, 286, 131774. [Google Scholar] [CrossRef] [PubMed]
  30. Junior, J.M.D.S.; Mariano, A.P. Gasification of Lignocellulosic Waste in Supercritical Water: Study of Thermodynamic Equilibrium as a Nonlinear Programming Problem. Eng 2024, 5, 1096–1111. [Google Scholar] [CrossRef]
  31. dos Santos Junior, J.M.; de Freitas, A.C.D.; Mariano, A.P. An AI-Assisted Thermodynamic Equilibrium Simulator: A Case Study on Steam Methane Reforming in Isothermal and Adiabatic Reactors. Processes 2025, 13, 2508. [Google Scholar] [CrossRef]
  32. Benazir, S.; Ghauri, M.; Chauhdary, S.T.; Asghar, A.B.; Jaffery, M.H.; Amentie, M.D. Accurate Forecasting of Bio-Syngas Composition via ANN in Catalytic Biomass Gasification. Discov. Comput. 2026, 29, 231. [Google Scholar] [CrossRef]
  33. Postawa, K.; Klimek, K.; Maj, G.; Kapłan, M.; Szczygieł, J. Advanced Dual-Artificial Neural Network System for Biomass Combustion Analysis and Emission Minimization. J. Environ. Manag. 2024, 349, 119543. [Google Scholar] [CrossRef] [PubMed]
  34. Sezer, S.; Kartal, F.; Özveren, U. Prediction of combustion reactivity for lignocellulosic fuels by means of machine learning. J. Therm. Anal. Calorim. 2022, 147, 9793–9809. [Google Scholar] [CrossRef]
  35. Abdoulmoumine, N.; Adhikari, S.; Kulkarni, A.; Chattanathan, S. A Review on Biomass Gasification Syngas Cleanup. Appl. Energy 2015, 155, 294–307. [Google Scholar] [CrossRef]
  36. Rey, J.R.C.; Longo, A.; Rijo, B.; Pedrero, C.M.; Tarelho, L.A.C.; Brito, P.S.D.; Nobre, C. A Review of Cleaning Technologies for Biomass-Derived Syngas. Fuel 2024, 377, 132776. [Google Scholar] [CrossRef]
  37. Channiwala, S. A.; Parikh, P. P. A unified correlation for estimating HHV of solid, liquid and gaseous fuels. Fuel 2002, v. 81(n. 8), 1051–1063. [Google Scholar] [CrossRef]
  38. Pedroso, D.T.; Machin, E.B.; Cabrera-Barjas, G.; Flores, M.; Urra, H.G.; de Carvalho, F.S.; Santos, M.I.S.d.; Machín, A.B.; Canettieri, E.V.; Pérez, N.P.; et al. Sugarcane Bagasse Torrefaction for Fluidized Bed Gasification. Appl. Sci. 2021, 11, 6105. [Google Scholar] [CrossRef]
  39. Čolnik, M.; Irgolič, M.; Perva, A.; Škerget, M. Hydrolytic Decomposition of Corncobs to Sugars and Derivatives Using Subcritical Water. Processes 2025, 13, 267. [Google Scholar] [CrossRef]
  40. Lei, J.; Ye, X.; Wang, H.; Zhao, D. Insights into Pyrolysis Kinetics, Thermodynamics, and the Reaction Mechanism of Wheat Straw for Its Resource Utilization. Sustainability 2023, 15, 12536. [Google Scholar] [CrossRef]
  41. Xu, Z.; Xiao, X.; Fang, P.; Ye, L.; Huang, J.; Wu, H.; Tang, Z.; Chen, D. Comparison of Combustion and Pyrolysis Behavior of the Peanut Shells in Air and N2: Kinetics, Thermodynamics and Gas Emissions. Sustainability 2020, 12, 464. [Google Scholar] [CrossRef]
  42. Moreira, J.; Carneiro, A.; Oliveira, D.; Santos, F.; Guerra, D.; Nogueira, M.; Rocha, H.; Charvet, F.; Tarelho, L. Thermochemical Properties for Valorization of Amazonian Biomass as Fuel. Energies 2022, 15, 7343. [Google Scholar] [CrossRef]
  43. Cavalaglio, G.; Fabbrizi, G.; Cardelli, F.; Lorenzi, L.; Angrisano, M.; Nicolini, A. Lignocellulosic Residues from Fruit Trees: Availability, Characterization, and Energetic Potential Valorization. Energies 2024, 17, 2611. [Google Scholar] [CrossRef]
  44. de Castro, D.A.R.; Ribeiro, H.J.d.S.; Guerreiro, L.H.H.; Assunção, F.P.d.C.; Bernar, L.P.; Silva, N.P.d.; Guimarães, D.M.D.; Monteiro, M.C.; Pizarro Borges, L.E.; Kuchta, K.; et al. Catalytic Pyrolysis of Açaí (Euterpe oleracea Mart.) Seeds: Circular Economy for Agro-Industrial Waste-to-Energy in the Amazon. Catalysts 2026, 16, 485. [Google Scholar] [CrossRef]
  45. Monteles, M.S.; dos Reis, L.P.; da Costa, D.S.; dos Santos Júnior, J.M.; da Silva, D.C.; Barros, T.V.; Souza Vidotti, A.D.; Daltro de Freitas, A.C.; Guirardello, R. Experimental and Thermodynamic Assessment of Caryocar brasiliense (Pequi) Stone for Hydrogen Production via Supercritical Water Gasification. Chem. Eng. Trans. 2026, 125, 145–150. [Google Scholar] [CrossRef]
  46. Cruz-Reina, L.J.; Velásquez, F.; Espitia, J.; Villagrán, E.; Rodríguez, J. A Review of Thermochemical, Physical, and Chemical Conversion Pathways of Coconut and Açaí Residues: Technological Progress and Readiness Assessment. Processes 2026, 14, 1314. [Google Scholar] [CrossRef]
  47. Costa, Y.C.M.D.; Coelho, A.C.P.P.; Santos Júnior, J.M.D.; Vidotti, A.D.S.; Freitas, A.C.D.D.; Guirardello, R. Thermochemical Valorization of Açaí (Euterpe oleracea) Seeds: A Study Using Gibbs Energy Minimization and Entropy Maximization Methods. Chem. Eng. Trans. 2025, 117, 943–948. [Google Scholar] [CrossRef]
  48. Yan, Q.; Guo, L.; Lu, Y. Thermodynamic analysis of hydrogen production from biomass gasification in supercritical water. Energy Convers. Manag. 2006, 47, 1515–1528. [Google Scholar] [CrossRef]
  49. Castello, D.; Fiori, L. Supercritical water gasification of biomass: A stoichiometric thermodynamic model. Int. J. Hydrogen Energy 2015, 40, 6771–6781. [Google Scholar] [CrossRef]
  50. Lee, N.-G.; Kim, M.-S.; Ihm, S.-K. Gasification of glucose in supercritical water. Ind. Eng. Chem. Res. 2002, 41, 1182–1188. [Google Scholar] [CrossRef]
  51. dos Santos Junior, J.M.; Vidotti, A.D.; Freitas, A.C.; Guirardello, R. Black Liquor Gasification in Supercritical Water: Thermodynamic Study for Isothermal Systems. Chem. Eng. Trans. 2024, 109, 205–210. [Google Scholar] [CrossRef]
  52. Okolie, J.A.; Nanda, S.; Dalai, A.K.; Kozinski, J.A. Optimization and modeling of process parameters during hydrothermal gasification of biomass model compounds to generate hydrogen-rich gas products. Int. J. Hydrogen Energy 2020, 45, 18275–18288. [Google Scholar] [CrossRef]
  53. Okolie, J.A.; Nanda, S.; Dalai, A.K.; Kozinski, J.A. Optimization studies for hydrothermal gasification of partially burnt wood from forest fires for hydrogen-rich syngas production using Taguchi experimental design. Environ. Pollut. 2021, 283, 117040. [Google Scholar] [CrossRef] [PubMed]
  54. Okolie, J.A.; Nanda, S.; Dalai, A.K.; Kozinski, J.A. Hydrothermal gasification of soybean straw and flax straw for hydrogen-rich syngas production: Experimental and thermodynamic modeling. Energy Convers. Manag. 2020, 208, 112545. [Google Scholar] [CrossRef]
  55. Teixeira, I.R.; Dias, I.M.; Mourão, L.C.; Andrade, L.A.; Pavão, L.V.; Abelleira-Pereira, J.M.; Souza, G.B.M.; Cardozo-Filho, L.; Alonso, C.G.; Guirardello, R. Continuous Hydrogen Production via Hydrothermal Gasification of Biodiesel Industry Wastewater: Experimental Optimization and Energy Integration Simulation. Water 2023, 15, 4062. [Google Scholar] [CrossRef]
  56. González-Arias, S.; Zúñiga-Moreno, A.; García-Morales, R.; Elizalde-Solis, O.; Verónico-Sánchez, F.J.; Flores-Valle, S.O. Gasification of Psidium guajava L. Waste Using Supercritical Water: Evaluation of Feed Ratio and Moderate Temperatures. Energies 2021, 14, 2555. [Google Scholar] [CrossRef]
  57. Ding, N.; Azargohar, R.; Dalai, A.K.; Kozinski, J.A. Catalytic gasification of cellulose and pinewood to H₂ in supercritical water. Fuel 2014, 118, 416–425. [Google Scholar] [CrossRef]
  58. Moravvej, Z.; Bazargani, Z.; Esmaeilzadeh, F. Thermodynamic Modeling and Optimization of Biomass and Bio-Renewable Organic Source Gasification in Supercritical Water Using Gibbs Free Energy Minimization. Water 2024, 16, 2123. [Google Scholar] [CrossRef]
  59. Li, H.; Lei, J.; Jia, M.; Xu, H.; Wu, S. High Dimensional Model Representation Approach for Prediction and Optimization of the Supercritical Water Gasification System Coupled with Photothermal Energy Storage. Processes 2023, 11, 2313. [Google Scholar] [CrossRef]
  60. Moretti, L.; Arpino, F.; Cortellessa, G.; Di Fraia, S.; Di Palma, M.; Vanoli, L. Reliability of Equilibrium Gasification Models for Selected Biomass Types and Compositions: An Overview. Energies 2022, 15, 61. [Google Scholar] [CrossRef]
  61. Tang, H.; Kitagawa, K. Supercritical Water Gasification of Biomass: Thermodynamic Analysis with Direct Gibbs Free Energy Minimization. Chem. Eng. J. 2005, 106, 261–267. [Google Scholar] [CrossRef]
Figure 1. Workflow adopted for the development of the ANN surrogate model for predicting the equilibrium gas composition during supercritical water gasification of açaí seed biomass from MinG thermodynamic data.
Figure 1. Workflow adopted for the development of the ANN surrogate model for predicting the equilibrium gas composition during supercritical water gasification of açaí seed biomass from MinG thermodynamic data.
Preprints 235313 g001
Figure 2. Schematic representation of the ANN model-selection methodology. Networks with one hidden layer were evaluated using 5, 8, 10, 12, 15 and 20, while two-hidden-layer architectures were tested with 8–4, 12–6, 16–8, 20–10, and 24–12 neurons.
Figure 2. Schematic representation of the ANN model-selection methodology. Networks with one hidden layer were evaluated using 5, 8, 10, 12, 15 and 20, while two-hidden-layer architectures were tested with 8–4, 12–6, 16–8, 20–10, and 24–12 neurons.
Preprints 235313 g002
Figure 3. Architecture of the selected artificial neural network (ANN) model (10–12–6–4), comprising 10 input neurons, two hidden layers with 12 and 6 neurons using hyperbolic tangent sigmoid (tanh) activation functions, and an output layer with 4 neurons using a linear activation function. The four outputs correspond to the predicted yields of H₂, CO, CO₂, and CH₄.
Figure 3. Architecture of the selected artificial neural network (ANN) model (10–12–6–4), comprising 10 input neurons, two hidden layers with 12 and 6 neurons using hyperbolic tangent sigmoid (tanh) activation functions, and an output layer with 4 neurons using a linear activation function. The four outputs correspond to the predicted yields of H₂, CO, CO₂, and CH₄.
Preprints 235313 g003
Figure 4. Regression analysis of the selected ANN for the training, validation, testing, and overall datasets. The plots compare the ANN outputs with the corresponding target values, with the Y = X
Figure 4. Regression analysis of the selected ANN for the training, validation, testing, and overall datasets. The plots compare the ANN outputs with the corresponding target values, with the Y = X
Preprints 235313 g004
Figure 5. Standardized residual analysis of the selected ANN for the prediction of H₂, CO, CO₂, and CH₄ yields. The dashed red lines represent the ±2 standardized residual limits, while the dashed magenta lines indicate the ±3 limits. The corresponding R2, RMSE, and percentage of outliers are reported for each gas species.
Figure 5. Standardized residual analysis of the selected ANN for the prediction of H₂, CO, CO₂, and CH₄ yields. The dashed red lines represent the ±2 standardized residual limits, while the dashed magenta lines indicate the ±3 limits. The corresponding R2, RMSE, and percentage of outliers are reported for each gas species.
Preprints 235313 g005
Figure 6. Comparison between TeS-MinG predictions and ANN responses for H₂, CO, CO₂, and CH₄ yields under variations in (a) water-to-biomass molar ratio, (b) pressure, and (c) temperature. Black lines represent TeS-MinG predictions, whereas red markers indicate ANN predictions. RMSE values are given for each gas species.
Figure 6. Comparison between TeS-MinG predictions and ANN responses for H₂, CO, CO₂, and CH₄ yields under variations in (a) water-to-biomass molar ratio, (b) pressure, and (c) temperature. Black lines represent TeS-MinG predictions, whereas red markers indicate ANN predictions. RMSE values are given for each gas species.
Preprints 235313 g006aPreprints 235313 g006b
Figure 7. External validation of the ANN against TeS-MinG predictions for H₂, CO, CO₂, and CH₄ yields. The dashed line represents the ideal Y = X relationship, while the color scale indicates the simulation temperature.
Figure 7. External validation of the ANN against TeS-MinG predictions for H₂, CO, CO₂, and CH₄ yields. The dashed line represents the ideal Y = X relationship, while the color scale indicates the simulation temperature.
Preprints 235313 g007
Figure 8. Effect of ((a) water-to-biomass molar ratio, (b) pressure, and (c) temperature on H₂, CO, CO₂, and CH₄ yields predicted by the TeS-MinG and ANN models for independent açaí seed biomass. The solid black line represents TeS-MinG predictions, whereas orange circles and lines represent ANN predictions.
Figure 8. Effect of ((a) water-to-biomass molar ratio, (b) pressure, and (c) temperature on H₂, CO, CO₂, and CH₄ yields predicted by the TeS-MinG and ANN models for independent açaí seed biomass. The solid black line represents TeS-MinG predictions, whereas orange circles and lines represent ANN predictions.
Preprints 235313 g008aPreprints 235313 g008b
Figure 9. ANN-predicted response maps for açaí seed biomass (Sample 4) at constant Pressure of 230 bar as a function of temperature and H₂O/biomass molar ratio: (a) H₂, (b) CO, (c) CO₂, and (d) CH₄.
Figure 9. ANN-predicted response maps for açaí seed biomass (Sample 4) at constant Pressure of 230 bar as a function of temperature and H₂O/biomass molar ratio: (a) H₂, (b) CO, (c) CO₂, and (d) CH₄.
Preprints 235313 g009
Table 1. Proximate and elemental composition, atomic ratios, and higher heating value (HHV) of the biomass samples used for ANN training and external validation.
Table 1. Proximate and elemental composition, atomic ratios, and higher heating value (HHV) of the biomass samples used for ANN training and external validation.
Training Samples External Validation Sample
Analysis Sample 1 (S1) Sample 2 (S2) Sample 3 (S3) Sample 4 (S4)
Proximate analysis (%)
Moisture 28.864 27.987 27.950 28.143
Volatile matter 66.187 67.393 67.140 66.812
Ash 1.220 0.980 1.040 1.201
Fixed carbon 3.729 3.640 3.870 3.844
Elemental analysis
Carbon 36.270 35.78 35.57 35.821
Hydrogen 4.758 4.768 4.719 4.753
Sulfur 0.090 0.081 0.073 0.083
Nitrogen 4.350 1.910 1.85 3.145
Oxygen1 53.312 56.481 56.748 56.783
O/C ratio 1.1035 1.1851 1.1977 1.1888
H/C ratio 1.5631 1.5879 1.5808 1.5923
N/C ratio 0.1028 0.0458 0.0446 0.0752
HHV (MJ/Kg)* 12.673 12.227 12.068 12.168
Pseudocomponents CH₁.₅₆₃₁O₁.₁₀₃₅N₀.₁₀₂₈ CH₁.₅₈₇₉O₁.₁₈₅₁N₀.₀₄₅₈ CH₁.₅₈₀₈O₁.₁₉₇₇N₀.₀₄₄₆ CH₁.₅₉₂₃O₁.₁₈₈₈N₀.₀₇₅₂
Table 2. Predictive performance of ANN configurations with one hidden layer (5, 8, 10, 12, 15 and 20 neurons) for H₂, CO, CO₂, and CH₄ using different training algorithms.
Table 2. Predictive performance of ANN configurations with one hidden layer (5, 8, 10, 12, 15 and 20 neurons) for H₂, CO, CO₂, and CH₄ using different training algorithms.
Hidden layer Training algorithm Training time (s) Prediction time (s) R² H₂ R² CO R² CO₂ R² CH₄ Mean R²
5 trainlm 21.332 0.500569 0.99822 0.98913 0.99749 0.99876 0.99590
5 trainbr 8.5610 0.253355 0.99920 0.98956 0.99882 0.99920 0.99670
5 trainscg 11.524 0.175080 0.99439 0.98025 0.99227 0.99522 0.99053
5 trainbfg 22.140 0.559789 0.99569 0.98158 0.99281 0.99495 0.99126
8 trainlm 30.461 0.632996 0.99969 0.99166 0.99943 0.99952 0.99758
8 trainbr 39.747 0.840742 0.99966 0.99519 0.99920 0.99950 0.99839
8 trainscg 6.8040 0.352583 0.99877 0.98909 0.99785 0.99867 0.99610
8 trainbfg 12.929 0.436136 0.99866 0.99007 0.99806 0.99813 0.99623
10 trainlm 10.589 0.103455 0.99978 0.99212 0.99945 0.99943 0.99770
10 trainbr 12.247 0.057034 0.99989 0.99543 0.99955 0.99963 0.99863
10 trainscg 6.4470 0.122126 0.99872 0.99001 0.99860 0.99785 0.99629
10 trainbfg 12.198 0.761908 0.99815 0.98925 0.99803 0.99724 0.99567
12 trainlm 21.171 0.246793 0.99990 0.99539 0.99959 0.99959 0.99862
12 trainbr 23.742 0.091975 0.99989 0.99585 0.99960 0.99969 0.99876
12 trainscg 12.857 0.431646 0.99924 0.99134 0.99922 0.99843 0.99706
12 trainbfg 8.0220 0.105744 0.99912 0.98933 0.99903 0.99819 0.99642
15 trainlm 23.544 0.723533 0.99992 0.99541 0.99961 0.99968 0.99866
15 trainbr 50.338 0.867572 0.99996 0.99611 0.99963 0.99973 0.99886
15 trainscg 9.4660 0.032820 0.99925 0.99073 0.99893 0.99905 0.99699
15 trainbfg 8.8460 0.079559 0.99889 0.98922 0.99869 0.99864 0.99636
20 trainlm 35.238 0.758156 0.99993 0.99615 0.99958 0.99966 0.99883
20 trainbr 66.177 0.479117 0.99995 0.99716 0.99961 0.99970 0.99911
20 trainscg 35.106 0.984398 0.99953 0.99060 0.99914 0.99928 0.99714
20 trainbfg 11.280 0.070801 0.99946 0.99039 0.99919 0.99934 0.99710
Table 3. Predictive performance of ANN configurations with two hidden layers (8–4, 12–6, 16–8, 20–10, and 24–12 neurons) for H₂, CO, CO₂, and CH₄ using different training algorithms.
Table 3. Predictive performance of ANN configurations with two hidden layers (8–4, 12–6, 16–8, 20–10, and 24–12 neurons) for H₂, CO, CO₂, and CH₄ using different training algorithms.
Hidden layer Training algorithm Training time (s) Prediction time (s) R² H₂ R² CO R² CO₂ R² CH₄ Mean R²
8–4 trainlm 22.985 0.432508 0.99901 0.99198 0.99863 0.99958 0.99703
8–4 trainbr 32.850 0.397486 0.99991 0.99532 0.99958 0.99967 0.99914
8–4 trainscg 21.047 0.758287 0.99651 0.98608 0.99660 0.99407 0.99317
8–4 trainbfg 25.490 0.484113 0.99847 0.99066 0.99804 0.99887 0.99716
12–6* trainlm 17.608 0.561046 0.99995 0.99563 0.99960 0.99971 0.99976
12–6 trainbr 54.589 0.201221 0.99992 0.99825 0.99960 0.99968 0.99986
12–6 trainscg 11.312 0.137523 0.99892 0.99113 0.99890 0.99945 0.99789
12–6 trainbfg 20.196 0.129917 0.99927 0.99100 0.99849 0.99912 0.99795
16–8 trainlm 33.535 0.736446 0.99991 0.99554 0.99962 0.99973 0.99942
16–8 trainbr 91.145 0.283723 0.99993 0.99917 0.99956 0.99969 1.00000
16–8 trainscg 9.447 0.061154 0.99868 0.99083 0.99809 0.99884 0.99725
16–8 trainbfg 13.325 0.157019 0.99884 0.99118 0.99834 0.99914 0.99737
20–10 trainlm 47.041 0.439656 0.99994 0.99720 0.99961 0.99972 0.99964
20–10 trainbr 159.653 0.235410 0.99997 0.99958 0.99964 0.99977 1.00000
20–10 trainscg 14.468 0.283899 0.99901 0.98958 0.99832 0.99928 0.99797
20–10 trainbfg 22.015 0.362116 0.99832 0.98923 0.99762 0.99867 0.99653
24–12 trainlm 148.570 1.039713 0.99996 0.99947 0.99965 0.99976 1.00000
24–12 trainbr 265.681 0.465043 0.99996 0.99974 0.99965 0.9998 1.00000
24–12 trainscg 16.221 0.049170 0.99956 0.99116 0.99922 0.99935 0.99786
24–12 trainbfg 66.012 1.073456 0.99974 0.99167 0.99939 0.99945 0.99865
* Bold values indicate the selected configuration providing the best balance between predictive performance and computational efficiency among the one or two-hidden-layer tested networks.
Table 4. Computational time required for different numbers of TeS-MinG simulations based on the benchmark reported for TeS-Thermo.
Table 4. Computational time required for different numbers of TeS-MinG simulations based on the benchmark reported for TeS-Thermo.
Number of simulations* TeS-MinG (TeS-Thermo) ANN
1,000 30.6 s 0.069 s
10,000 5.1 min 0.075 s
100,000 51 min 0.084 s
1,000,000 8.5 h 0.2317 s
*ANN and TeS-MinG inference time determined using a MacBook Air M2 with 16 GB of RAM.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.