Machine Learning in the Macromolecular Sciences: From Methods to Applications. A Review

1. Introduction

1.1. A brief history of machine learning

The origins of machine learning (ML) are deeply rooted in the broader field of artificial intelligence (AI), which emerged in the 1940s and 1950s with the goal of creating machines capable of simulating human intelligence. Early AI research focused on symbolic reasoning and rule-based systems, epitomized by the work of Alan Turing and his foundational concepts like the Turing Test [1]. By contrast, ML began to take shape as a distinct subfield in the 1950s when Arthur Samuel coined the term while developing a checkers-playing program that improved its performance through experience [2]. Early ML approaches were primarily based on statistical methods, such as linear regression and nearest neighbors, which laid the foundation for modern algorithms [3].

Despite progress, both AI and ML experienced periods of stagnation, known as the "AI winters" of the 1970s and 1990s, caused by limited computational resources and unmet expectations. During these periods, symbolic AI faced challenges in handling complex tasks like image recognition or natural language understanding, while ML quietly advanced with the development of methods like decision trees, support vector machines (SVMs), and neural networks (NNs). The resurgence of ML in the 2000s was driven by the confluence of abundant data, increased computational power (particularly GPUs), and breakthroughs in algorithms, particularly deep learning and reinforcement learning. Today, ML underpins many transformative technologies, including natural language processing, autonomous vehicles, and personalized recommendations.

While ML is a subset of AI, the two fields differ in focus and approach. AI encompasses the broader aim of replicating human intelligence, including techniques like symbolic reasoning and robotics, which do not necessarily rely on data-driven learning. In contrast, ML focuses specifically on enabling systems to learn patterns and make predictions from data. By leveraging supervised, unsupervised, and reinforcement learning [4], ML has become the engine driving most modern AI applications, making it a cornerstone of advancements in technology and science.

Reflecting on our early works, particularly the 1996 articles titled "Optimization of classification trees: strategy and algorithm improvement" and "A novel algorithm to optimize classification trees" [5,6], I am motivated to write this review. In that papers, we explored the application of classification and regression trees (CART) in data analysis, highlighting their effectiveness in handling complex datasets without relying on parametric assumptions. The study demonstrated how CART could identify significant variables in medical diagnosis, providing a robust alternative to traditional statistical methods.

1.2. Why another review?

The rapid advancements in ML have revolutionized numerous scientific domains, including chemistry, biology, and materials science. While there are already several comprehensive reviews on ML in materials science, the unique challenges and opportunities presented by polymers and macromolecules necessitate a focused and timely discussion. Polymers exhibit unparalleled complexity due to their diverse architectures, multiscale behavior, and intricate structure-property relationships. This review addresses these distinct aspects, providing insights into how ML is tailored to overcome such challenges in polymer science. Moreover, the transformative potential of ML has been highlighted by the 2024 Nobel prizes in Physics, emphasizing the profound influence of data-driven approaches in advancing scientific discovery. These accolades underscore the need for domain-specific reviews that guide researchers through the nuanced applications of ML, particularly in emerging fields like macromolecular science.

A significant feature of this review is its focus on recent developments, providing an up-to-date synthesis of the latest advancements in ML techniques and their applications to polymers. This includes breakthroughs in predictive modeling, high-throughput screening, and multi-objective optimization, as well as insights into techniques that extend beyond conventional methodologies. Existing reviews have dealt with emerging trends in ML [7], ML challenges, progress, and potential in polymer science [8], ML interpretability methods [9], methods and applications of ML for materials design [10], ML towards multiscale soft materials design [11], or trends in extreme learning machines [12], to mention a few. For those who prefer to read a more extensive introduction to the various ML methods for polymer informatics, see [13]. The present review adopts a dual perspective by classifying contributions both by methods and techniques (e.g., deep learning, Bayesian optimization, graph NNs) and by application types (e.g., mechanical properties, thermal stability, biodegradability). This structured approach offers researchers a clear framework for understanding how different ML approaches align with specific goals in polymer design and engineering. Through this focused and methodical exploration, this review aims to bridge gaps in the literature, offer fresh perspectives, and inspire innovative research at the intersection of ML and polymer science.

Figure 1. Cumulative number of publications in the course of time belonging to the five areas: Supervised learning (Section 2.1), Unsupervised learning (Section 2.2), Reinforcement learning (Section 2.3), Generative models (Section 2.4), and Deep learning (Section 2.5).

Figure 1. Cumulative number of publications in the course of time belonging to the five areas: Supervised learning (Section 2.1), Unsupervised learning (Section 2.2), Reinforcement learning (Section 2.3), Generative models (Section 2.4), and Deep learning (Section 2.5).

1.3. Outline

ML has redefined scientific inquiry by providing tools to analyze complex data and predict outcomes with precision. In polymer science, where material performance often depends on subtle molecular variations, ML offers unparalleled opportunities. Early studies focused on property prediction using statistical models, evolving into deep learning applications for polymer informatics. Emerging trends include the integration of generative AI for creating novel polymer architectures. Macromolecular and polymeric sciences have long been at the forefront of material innovation, contributing significantly to industries such as automotive, healthcare, and packaging. However, the field faces challenges that require increasingly complex and high-throughput experimental workflows. The rise of ML offers solutions by enabling predictive modeling, automated optimization, and data-driven decision-making. Traditional experimental approaches are inherently resource-intensive and time-consuming. For example, the iterative process of synthesizing and testing polymer candidates can take weeks or months. ML addresses these limitations by leveraging computational algorithms to analyze large datasets, uncover structure-property relationships, and optimize synthesis pathways.

This review provides a comprehensive exploration of ML applications in polymer science, focusing on: (i) Core ML methodologies, including supervised, unsupervised, reinforcement, generative and deep learning. (ii) Applications in property prediction, synthesis optimization, and sustainable materials design. (iii) Challenges and limitations, such as data scarcity and interpretability. (iv) Future directions. For each of the methods we have created a schematic drawing along with a short description of its distinctive feature. For those readers primarily interested in a specific application rather than a method, we prepared Table 1. While this review attempts to present and discuss the primary ML methods of relevance for polymer research, additional methods and variations (such as principle component analysis, domain adaption, multitask learning, sample selection bias, covariate shift, semi-supervised and self-supervised learning, anomaly detection, feature learning, sparse dictionary learning, robot learning, association rule learning etc.) certainly exist, and their amount keeps growing on a monthly basis. We suspect that the diversity of machine learning "methods" is nearly matched by the multitude of programs developed to tackle problems that could often be addressed using variations, combinations, or adaptations of existing approaches. Thats why we focus on the core ML methodolgies.

2. Current Machine Learning Strategies in Polymer Science

ML encompasses a diverse array of algorithms and approaches, each suited to specific tasks. Here we provide a detailed overview of the primary ML techniques used in the field along with typical applications.

2.1. Supervised Learning

Supervised learning (Figure 2) is the most widely used ML technique in polymer science. It involves training a model on labeled data, where the relationship between input features (e.g., polymer composition) and output targets (e.g., tensile strength) is known [14]. Kotsiantis [14] reviewed various supervised ML classification techniques, emphasizing their role in building models that generalize from labeled data to predict future instances. The paper covers key algorithms, such as decision trees, SVMs, and ensemble methods, and discusses their theoretical foundations and practical applications. The study underscores the importance of feature selection and bias-variance trade-offs in optimizing classifier performance, providing a foundational guide for researchers exploring supervised learning methods.

(A) Neural Networks (NNs), particularly convolutional NNs (CNNs), which will be discussed in detail in Section 2.5, excel in analyzing polymer microstructures. The principle behind the learning feature of a NN (Figure 4) is iterative optimization. NNs learn by adjusting their internal parameters (weights and biases) to minimize the error between predicted outputs and actual targets. This process involves forward propagation, where data passes layer by layer through the network, and backpropagation, which computes gradients of the loss function with respect to network parameters using the chain rule of calculus. These gradients guide an optimizer in updating the parameters over multiple epochs, enabling the network to learn general patterns in the data while avoiding overfitting.

Recent contributions demonstrate the versatility of NNs across various applications in polymer science. In polymer property prediction, Saingam et al. [15] used NNs to predict the compressive strength of hemp fiber-reinforced polymer composites, showing strong agreement with experimental results. Alhulaybi and Otaru [16] utilized deep NNs (DNNs) to model the thermal decomposition behavior of biodegradable composites, achieving near-perfect correlation with experimental data. Similarly, Malashin et al. [17] optimized NNs and SVMs for predicting textile polymer composite properties using multi-objective optimization techniques. In polymer processing and structural analysis, Yang et al. [18] employed multitask NNs to predict gas permeability and selectivity in polymer chemistry, leading to novel membrane designs. Cassola et al. [19] reviewed ML’s integration with Computer Aided Engineering tools for efficient simulation of polymer composite processes. Bejagam et al. [20] introduced a novel ANN framework for force-field optimization in coarse-grained molecular dynamics (CG-MD), bridging simulation data with experimental properties. NNs have also found utility in advanced material design. Axelrod et al. [21] explored their integration with atomistic simulations, advancing the design of materials like polymer electrolytes. Alesadi et al. [22] leveraged NNs to predict the glass transition temperature of conjugated polymers, utilizing molecular dynamics and chemical structure features. Wu et al. [23] introduce a ML-enhanced computational reverse-engineering analysis method for analyzing small-angle scattering profiles of polymer solutions containing semiflexible fibrils with variable diameters, such as methylcellulose fibrils. By incorporating neural networks (NN), the improved method enhances workflow speed while maintaining accuracy. Validation with in silico scattering profiles confirms its precision in determining fibril dimensions. Application to experimental data from methylcellulose fibrils yields diameter distributions consistent with previous analytical model fits (15–20 nm). Several reviews highlight broader implications of NNs in polymer science. McDonald et al. [24] emphasize their role in accelerating biomaterial design, addressing challenges in data standardization and characterization. Malashin et al. [25,26] explore NN techniques in 3D/4D printing and LSTM applications for polymer property prediction. Loh et al. [27] discuss the integration of NNs with traditional methods for modeling damage in fiber-reinforced composites. These studies underscore the transformative potential of NNs in polymer science, from property prediction and material design to advanced processing and simulation.

(B) Random Classification Tree is a type of decision tree used for categorizing data into distinct classes (Figure 5). These parameter-free, supervised learning algorithms model data by recursively splitting it into subsets based on feature values, following a tree-like structure. At each node, a decision partitions the data into groups, ultimately assigning a classification label at the leaf nodes. The algorithm evaluates all possible splits at each node, aiming to maximize class separation. Known for their clarity and interpretability, classification trees represent each decision as a simple rule, making them easy to visualize and understand [14].

Fakhry et al. [28] employed decision tree models to enhance the prediction of Cr(VI) sorption by fungal biomass and extracellular melanin. By integrating equilibrium isotherms and kinetic studies, they showcased the efficacy of decision trees in biosorption predictions, complementing traditional methods. Saber et al. [29] utilized a decision tree learning algorithm alongside Taguchi’s optimization approach to enhance the biosynthesis of pullulan, a biodegradable hydrogel polymer. This method identified key process variables, reducing sucrose usage by 33% while maintaining high yield, achieving 7.23% pullulan under optimal conditions, and highlighting AI-driven methods’ potential in fermentation optimization. Koinig et al. [30] applied classification methods to improve the sorting of monolayer and multilayer films in recycling. By combining near-infrared spectroscopy with Fourier transforms to optimize spectral data, they significantly enhanced classification accuracy for material separation. Bhowmik et al. [31] use decision tree and principal component analysis, to predict polymers’ specific heat at room temperature, providing insights into the influence of polymer descriptors and guiding the design of novel materials with tailored thermal properties.

(C) Random Forests are a popular parameter-free ML algorithm used for classification and regression tasks (Figure 6). This ensemble learning method constructs multiple decision trees and combines their outputs to enhance performance, stability, and robustness compared to individual trees. While individual decision trees are interpretable, Random Forests sacrifice some interpretability due to their ensemble nature.

Several studies have demonstrated the versatility of Random Forests in addressing diverse challenges across polymer sciences. In the context of material sustainability, Guarda et al. [32] reviewed the application of ML algorithms, including Random Forests, to improve plastics’ life cycle sustainability, addressing challenges from production to end-of-life processes. Similarly, Nelon et al. [33] emphasized Random Forests’ predictive capabilities for damage mechanism detection and life estimation in fiber-reinforced composite materials when combined with structural health monitoring and nondestructive evaluation methods. Random Forests have proven effective in predicting material properties and optimizing processes. Munir et al. [34] employed RFE-RF models to monitor molecular weight degradation in PLA during extrusion, providing interpretable and robust predictions across varying process conditions. Similarly, Pelzer et al. [35] leveraged Random Forests to analyze process parameter dependencies in extrusion-based additive manufacturing, identifying optimal settings and key parameter interactions. Chavez-Angel et al. [36] applied Random Forests in conjunction with k-nearest neighbor algorithms to generate hyperspectral temperature maps of polymer thin films, enabling precise heat distribution analysis for applications in CMOS technologies. Optimization and anomaly detection also benefit from Random Forest applications. Huang et al. [37] integrated a mixed whale optimization algorithm (MWOA) with Random Forests to predict the chloride permeability coefficient of rubber concrete, achieving a 62.9% improvement in prediction accuracy. Gope et al. [38] used Random Forests to detect abnormal processing parameters in polypropylene fiber melt spinning, enhancing parameter identification and quality control. In structural and mechanical property prediction, Random Forests have shown exceptional accuracy. Amin et al. [39] predicted the bond strength of FRP laminates with concrete ($R^2=0.82$), while Anjum et al. [40] achieved an $R^2$ of 0.91 for compressive strength prediction of fiber-reinforced nano-silica concrete. Alabdullh et al. [41] proposed a hybrid ensemble (HENS) model for FRP bond strength prediction, where Random Forest served as a baseline, achieving robust results before HENS demonstrated superior accuracy ($R^2=0.9663$). Random Forests have also been applied in composite material performance prediction. Karamov et al. [42] predicted fracture toughness of pultruded composites, correlating strongly with bending and tension properties. Similarly, Joo et al. [43] used Random Forests to predict physical properties of polypropylene composites, supporting recipe optimization for industrial applications. Aminabadi et al. [44] integrated Random Forests in an Industry compliant injection molding setup, effectively predicting part geometry and surface quality for automated quality control. Lastly, Random Forests have proven invaluable in concrete reinforcement studies. Khan et al. [45] applied Random Forests to predict the compressive strength of steel-fiber-reinforced concrete (SFRC), achieving an $R^2=0.96$ and identifying cement content as the most influential factor. Another study by Khan et al. [46] demonstrated Random Forests’ reliable performance ($R=0.97$) in predicting the flexural strength of FRP-reinforced beams, using sensitivity analysis to reveal critical influencing factors.

(D) Gradient Boosting is a powerful ML technique widely used for regression, classification, and ranking tasks. It constructs an ensemble of weak learners (typically decision trees) sequentially, where each model corrects the residual errors of its predecessors (Figure 7). By iteratively optimizing a loss function, Gradient Boosting achieves high predictive accuracy and robustness, making it an effective tool across various domains.

For property prediction, Ascencio-Medina et al. [47] used Gradient Boosting Regressor models to predict dielectric permittivity in polymers, achieving high $R^2$ coefficients and providing molecular insights through Accumulated Local Effect analysis. Fatriansyah et al. [48] investigated extreme gradient boosting (XGBoost) for predicting the glass transition temperature of polymers, demonstrating stable performance ($R^2=0.774$) while emphasizing SMILES descriptor optimization for improved reliability. Yan et al. [49] employed Gradient Boosting Regression to predict tribological properties of PTFE composites, effectively capturing the influence of load and speed on wear and friction rates. In composite materials, Zhao et al. [50] utilized an extreme gradient boosting model to predict the residual compressive strength of composite laminates post-impact, achieving exceptional accuracy ($R^2=0.9910$) and showcasing acoustic emission parameters’ utility for structural health monitoring. Similarly, Hu et al. [51] integrated XGBoost and Gradient Boosting to optimize the stacking sequence and orientation of CFRP/metal composite laminates, achieving reliable predictions for tensile and bending strengths. Gradient Boosting has also advanced sustainable material design and process optimization. Cao and Xu [52] employed Gradient Boosting Decision Trees to predict photovoltaic performance in ternary polymer solar cells, achieving a relative error of less than 5%. Okada et al. [53] combined Gradient Boosting with feature selection to analyze molecular descriptors and TD-NMR data, enabling accurate hydrophilic polymer coating predictions for water treatment membranes. Mayorova et al. [54] applied Gradient Boosting alongside Random Forest and Raman spectroscopy to study structural changes in whey protein isolate-hyaluronic acid complexes, identifying key Raman bands indicative of modifications. Gradient Boosting has proven effective in manufacturing and machining applications. Biruk-Urban et al. [55] employed Gradient Boosting Regressor to simulate drilling parameters in GFRP composites, identifying feed per tooth as the most influential factor on delamination and optimizing machining processes. Rahimi et al. [56] used Gradient Boosting to classify moisture content in kiln-dried wood, aiding in moisture uniformity optimization and effectively grouping timber into acceptable, over-, and under-dried categories. In polymer-based systems, Deshpande et al. [57] achieved high prediction accuracy ($R^2=0.97$) for the specific wear rate of glass-filled PTFE composites using Gradient Boosting, identifying key parameters via Pearson’s correlation. Chen et al. [58] utilized extreme gradient boosting to analyze molecular dynamics simulations of natural rubber, identifying hydrogen-bond interactions as critical to stress-strain behavior and strain-induced crystallization. Martinez et al. [59] leveraged Bayesian Additive Regression Trees (BART) and Gradient Boosting to predict dispersancy efficiency in oil and lubricant additives, enhancing discovery through interactive tools for chemical library screening. Ethier et al. [60] use ML to predict polymer-solvent phase behavior, utilizing a curated database of 21 polymers, 61 solvents, and 6524 cloud point temperatures. Models incorporating molecular descriptors and Hansen solubility parameters achieved high accuracy in cloud point predictions and replicated phase diagrams. While extrapolation to dissimilar polymers is limited, minimal additional data improves predictions.

(E) Support Vector Machines (SVMs) are particularly effective for classification tasks, aiming to find the hyperplane that best separates data points of different classes in a high-dimensional space. By maximizing the margin between the closest points (support vectors) of each class, SVMs achieve robust classification performance (Figure 8). Through the use of kernel functions, they can handle non-linear separations by mapping data to higher-dimensional spaces where linear separation becomes possible.

In non-destructive testing, Ashebir et al. [61] reviewed advanced techniques for defect detection in fiber-reinforced thermoplastic composites. Integrating ML algorithms, including SVMs, improved defect detection sensitivity by up to 30%, enhancing the structural integrity of additively manufactured components. Liu et al. [62] used support vector regression to analyze 3D printing parameters’ influence on outcomes, identifying critical factors such as extrusion expansion ratio, elastic modulus, and elongation at break through feature importance and SHapley Additive exPlanations (SHAP) values. In process optimization, Malashin et al. [63] combined finite element modeling with ML methods, including SVMs, to predict tensile properties of selective laser-sintered polyamide components under various loading conditions. Similarly, Subeshan et al. [64] highlighted the potential of SVMs in optimizing electrospinning parameters to produce nanofibers with precise properties for applications ranging from tissue engineering to sensors. Chen et al. [65] used SVMs to predict piezoelectric output from material and synthesis parameters in electrospun PVDF and composite fibers, advancing high-performance energy-harvesting systems. SVMs also play a critical role in materials design and property prediction. Yue et al. [66] integrated high-throughput stochastic breakdown simulation with SVMs to predict energy storage performance in polymer composites, providing valuable insights for designing high-density capacitors. Shen et al. [67] combined phase-field modeling with ML, including SVMs, to study breakdown mechanisms in polymer-based dielectrics. This approach yielded predictive models for breakdown strength and identified optimal nanofiller properties, verified through experiments, contributing to the design of enhanced dielectric materials. In broader polymer applications, Jongyingcharoen et al. [68] applied ML models, including SVMs, to classify crosslink density in para rubber gloves using near-infrared spectroscopy, achieving reliable classification for production screening. Qin et al. [69] developed an AI-assisted approach for analyzing bicomponent fibers, utilizing SVMs for accurate and cost-effective component identification in industrial testing. Uddin and Fan [70] used SVMs within an interpretable ML framework to predict the glass transition temperature of polymers, combining feature elimination and SHAP analysis for enhanced interpretability.

2.2. Unsupervised Learning

Unsupervised learning techniques are essential for exploring large, unlabeled datasets (Figure 9). They enable researchers to uncover patterns and groupings that may not be immediately apparent.

(A) Clustering Algorithms, such as k-means and hierarchical clustering (Figure 10), are widely used in polymer informatics for tasks like classifying polymers by molecular weight distributions or chemical compositions [71]. These methods are essential for exploratory data analysis, uncovering hidden patterns in datasets, and aiding in hypothesis generation [13].

Clustering plays a critical role in recycling and material optimization. Olivieri et al. [72] developed a classification system for recycled polymers using industrial sorting data, setting the stage for AI-driven databases to optimize recycling processes and enhance material properties. Wolf and Stammer [73] examined chemical recycling strategies for silicone polymers, emphasizing depolymerization techniques and the growing potential for closed-loop systems to minimize waste and conserve resources. Neo et al. [74] employed Random Forest Regression and Bagging to predict offline color data of extruded thermoplastic resins, significantly reducing time and material waste in color-matching processes. In polymer mechanics and composite design, Dairabayeva et al. [75] investigated the mechanical performance of continuous carbon-fiber reinforced composites, analyzing tensile and flexural properties alongside cost and printing time optimizations for additive manufacturing. Bounjoum et al. [76] used clustering techniques to identify distinct failure patterns in CFRP-reinforced concrete structures, revealing a 45% improvement in flexural strength and informing composite-reinforced infrastructure designs. Lee et al. [77] studied epoxy resin and crumb rubber powder-modified cement asphalt mortar, showcasing improved flowability, stability, and environmental resistance with applications in railway infrastructure durability. Unsupervised learning has also been used for novel analytical and structural insights. Scatigno et al. [78] applied clustering techniques to classify spectroscopic benchmarks in Italian pictorialism photography, identifying distinct molecular and elemental fingerprints to uncover stylistic practices. Younes et al. [79] used Principal Component Analysis (PCA) to study lignin degradation in tropical peatlands, linking molecular fingerprints to carbon and methane emission dynamics in anoxic environments. Clustering algorithms also enable molecular and mesoscopic structural analysis. Doi et al. [80] introduced the ML-aided Local Structure Analyzer (ML-LSA) for classifying mesoscopic structures of liquid crystal polymers (LCPs) using molecular dynamics simulations. This approach identified nematic- and smectic-like structures with high accuracy, revealing structural transitions without predefined order parameters. Similarly, Bejagam et al. [81] integrated nonmetric multidimensional scaling (NMDS) with CG-MD simulations to study the coil-to-globule transition in poly(N-isopropylacrylamide) (PNIPAM). [82] Their analysis uncovered multiple metastable states during the transition, providing mechanistic insights into polymer behavior above its lower critical solution temperature (LCST).

(B) Dimensionality Reduction techniques, such as principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE), simplify high-dimensional datasets, enabling more efficient analysis and visualization (Figure 11). For instance, [83] applied PCA to visualize spectroscopic data, revealing correlations between polymer structure and functionality. These methods are particularly valuable in polymer science for managing complex datasets and uncovering key structure-property relationships.

In materials discovery and polymer design, dimensionality reduction plays a crucial role. Zhao et al. [84] emphasized its use in managing large datasets, improving property prediction models, and accelerating polymer discovery. Martinez et al. [85] introduced a pipeline that leverages ML and dimensionality reduction to predict molecular weight, polydispersity index, and conversion rates in polymer synthesis, achieving $R^2$ values up to 0.93. Patra [86] highlighted its potential in navigating the vast configurational and chemical space of polymers, aiding in the identification of key structure-property relationships for data-driven design. Dimensionality reduction also enhances simulation and experimental workflows. Karatrantos et al. [87] explored molecular simulations, emphasizing ML’s role in dimensionality reduction to study dynamic bond exchange reactions in vitrimers and covalent adaptable networks. Trienens et al. [88] used ML-assisted simulation models to predict melt quality during polymer extrusion, correlating process parameters with thermal and material homogeneity. Similarly, Knoll and Heim [89] proposed a "machine fingerprint" approach to reduce disparities between simulation and experimental injection molding results, enhancing transfer learning applications. For soft materials and nanoscale systems, dimensionality reduction has proven transformative. Noid [90] and Dhamankar and Webb [91] reviewed coarse-grained (CG) modeling techniques, highlighting the use of dimensionality reduction to enhance chemically specific simulations and predictive capabilities. Jacob et al. [92] developed an analytical model for FRET in flexible peptides, simplifying diffusion analysis through fluorescence spectroscopy. Jackson et al. [11] integrated ML with multiscale modeling frameworks, emphasizing dimensionality reduction to distill critical chemical and morphological features for inverse materials design. Ziolek et al. [93] reveal the detailed distribution of polymer conformations within Pluronic and Tetronic micelles, highlighting the structural organization in the disordered micellar corona. Their analysis demonstrates how the block architecture of these amphiphilic copolymers influences chain arrangement in core-shell morphologies, providing new insights into their nanoscale structure and dynamics. Dimensionality reduction also supports optimization in advanced material systems. Saleh et al. [94] optimized TPMS lattice structures using adaptive neuro-fuzzy inference systems, achieving superior accuracy in mechanical property predictions. Yan et al. [95] applied response surface methodology (RSM) and central composite design (CCD) to optimize the preparation of CNT-GN nanocomposites, achieving high sensitivity and modulus while minimizing experimental costs. Bessa et al. [96] employed dimensionality reduction in computational frameworks for uncertainty modeling, mitigating the "curse of dimensionality" and enhancing design space exploration. Lastly, dimensionality reduction facilitates property optimization in polymer dielectrics and energy systems. Wang et al. [97] reviewed quantum mechanical and molecular dynamics strategies, highlighting the role of QSPR and ML in dielectric property optimization. Li et al. applied PCA and t-SNE to study molecular orientation tensors, uncovering mechanisms of shear thinning in squalane under high strain rates.

2.3. Reinforcement Learning

Reinforcement learning (RL) is a promising but underexplored area in polymer science. It involves training an agent to make sequential decisions by maximizing a reward signal (Figure 12). By autonomously exploring large decision spaces, RL has shown potential for optimizing polymerization pathways [98] and advancing data-driven approaches in material design.

The application of RL and related methodologies in polymer science is gaining attention. Gomez-Flores et al. [99] critically reviewed the role of ML in fiber and nanocomposites, emphasizing the underutilized potential of RL. They proposed a control loop framework to enhance process integration across scales, enabling more efficient and adaptive design strategies. Wang et al. [100] demonstrated the potential of RL-inspired Bayesian optimization (BO) integrated with CG-MD to design solid polymer electrolytes (SPEs) with enhanced ionic conductivity. Their framework autonomously explored high-dimensional design spaces using physically interpretable descriptors, uncovering molecular-level insights into how properties such as molecule size and interaction strength affect conductivity. This approach provided actionable guidance for improving SPE components and showcased RL’s utility in materials optimization. RL-related techniques have also been applied to industrial polymer processing. Malashin et al. [101] integrated multi-objective evolutionary algorithms with data mining to optimize extrusion barrier screw geometry for processing low-density polyethylene and polypropylene. This study highlights the role of AI in refining complex industrial processes and improving material efficiency. Jeon et al. [102] developed ML models incorporating energy flow-related features to predict melt temperature during injection molding, providing insights into process control and optimization.

2.4. Generative Models

Generative models, such as generative adversarial networks (GANs) and variational autoencoders (VAEs), are revolutionizing polymer design by enabling the creation of novel polymer structures and architectures (Figure 13). GANs, for instance, have been used to generate realistic polymer architectures, opening new possibilities for material innovation [103]. These models excel in exploring vast design spaces, accelerating discovery, and optimizing material properties.

Fan et al. [104] reviewed smart-responsive hydrogels for skin wound therapy, highlighting the integration of AI to optimize material properties for improved clinical outcomes. Adetunji and Erasmus [105] discussed green synthesis techniques for bioplastics derived from microalgae, emphasizing AI’s role in optimizing sustainable production. Zhu et al. [106] reviewed the application of generative models in the design of polymer-based dielectrics, focusing on inverse design methods to establish structure-property linkages. Similarly, Zhou et al. [10] emphasized the transformative role of big data and ML in materials design, showcasing how generative models facilitate large-scale screening and rational design of advanced polymers and porous materials. Generative models have also enhanced molecular simulations and multiscale modeling by bridging scales and uncovering complex structural dynamics. Li et al. [107] employed conditional GANs (cGANs) for backmapping CG macromolecules to atomistic resolutions, achieving high fidelity in modeling cis-1,4 polyisoprene melts. Pan et al. [108] applied dissipation particle dynamics simulations to study lipid-nanoparticle mixtures, providing insights into interfacial tension and mechanical properties for molecular-level composite design. Guda et al. [109] used ML-driven approaches to analyze X-ray absorption near-edge structure (XANES) spectra, revealing active site dynamics in catalytic systems and demonstrating the ability of generative models to uncover structural features in complex materials. In soft material and composite design, Chiu et al. [110] combined a Genetic Algorithm (GA) with a Conditional Variational Autoencoder (CVAE) to design bioinspired composite materials with optimized stiffness and toughness. Their approach was validated through finite element simulations and 3D-printed tensile tests, showcasing generative models’ efficiency in soft material placement optimization. Wang et al. [111] developed a polyurethane-based triboelectric nanogenerator incorporating zirconium boride, demonstrating AI’s role in improving thermal conductivity and electrical output for wearable electronics. Tan et al. [112] introduced a conductive hydrogel system for targeted drug delivery, leveraging generative design to enhance therapeutic outcomes and material performance. Generative models are also driving advancements in high-throughput material discovery. Tao et al. [113] utilized DNNs to discover over 65,000 high-temperature polymer candidates ($T_g>200°\mathrm{C}$) through high-throughput screening. Zhang et al. [114] investigated fluorine-containing benzoxazine and epoxy co-curing, achieving low dielectric constants and high thermal stability, showcasing applications in integrated circuits. Patra et al. [115] introduced a NN-biased Genetic Algorithm (NBGA) to optimize material properties, combining the strengths of genetic algorithms and NNs for accelerated discovery in soft material design. Generative models are reshaping how polymers and materials are developed, enabling efficient exploration of design spaces, optimizing properties, and uncovering structure-property relationships. By integrating these models with domain-specific knowledge and experimental validation, researchers can accelerate the discovery of next-generation materials and expand the possibilities of polymer science and engineering.

2.5. Deep Learning

Deep Learning represents a subset of ML characterized by the use of multi-layered artificial NNs that automatically learn complex patterns from data (Figure 14). Unlike traditional supervised or unsupervised learning methods, which often rely on manual feature engineering or simpler models, deep learning leverages hierarchical representations to extract features directly from raw data, enabling breakthroughs in tasks such as image recognition, natural language processing, and generative modeling.

While the earlier sections focus on more traditional ML approaches, such as supervised methods (e.g., Random Forests, SVMs) and unsupervised techniques (e.g., clustering, dimensionality reduction), deep learning introduces highly flexible architectures, such as CNNs, Generative Adversarial Networks, and Transformer models. These methods excel in handling large-scale datasets and unstructured data (e.g., images, text) and have applications that often blur the boundaries between supervised, unsupervised, and reinforcement learning paradigms.

(A) Transfer Learning is a ML approach where a model trained on one task is adapted to perform a related but distinct task [116]. By leveraging knowledge from large datasets—commonly used in fields like image classification or natural language processing—transfer learning significantly reduces computational and data requirements for training models on specialized tasks (Figure 15). Pre-trained models, such as those based on ImageNet or language models like BERT, can be fine-tuned for domain-specific applications, making this method highly effective in scenarios with limited labeled data.

In polymer science, transfer learning has emerged as a powerful tool for improving material and process predictions with minimal experimental data. Lockner et al. [117] employed transfer learning to optimize machine parameters for injection molding across different polymer materials. By reusing prior process data, the study achieved improved model accuracy and quality even with small datasets. Similarly, Zhang et al. [118] developed a novel transfer learning framework combined with optimal transport to predict stress-strain curves for fiber-reinforced polymer composites, demonstrating accurate performance with limited experimental inputs. In the domain of material property prediction, Lee et al. [119] utilized transfer learning with GNNs to model the optoelectronic properties of conjugated oligomers. By pretraining on short oligomer datasets, the study achieved high accuracy in predicting excited-state energies despite data scarcity. Hu et al. [120] introduced a machine-learning-assisted materials genome approach that integrated graph convolutional networks, transfer learning, and classical gel theory. This approach enabled rapid high-throughput screening and experimental validation of highly tough thermosetting polymers, significantly accelerating the design cycle. Transfer learning has also been applied to bridge the gap between simulated and real-world performance. Huang et al. [121] applied transfer learning with backpropagation NNs (BPNNs) to predict shrinkage and warpage characteristics of injection-molded products. By combining computer-aided engineering (CAE) data with real-world machine performance, the study improved prediction reliability and demonstrated the practical utility of transfer learning in industrial polymer processing.

These studies underscore the versatility of transfer learning in polymer science, enabling accurate predictions for mechanical, thermal, and optoelectronic properties. By reducing data and computational demands, transfer learning is particularly suited for data-scarce domains, supporting adaptive process control and high-throughput material design. The integration of transfer learning with techniques like SHAP analysis further enhances feature reliability and interpretability.

(B) Convolutional NNs (CNNs) are a specialized type of NN designed to process data with a grid-like topology, such as images. CNNs are particularly effective in tasks involving spatial data due to their ability to capture spatial hierarchies and features (Figure 16). By applying convolution operations with sliding filters (kernels) over the input, CNNs compute feature maps that highlight essential patterns in the data [122].

CNNs have been extensively applied to polymer science and engineering, particularly in defect detection, process optimization, and structure-property predictions. Liu et al. [122] demonstrated the use of CNNs for defect detection and impact dynamics in fiber composite materials, highlighting their ability to extract mechanical performance information from images. Similarly, Sepasdar et al. [123] employed CNNs in a physics-informed deep learning framework to predict full-field nonlinear stress distributions and failure patterns in composite materials, achieving an impressive 90% accuracy using synthetically generated microstructural data. Applications of CNNs in process optimization have also been notable. Schmid et al. [124] developed ML models incorporating CNNs to predict key parameters in the plasticizing process during injection molding, achieving a mean absolute error of 0.27% and validating the model in real-machine experiments. Gim and Rhee [125] utilized NN-based interpretations, including CNNs, to analyze cavity pressure profiles in injection molding processes, linking process state points to part quality and demonstrating the industrial potential of CNNs. In structure-property prediction, Kojima et al. [126] applied CNNs to analyze microstructure-property relationships in filled rubber, training models to identify stress-contributing filler morphologies. Their results were validated using CG-MD simulations, providing insights into the mechanics of filled rubber systems. Rahman et al. [127] used a CNN model trained on molecular dynamics simulation data to predict the interfacial shear strength of carbon nanotube-polymer interfaces, with applications in lightweight aerospace materials. CNNs have also facilitated material behavior predictions. Chen et al. [128] employed a chemical language processing model utilizing SMILES embeddings and recurrent NNs to predict the glass transition temperature ($T_g$) of polymers. While not a CNN-specific application, their approach aligns with CNN methodologies in leveraging sequential data for property prediction, achieving high-throughput screening potential. Sepasdar et al. [123] further illustrated CNNs’ role in predicting nonlinear material behaviors, emphasizing their adaptability in synthetic data scenarios. Yan et al. [129] addressed challenges in applying ML to thermoset shape memory polymers (TSMPs) by introducing methodologies like BigSMILES-based fingerprinting, mixed-dimension inputs, and a dual-convolutional model framework. The ML model, trained on a small dataset with supplemented approximations, predicted recovery stresses and identified 14 novel TSMPs with superior recovery stress. Validation through synthesis, testing, and molecular dynamics simulations confirmed the approach’s effectiveness, showcasing its potential for discovering TSMPs with tailored properties.

(C) Generative Adversarial Networks (GANs) are a class of deep learning models introduced by Goodfellow et al. [130] that consist of two NNs—a generator and a discriminator—trained simultaneously in a competitive setting (Figure 17). The generator aims to produce synthetic data that mimics the real data distribution, while the discriminator evaluates whether a given data sample is real or generated. This adversarial process drives both networks to improve iteratively, resulting in highly realistic synthetic data.

GANs have been widely applied in tasks such as image synthesis, style transfer, and data augmentation, as well as in generating high-quality samples for domains like audio and video processing.

(D) Transformer Architectures, introduced by Vaswani et al. [131], represent a significant breakthrough in deep learning, particularly for sequence-to-sequence tasks in natural language processing (NLP). Unlike traditional recurrent or convolutional architectures, Transformers rely on a self-attention mechanism, which dynamically weighs the importance of different input tokens (Figure 18). This design enables efficient parallel processing of data and better handling of long-range dependencies, making Transformers particularly effective for tasks involving complex relationships. Transformers have become the backbone of state-of-the-art models such as BERT, GPT, and T5, revolutionizing applications like machine translation, text summarization, and language modeling. Their flexibility has extended beyond NLP to domains like computer vision and protein structure prediction, showcasing their broad applicability across diverse fields.

In polymer science, Transformer architectures are emerging as powerful tools for predictive modeling and material characterization. Lee et al. [132] applied a transformer-based model to predict mechanical properties of polymer matrix composites, achieving superior performance in modeling complex material systems. For delamination detection, Liu et al. [133] employed a transformer-based NN combined with enhanced terahertz time-domain spectroscopy (THz-TDS) techniques to analyze quartz fiber-reinforced polymer structures. This approach enabled accurate and automated characterization, addressing challenges in non-destructive testing. Additionally, Han et al. [134] introduced a multimodal transformer framework for property prediction in polymers, effectively integrating diverse data types—such as molecular structures, processing parameters, and experimental properties—to improve predictive accuracy. This study highlights the versatility of Transformers in handling heterogeneous data and providing insights into polymer design.

(E) Few-shot and Zero-shot Learning are advanced paradigms in machine learning designed to address the challenge of training models with limited or no labeled data for a target task [135]. Few-shot learning enables models to generalize from a small number of labeled examples by leveraging prior knowledge obtained through large-scale pretraining or related tasks (Figure 19). Zero-shot learning extends this capability by allowing models to perform tasks they have not encountered during training, relying on semantic or contextual information provided by embeddings or pre-trained language models. These approaches have gained prominence with the development of models like GPT and CLIP, which utilize pre-trained representations to generalize across tasks. They are particularly valuable in scenarios where labeled data is scarce or expensive to obtain, making them powerful tools for text classification, image recognition, and natural language understanding.

Few-shot learning has also found impactful applications in materials science, including addressing the challenges of data scarcity in polymer research. Chen et al. [136] reviewed advancements in few-shot learning, emphasizing its role in transfer learning and data augmentation to improve material design workflows. In polymer science, Wu et al. [137] demonstrated an AI-guided few-shot inverse design framework for HDP-mimicking polymers aimed at combating drug-resistant bacteria. By leveraging multi-modal representations and reinforcement learning, the study optimized polymer generation with limited data. Similarly, Waite et al. [138] applied few-shot learning to atomic force microscopy (AFM), enabling automated and sample-efficient analysis of force-extension curves for single-molecule interactions.

(F) Federated Learning is a distributed ML paradigm that enables training models across multiple decentralized devices or servers holding local data, without the need to share the data itself (Figure 20). Introduced by McMahan et al. [139], this approach is designed to address privacy concerns, data security, and regulatory compliance by keeping data localized while only sharing model updates with a central server.

Federated Learning has been widely adopted in applications like mobile device personalization, healthcare, and IoT, where sensitive data cannot be easily centralized. Challenges in Federated Learning include handling data heterogeneity, communication efficiency, and ensuring model performance under limited data availability and system constraints. We are not aware of any application of this principle in the area of polymer research. There are, however, applications using federated learning-based graph convolutional networks for non-Euclidean spatial data, including protein structures [140].

(G) Graph Neural Networks (GNNs) are a class of deep learning models designed to process and analyze data structured as graphs, which consist of nodes and edges (Figure 21). GNNs leverage graph convolutions and message-passing mechanisms to learn representations of nodes, edges, or entire graphs by aggregating information from their neighbors. This makes GNNs particularly effective for tasks where the relationships between entities are as important as the entities themselves. Applications of GNNs span diverse domains, including social network analysis, molecular property prediction, recommendation systems, and knowledge graph reasoning. The foundational work by Kipf and Welling [141] on graph convolutional networks (GCNs) is widely regarded as a milestone in the development of GNNs.

Kai et al. [142] present a ML workflow integrated with quantum calculations and a GNNs to identify ionic liquids (ILs) for ionic polymer electrolytes (IPEs) in lithium metal batteries. The resulting IPE membranes exhibit exceptional mechanical strength and electrochemical performance, enabling high critical current densities, outstanding capacity retention, fast charge/discharge capabilities, and excellent efficiency without flammable organics. Xie at al. [143] introduce a multi-task GNN to accelerate the screening of polymer electrolytes for lithium-ion batteries, overcoming the high computational cost of MD simulations in amorphous systems. By leveraging noisy, short MD data and limited long MD data, the method accurately predicts converged properties, enabling the exploration of a vastly larger polymer space and uncovering design principles for polymer electrolytes.

(H) Bayesian Deep Learning (BDL) combines the flexibility of DNNs with the principled framework of Bayesian inference, enabling models to quantify uncertainty in their predictions [144]. Unlike traditional deterministic NNs, BDL approaches treat model parameters as probabilistic distributions, allowing for the capture of both epistemic (model-related) and aleatoric (data-related) uncertainties (Figure 22). This capability is particularly valuable in high-stakes applications such as healthcare, autonomous driving, and financial modeling, where robust decision-making requires an understanding of prediction uncertainty.

In polymer science, Bayesian Deep Learning has emerged as a powerful tool for addressing complex design and optimization challenges. Wheatle et al. [145] utilized Bayesian optimization to design polymer blend electrolytes, balancing ionic conductivity and mechanical properties. By incorporating molecular simulations and proxy variables, their method optimized trade-offs efficiently, paving the way for high-performance material development. Bayesian methods have also been instrumental in computational phase discovery and molecular simulations. Dorfman [146] reviewed the use of Bayesian optimization in computational phase discovery for block polymers, improving predictions from self-consistent field theory (SCFT)-based models. Similarly, Wheatle et al. [145] applied Bayesian optimization to explore multi-objective design spaces in polymer blend electrolytes, identifying high-performance blends by navigating trade-offs between mechanical and electrical properties. In molecular simulations, Peng et al. [147] used Bayesian inference to backmap CG molecular models into atomic-resolution structures. This approach enhanced the accuracy of molecular dynamics simulations, particularly for biomolecules, by refining the structural details of CG models.

(I) Neural Architecture Search (NAS) is a technique for automating the design of NN architectures, enabling the discovery of optimal configurations tailored to specific tasks. By systematically exploring a search space of candidate architectures, NAS eliminates much of the trial-and-error associated with manual network design (Figure 23). Popular NAS methods include reinforcement learning [148], evolutionary algorithms [149], and gradient-based optimization [150], which iteratively refine architectures to achieve superior performance.

NAS has demonstrated remarkable success in domains such as image classification and object detection, often surpassing human-designed networks [151]. By automating the architecture search, NAS reduces the need for domain-specific expertise while uncovering novel and efficient designs. However, challenges such as high computational costs and the development of efficient search strategies remain active areas of research.

In polymer science and materials engineering, the adoption of NAS is an emerging trend, promising to accelerate the development of predictive models and optimize performance in data-scarce environments. While specific applications of NAS in polymer research are still developing, its proven ability to enhance NN performance in other domains underscores its potential to drive innovation in polymer informatics and related fields.

2.6. Coarse-Grained (CG) Models

CG models have become invaluable in the study of soft materials, offering computational efficiency by simplifying atomically detailed systems while retaining critical observable properties. These models have facilitated significant progress in exploring complex phenomena. Noid [90] provides a comprehensive review of predictive CG models, highlighting their success in modeling liquids and polymers. However, challenges such as transferability and thermodynamic accuracy persist, particularly in complex biomolecular systems. Advances in CG mapping techniques, modeling of many-body interactions, and addressing state-point dependence of effective potentials have paved the way for accurate and transferable CG models, combining rigorous theoretical approaches with modern computational tools.

The integration of machine learning (ML) has enhanced CG methodologies. Bhattacharya et al. [152] introduce dPOLY, a ML framework leveraging DNNs to analyze molecular dynamics (MD) trajectories and predict polymer phases and phase transitions. Using CG-MD, dPOLY accurately identifies critical temperatures for coil-to-globule transitions across various polymer sizes. This generic method can be extended to study other phase transitions and significantly accelerate phase prediction and characterization in polymers and soft materials. Axelrod et al. [21] discuss the application of ML to materials design, emphasizing its potential in identifying high-quality candidates and accelerating simulations. Similarly, Ye et al. [153] review ML techniques for CG modeling of organic molecules and polymers, focusing on descriptors, training datasets, and loss function optimization to improve model accuracy and thermodynamic consistency. These studies underscore ML’s ability to address challenges of transferability and efficiency in CG modeling. Alessandri and de Pablo [154] introduce a ML-enhanced CG method to efficiently predict electronic properties of soft electronic materials, bypassing computationally demanding backmapping and quantum-chemical calculations. Applied to radical-containing polymers for energy storage, the method trains models to link electronic properties with CG conformations, enabling direct prediction from CG simulations. Validation shows strong agreement with traditional methods, highlighting the approach’s potential to accelerate multiscale workflows and design CG models retaining electronic structure details. Karuth et al. [155] integrate cheminformatics and CG-MD simulations to predict the glass-transition temperature $T_g$ of diverse polymers. The quantitative structure-property relationsip model identifies key molecular descriptors influencing $T_g$, while simulations reveal how cohesive interactions, chain stiffness, and grafting density affect $T_g$, offering mechanistic insights. Dhamankar and Webb [91] explore chemically specific CG modeling for polymers, emphasizing the ability to capture polymer behavior over relevant spatiotemporal scales. Joshi and Deshmukh [156] review advancements in CG-MD simulations, showcasing transferable models for solvents, polymers, and biomolecules. Their work illustrates how ML and hybrid approaches have enabled studies of solute-solvent interactions and polymer architectures, offering experimentally verifiable insights. Applications of ML-enhanced CG models have also driven innovation in materials design. Wheatle et al. [145] combined ML with Bayesian optimization to design polymer blend electrolytes, revealing trade-offs between ionic conductivity and mechanical properties. Wang et al. [100] integrated CG-MD with ML and Bayesian optimization to uncover relationships between lithium conductivity and intrinsic material properties. These studies demonstrate the value of combining molecular simulations with ML to optimize material performance for energy applications. Innovative techniques continue to expand CG modeling capabilities. Bejagam et al. [81] developed a temperature-independent CG model to study the coil-to-globule transition of PNIPAM, leveraging ML-based nonmetric multidimensional scaling (NMDS) to analyze simulation trajectories and uncover new insights into polymer transitions. Bejagam et al. [20] introduced a framework combining particle swarm optimization, artificial NNs (ANN), and molecular dynamics simulations to optimize force-field parameters for CG models, demonstrating its utility for solvents like D₂O and DMF. Li et al. [107] tackled the challenge of backmapping CG configurations to atomistic structures using conditional generative adversarial networks (cGANs), demonstrating a versatile and efficient approach for multiscale modeling across varying molecular weights. Lenzi et al. [157] discuss the challenges and progress in modeling aliphatic isocyanates and polyisocyanates, emphasizing the potential of ML in overcoming barriers to accurate parametrizations. Schneider et al. [158] introduce a ML model to simulate defect kinetics during the self-assembly of symmetric diblock copolymers into lamellar phases. Trained on intermediate time-scale simulations of a CG model, the artificial neural network (ANN) predicts time-independent transition probabilities based on the Markov chain framework. This approach enables efficient generation of long-time trajectories, providing novel insights into late-stage defect dynamics and relaxation. The ANN’s input-size independence allows training on small systems and application to larger ones. The method is demonstrated by analyzing defect motion and lifetimes during long-range ordering processes.

ML methods have increasingly been used also to perform atomistic computations of energies and atomic forces, with greater accuracy than empirical functions [159,160,161,162,163,164,165]. Mariloa et al. [166] present a staggered neural network force field (NNFF) architecture that accelerates molecular dynamics simulations by directly predicting atomic force vectors without relying on computationally expensive spatial derivatives. By combining rotation-invariant and rotation-covariant features, the method achieves a 2.2x speed-up over state-of-the-art engines and enables efficient modeling of complex multi-element systems, such as long polymer chains and surface chemical reactions, with broader applicability beyond computational material science.

3. Summary and Conclusions

ML has started to become an indispensable tool in polymer and macromolecular science, offering a suite of techniques that complement and surpass traditional experimental and computational methods. The relevance of these methods varies depending on the specific challenges in polymer research, but several approaches stand out for their transformative potential.

NNs and ensemble methods have been applied to predict tensile strength [42,43,51,61,63,68,75,110,167,168,169,170,171,172,173,174,175,176,177,178,179,180,181,182,183,184,185] and elastic modulus [38,39,42,46,62,75,94,95,122,170,173,177,180,182,186,187,188,189,190,191,192,193,194,195,196], and other rheological properties [87,114,191,197,198,199]. Random forests and gradient boosting algorithms have proven effective for predicting glass transition temperatures [22,48,70,87,113,128,167,176,180,200,201], thermal stability [114,202,203,204,205,206], and thermal conductivity [111,207,208]. These models rely on structure-property relationships derived from experimental datasets. Recent studies, such as [209], have demonstrated the application of SVMs in predicting the refractive index [200], dielectric constants [66,114,194], and conductivity [112,208] of polymeric materials. Emerging trends such as nanoparticle doping, photosensitive polyimides, and ML-assisted molecular design were identified as key to advancing PI films for flexible displays, solar cells, and high-performance devices [202]. Optimizing polymer synthesis pathways is another critical application of ML. By reducing the number of experimental trials, ML-driven techniques enable rapid prototyping and material discovery. Bayesian optimization techniques [59,62,100,145,182] have been successfully applied to minimize experimental costs while maximizing efficiency [83,209]. These methods utilize prior knowledge to iteratively improve synthesis outcomes. Current approaches leverage combinatorial and high-throughput experimental designs to mitigate data limitations; however, challenges persist due to the lack of standardized characterization for critical biomedical parameters like degradation time and biocompatibility. Significant gaps at the intersection of applied ML and biomaterial design, offering insights into current advancements and future directions for overcoming these obstacles have been identified [24]. Reinforcement learning has emerged as a powerful tool for optimizing complex reaction networks. It has been used to identify optimal conditions and reaction pathways for synthesizing high-performance polymers [85,98,109,210]. Ensemble methods, such as gradient boosting, are increasingly used to design scalable manufacturing processes for specialty polymers [85,206]. As environmental concerns grow, ML has become an invaluable tool for designing sustainable and biodegradable polymers. Researchers are leveraging ML to predict degradation rates, optimize recyclability, and reduce the environmental footprint of materials. SVMs and NNs have been applied to classify polymers based on their biodegradability and polymer degradation rates [16,105,203,211]. Studies such as [30,71,212] focus on predicting the recyclability of polymers, enabling more effective recycling strategies. ML techniques have also been used to conduct lifecycle assessments, evaluating the environmental impacts of polymer production and disposal [205,213,214]. So-called multitask ML has been employed to predict polymer-solvent miscibility using Flory-Huggins interaction parameters [215]. Chen et al. [216] demonstrate a novel approach to automated scientific discovery by integrating machine learning with physical principles to construct macroscopic dynamical models directly from microscopic trajectory observations. Using a framework based on the generalized Onsager principle, the authors identify interpretable thermodynamic coordinates and map the dynamical landscape of stochastic dissipative systems. The method is validated experimentally by studying polymer stretching in elongational flow, revealing stable and transition states while enabling control over the stretching rate, with broad applicability to complex scientific phenomena.

Critical evaluations of these works reveal both opportunities and challenges. While supervised and unsupervised learning methods have achieved significant successes, the scarcity of high-quality, labeled datasets remains a bottleneck [129,129]. Moreover, the interpretability of complex ML models, such as DNNs, continues to be a concern for researchers and practitioners. Addressing these limitations will be essential for advancing the integration of ML into polymer science further.

Overall, the integration of ML into polymer science not only enhances our ability to model and predict polymer behavior but also accelerates the discovery and optimization of new materials. Future advancements will likely stem from hybrid approaches that combine different ML methods, leveraging their unique strengths to tackle increasingly complex challenges in polymer research. While the reviewed studies provide a strong foundation for ML-driven polymer design, significant opportunities remain to address limitations in data diversity, model interpretability, and real-world validation.