Engineering

Abstract: Learning reusable local operators can accelerate thermoelastic topology optimization, yet accurate stiffness prediction alone does not ensure accurate thermal-force transfer between fields. A hybrid Problem-Independent Machine Learning (PIML) framework combines reference-anchored thermal shape-function learning with mechanical residual-enriched static condensation. A fixed-trace error analysis exposes the mechanism: stiffness errors are quadratic about interior equilibrium, whereas thermal-load projection can retain first-order errors. This distinction motivates a field-specific strategy that learns the compact thermal response and corrects the larger mechanical response from current-material equilibrium. Shared Ritz factors supply stiffness, coupling and local state/adjoint products without mechanical training or explicitly forming the full mechanical basis. Across eight three-dimensional optimization benchmarks exceeding one million fine elements each, thermal designs have compliance gaps below 0.3% relative to full finite element method (FEM) reference designs; coupled designs have mixed-objective gaps below 0.027% relative to exact linear-boundary substructure references. Complete coupled optimizations are 4.75–5.05× faster than exact Q1 substructuring in total online time. In separate warmed benchmarks covering states, sensitivities and update proposals, the coupled hybrid gives a 186× speedup over matched Full FEM and reduces peak working-set memory by 94.4%. On a 32 GB laptop, a coupled plate with 30.752 million analysis cells is optimized in 60.54 min using 14.17 GiB peak resident memory. The framework extends reusable local learning from separate-field stiffness approximation to accurate, memory-efficient coupled design.

Abstract: Friction monitoring is essential for the precision control and fault diagnosis of electromechanical actuators (EMAs). This paper presents a friction estimator that reconstructs friction under various loads and velocities solely from motor current signals. The estimator consists of a preprocessing stage and an extrapolation training stage using a physics-informed neural network (PINN). In the preprocessing stage, basic Stribeck parameters are estimated from no-load constant-speed measurements. In the extrapolation training stage, a load-augmented Stribeck law is embedded as a soft constraint within the PINN, and friction is distinguished from reflected load torque by its velocity-structured baseline. Because the load dependence is encoded in the physical law, the estimator extrapolates to loads absent from training, reconstructing the friction map from limited samples; a theoretical analysis derives the conditions for well-posed extrapolation. In dynamic simulations, the preprocessing stage estimates the base Stribeck parameters with a mean relative error of 0.87%. On graded load-extrapolation experiments, the constrained PINN keeps the mean relative error close to 1% at loads far beyond the maximum training load, about 90% lower than data-driven methods, and predicts friction in untrained dynamic processes with a mean relative error below 4.8%. Few quasi-static points thus suffice to reconstruct the full velocity–load friction map without force or torque sensors, and periodic re-estimation can track friction over the service life.

Abstract: This study evaluates the static, dynamic, fatigue, and multi-objective performance of an Aluminum-Alumina functionally graded material (FGM) mono-leaf spring relative to a conventional 50R steel spring. Six power-law indices (k = 0.2, 0.5, 1, 2, 5, and 10) were analyzed using ANSYS Workbench. The FGM designs reduced mass by 60.5-65.1% while maintaining maximum von Mises stresses of 28.718-30.897 MPa, close to the steel value of approximately 28.6 MPa. All FGM cases exhibited higher natural frequencies, with the first mode increasing from 310.58 Hz for steel to 430.29-550.14 Hz for the FGM designs. Total elastic strain energy decreased from 327 mJ at k = 0.2 to 91 mJ at k = 10. The Basquin S-N model indicated superior high-cycle fatigue strength for steel, while the ANSYS fatigue-life predictions for the FGM cases were treated as model-dependent because they extend beyond the adopted S-N data range. A four-objective design assessment minimized mass and maximized fatigue factor of safety, first natural frequency, and total strain energy. All FGM designs satisfied the imposed constraints, and equal 25% weighting identified k = 0.2 as the best-balanced configuration, with a combined normalized score of 0.75. The results demonstrate that FGM gradation provides substantial lightweighting while introducing trade-offs among static, dynamic, energy-storage, and fatigue-related performance.

Abstract: Erythritol is a promising phase change material for medium-temperature thermal energy storage, but its low thermal conductivity leads to slow heat release from the material during solidification. A three-dimensional numerical study of erythritol solidification in a cylindrical shell-and-tube latent heat thermal energy storage unit with three different internal arrays – a finless annulus baseline, traditional pin fins, and a deterministic fractal-fin configuration – is presented. A one-fifth sector was modelled in ANSYS Fluent using the enthalpy-porosity method. Buoyancy forces were modelled using the Boussinesq approximation, and the mesh-independence, mushy-zone, and time-step sensitivity tests were performed for numerical validation, along with a natural-convection phase-change benchmark from the existing literature. The production mesh comprised 732,581 cells and 3,524,743 nodes, with a time step of 1 s. For the finned annulus, pin-fin, and fractal fin arrays, the quantities of added conductor were nearly the same, differing by only 0.11%, but the fractal array afforded 100.4% more fin-PCM interfacial area. The total solidification times were 41,000 s for the finless array, 38,500 s for the pin-finned array, and 36,500 s for the fractal array; the liquid fractions at 20,000 s were 0.348, 0.286, and 0.227, respectively, and the measured mean energy recovery rates at 20,000 s were 7.49 W, 7.51 W, and 8.03 W. The findings suggest that branching conductive structures by using a fractal network with an approximately constant volume of added conductor per generation shortens conduction paths during late-stage solidification of erythritol.

Abstract: In the realm of physically taxing manufacturing work, active exoskeletons serve as wearable robotic aids to bolster a worker's strength, endurance and ergonomics. Such devices are well suited to the demands of Industry 5.0 and human-robot collaboration, where an unobtrusive interface can be key to improving both safety and output. The present manuscript offers a synthesis of what has been written in research reviews and implementation studies on the conventional and less common applications of these exoskeletons in manufacturing. The current state of the art is a matter of systems-level balance: between the power of assistance and such factors as kinematic alignment, adaptive control and keeping distal mass to a minimum. Cable-driven, soft and pneumatic designs are preferred for this reason, with control methods moving in the direction of intent-aware, adaptive support. The most compelling data from conventional use comes in overhead operations, where active equipment has been shown to cut anterior deltoid activity, or in repetitive lifting, with back-support systems lowering muscle exertion. There are also unconventional applications making headway, from precision positioning and trunk-flexion to mobility-heavy handling in shared workcells with collaborative robots.

Abstract: Sustained operation near transient structural resonance remains a critical degradation factor for variable-speed rotating machinery, as conventional passive isolators cannot adapt to shifting forcing frequencies. To overcome the structural complexity and power demands of active systems, this study experimentally validates a novel, field-free method for real-time adaptive vibration isolation via controlled transverse strain induction in 60 Shore A elastomeric mount. A mathematical framework incorporating the Rectangular Deformation Hypothesis was formulated to map geometric nonlinearities, establishing a lightweight third-order polynomial approximation (Y=0.0002302X³+0.01494X²+1.71X+234) for immediate stiffness prediction. Testing was executed on a motor-driven rig across seven speed regimes (400–600 RPM) under incremental stepper-motor strokes (0.0–6.0 mm). Time-domain acceleration streams filtered via OriginPro reveal a highly non-monotonic response: an initial 1.5 mm displacement triggers transient resonance tracking, whereas higher compression limits (4.5–6.0 mm) achieve optimal structural impedance, yielding substantial multi-axial attenuation along both radial and transverse paths. Frequency spectral maps confirm that mechanical pre-strain purposefully shifts structural eigenvalues away from localized driving frequencies. This work successfully verifies the operability of the authors' patented concept, offering a scalable, mechanically realizable strategy for rotating machinery health management.

Abstract: Laser Powder Bed Fusion (LPBF) is one of the most established metal Additive Manufacturing technologies, yet obtaining predictable links between various process parameters and material properties is still difficult because of the complex and nonlinear nature of LPBF processing. In this study, an Explainable Machine learning framework is presented to predict and optimise the material performance in LPBF using the input parameters (grain size, hatch width, laser speed, layer thickness, and laser power). The total of 1000 LPBF samples developed from experimentally obtained samples was used for model development using Generative Adversarial Network (GAN) based augmentation. Three models: Linear Regression, Random Forest, and Extreme Gradient Boosting (XGBoost) were tested to predict tensile strength and surface roughness. The model performance was evaluated by the coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE) and five fold cross validation. The use of Explicable Artificial Intelligence (XAI) techniques, such as SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDP), were used to explain how models worked and to determine which process variables were most important. For this, another optimization system was developed based on a machine-learning framework to identify process conditions that maximize tensile strength. RF model showed the highest performance in the prediction of tensile strength with a mean CV R2 of around 0.80 which demonstrated a good prediction skill. Conversely, surface roughness did not exhibit good predictive power with the highest R2 value for the model being around 0.19. Explainability analysis has always shown laser power to be the most significant parameter which has influence on tensile strength. The optimization results showed that the process window, which was associated with the maximum predicted tensile strength of 475.69 MPa, was calculated. The proposed framework illustrates opportunities for explainable machine learning for LPBF process prediction, interpretation and optimization aiding in data-driven manufacturing decision making.

Abstract: Laser and electron beam powder bed fusion of metals (PBF-LB/M and PBF-EB/M) processes are among the most widely used metal Additive Manufacturing (AM) technologies, offering unprecedented design freedom by enabling the fabrication of complex geometries that are impossible or prohibitively expensive to produce using conventional manufacturing methods. However, ensuring consistent part quality remains a significant challenge, as defects such as lack of fusion and keyhole porosity, cracks, surface irregularities, and dimensional inaccuracies can arise from suboptimal process parameters. Traditional parameter optimization aimed at minimizing such defects relies on extensive experimental campaigns, which are costly, time-consuming, and often material-specific. Moreover, the optimized process parameters must not only suppress defect formation but also maintain high productivity and cost-efficient processing to meet industrial manufacturing requirements. Recently, machine learning (ML) approaches have emerged as powerful tools to model relationships between PBF-LB/EB process parameters and resulting defects. However, ML models typically require large datasets that are not available for newly developed alloys. Transfer learning (TL) provides a pathway to address this limitation. In essence, transfer learning refers to the reuse of knowledge gained from one domain (e.g., a well-characterized alloy with extensive process data) to improve learning performance in a related but data-scarce domain (e.g., a newly developed alloy). Instead of training a model from scratch, TL adapts pre-trained models by fine-tuning them using a limited amount of new data. This review summarizes recent advances in ML and TL approaches for PBF-LB/EB process optimization, with a particular focus on their role in improving part quality and enabling efficient parameter development for emerging alloy systems.

Abstract: This study introduces PCREM, a principal-component analysis/artificial neural network (PCA-ANN) framework for identifying reduced ethanol-combustion mechanism topologies from a two-dimensional OpenFOAM ethanol pool-fire analogue simulated using the 2016-12-14 San Diego Mechanism. For PCA and species ranking, 17 species are selected from the San Diego-generated fields according to the modified Millán-Merino retained-species set, with temperature included as an additional thermochemical scalar. PCA reduces the 18-dimensional thermochemical state to 10 principal components capturing 99.939% of the cumulative variance, while ANN models reconstruct temperature, species mass fractions, principal-component source terms, and diffusion information. A unity-weighted sensitivity analysis is subsequently used to rank the influence of the original thermochemical scalars on the selected quantities of interest. Based on this ranking, three candidate reduced-state mechanisms are proposed: PCREM-16, removing CH₃ CH₂O; PCREM-15, additionally removing CH₃CHOH; and PCREM-13, further removing HO₂ and C₂ H₂. Reaction-consistency analysis distinguishes directly closed pathways from those requiring ANN-based retrieval. The results establish an a priori, data-informed framework for reduced ethanol pool-fire chemistry and provide the basis for subsequent a posteriori OpenFOAM validation and computational-performance assessment.

Abstract: The general influence of the reheat pressure on the performance of the steam Rankine cycle with reheat, over a broad range from 0.1 to 12.5 MPa, is investigated. The response curves of several performance variables are computed as the reheat pressure changes while fixing some cycle parameters at reasonable reference values. The findings show that both the specific total input heat and the specific rejected heat monotonically decrease with the reheat pressure, approximately according to a logarithmic function; the specific net output shaft work has a peak value near the low reheat pressure of 0.2 MPa, and it can be approximated by a cubic function; the thermal cycle efficiency has an optimum peak value near a reheat pressure of 3.2 MPa, with a quartic regression expression for its natural-logarithm transformed quantity; the steam dryness fraction at the exit of the high-pressure turbine increases monotonically with the reheat pressure with a profile that closely a cubic polynomial; the steam dryness fraction at the exit of the intermediate-pressure turbine decreases monotonically with the reheat pressure, approximately according to a logarithmic profile; and the specific entropy generation decreases monotonically with the reheat pressure following nearly a power function. Also, our results show that there is a narrow range of reheat pressures during which the stream dryness fraction at the exit of both turbines is simultaneously above the acceptable threshold of around 90% (thus, 10% stream wetness fraction). This range does not include the optimal reheat pressure of maximum cycle efficiency.

Abstract: Dynamic stall produces different loads at the same incidence during pitch-up and pitch-down. We reconstruct this hysteresis from 128 synchronized University of Glasgow NACA 0012 pressure samples using a common normalized ordered-path coordinate adapted from the authors’ IEEE fuzzy-orbit framework. Four shared-antecedent, three-output first-order Takagi–Sugeno models—FCM–TS6, SC16–TS, SC9–TS, and ANFIS–TS9—represent lift, pressure drag, and quarter-chord moment. Across ten matched 70/30 within-cycle partitions, mean pooled normalized root-mean-square errors are 0.03468, 0.02754, 0.03552, and 0.01781, respectively. ANFIS–TS9 also gives the lowest error under 5% training-output contamination (0.03833); SC16–TS better preserves pressure-drag and moment signed-contour-integral magnitudes in full-data fits. Shared rules provide a common representation, but no accuracy advantage over independently trained single-output models has been established. Separate seven-condition assessments yield mean pooled NRMSE values of 0.16871 for fixed-bandwidth fuzzy-surrogate transfer and 0.10500 for nested measured-trajectory interpolation. These unmatched protocols do not support an algorithm ranking. The findings support interpretable reconstruction of recorded hysteresis, not prediction of arbitrary motion histories. Small sample size, within-cycle dependence, and limited unseen-frequency validation bound the conclusions.

Abstract: The aim of this study was to develop and preliminarily validate an accelerometer-based method for objectively assessing the fidelity of laboratory-replicated motorcycle road vibrations and supporting the operator’s decision to terminate or continue the replication process. Reliable reproduction of real-world road inputs is necessary for repeatable investigations of vehicle dynamics and safety-relevant loads. Vertical front- and rear-axle accelerations recorded on different road surfaces were used as targets for a two-channel electrohydraulic test rig. The method combines percentage errors in response minima and maxima, retention of the best result obtained, time- and frequency-domain signal-fidelity assessment, and machine-learning prediction of the next trial. The retained responses exhibited high lag-compensated correlation and high in-band coherence. The extremum-based criterion detected local degradation that RMS error did not consistently reveal. A random-forest regressor produced a lower forecast error than a persistence baseline; however, the limited number of independent sequences does not support autonomous stopping. The proposed method provides a transparent, verifiable operator-support layer for sensor-based motorcycle testing. It may improve the repeatability and credibility of laboratory assessments of vehicle response to real-world road inputs, but validation over repeated sessions is required before automation.

Abstract: Conventional reservoir-level measurement systems convert hydrostatic pressure into an electrical signal, which makes the entire sensing chain dependent on electronic components that are vulnerable to lightning-induced transients, power interruption, and long-term environmental degradation. This paper proposes a hydromechanical measurement architecture in which hydrostatic pressure is converted directly into an n-bit digital code through a purely mechanical successive-approximation procedure, with no electrical component anywhere in the sensing or comparison chain. A lever-balance comparator, pivoting on a knife-edge, is weighed against a reference column fed by a bank of individually calibrated, laminar-flow binary-pressure orifices; a fixed manual protocol, structurally identical to the classical single-pan binary-weight problem, halves the search interval at each of n steps, resolving the most significant bit first. A Monte Carlo simulation propagating a physically-derived, bit-dependent orifice calibration-retention error, differential fluid-density mismatch, and comparator friction through the full binary-search procedure identifies a practical operating point of n = 10 bits, giving an effective resolution of approximately 1.4 mm (7.7 effective bits) over a 1.024 m bench-scale range. The architecture extends our prior demonstration that the successive-approximation principle is independent of the specific physical quantity involved, showing here that it is also independent of the electrical domain altogether.

Abstract: This paper focused the degradation mechanism of microwave transmittance of glass fiber reinforced polymer (GFRP) plates subjected to laser ablation. During the tests, the wave transmittance experiments were performed on the laser-ablated GFRP plates in Ku- and K-bands (12-26.5 GHz). Based on the experimental results, a thermal-mechanical-electromagnetic numerical model was proposed to achieve the laser ablation and wave transmittance simulation, in which the UMAT and UMATHT user-defined subroutines were developed to analyze the pyrolysis characterization. The results indicated that pyrolysis and ablation products significantly raised the electrical conductivity and dielectric constant, which amplified the impedance mismatch. As the electromagnetic wave transmitted in the ablated the GFRP plate, the greater impedance mismatch promoted the stronger reflection. Furthermore, the pyrolysis char would produce the ohmic loss, which tended to be stable once the depth of the pyrolysis char reached a critical value.

Abstract: Haul road condition is recognised as a driver of tyre wear, fuel consumption and maintenance cost of mining haulage vehicles, but its link to reliability is usually stated qualitatively. Road surface irregularities are classified by wavelength and amplitude, and a dimensionless roughness index Q is defined from defect density and amplitude; for the categories of a published microprofile classification it gives Q ≈ 0.15, 0.50 and 1.00 for good, moderate and poor roads. Reliability is modelled by an additive hazard, which yields R(t|Q) = R0(t)exp(-γQt): the relative reliability decays exponentially in Q at a rate proportional to mission time and remains a probability for all Q ≥ 0. Parameters are estimated by Poisson regression of failure counts on exposure by road category, and a closed-form rule gives the number of failures needed to detect a given rate ratio (about 200 for a 1.5-fold difference at 80 % power). The case data comprise 2737 daily downtime records for four CAT 773E haul trucks at the Sayak mine from 1 January 2023 to 30 September 2026 and 251 component-monitoring records as of 1 September 2026. Consolidation of consecutive records with the same truck and classifier gives 295 technical episodes; tyre-related events are the largest group (30.2 %), while transmission and final-drive events account for 72.5 % of technical downtime. Because the records contain no exposure by road category, they characterise downtime structure and replacement priorities but do not quantify the road effect; a data collection protocol for doing so is proposed. Mid-range speeds of the road categories imply 33 % and 129 % longer travel times on moderate and poor roads.

Abstract: To address the insufficient core feature inheritance, weak control of morphological variation, and limited quantitative evaluation in product-family form design, this study proposes a constraint-based generative framework that integrates homologous gene theory and shape grammar. Using handheld lithium-ion drills, homologous product genes are extracted from structural, functional, and visual features and classified into rigid conservation and flexible variation layers based on their effects on functional stability and brand recognition. The rigid layer constrains the topology and functional layout of the core components to preserve the operational logic and family identity. The flexible layer governs local morphological elements and supports differentiated derivation through shape grammar rules. Based on this framework, a shape grammar rule library for lithium-ion drills is established, and two-dimensional line-drawing schemes are generated. A model combining sequence similarity, topological conservation, functional relevance, and visual morphological similarity was constructed, and visual similarity was calculated using multi-algorithm fusion. The results showed that the homology index ranged from approximately 0.84 to 0.88, indicating that the generated schemes maintained high visual and homologous similarity to the original scheme. Overall, the proposed method enables morphological innovation while preserving core product-family genes, thus providing an evaluable approach to product-family styling design.

Abstract: Owing to uneven ground conditions in hilly-mountainous regions, a profiling-type shaping mechanism was adopted to follow terrain fluctuations and simultaneously complete ridge shaping. The influencing factors of ridging and profiling-type shaping were investigated, and their coupling relationship was analyzed. An EDEM–RecurDyn coupled discrete element–multibody dynamics model was established to investigate soil particle motion and ridge formation. Single-factor tests and a three-factor, three-level Box–Behnken response surface experiment were conducted using the horizontal inclination angle of the soil-covering plate, ridging blade installation angle, and blade rotational speed as factors, with the qualification rates of ridge dimensions and soil compaction as evaluation indices. As the factor levels increased, both indices initially increased and then decreased. The ridging blade installation angle had the greatest effect on ridge dimensions, whereas the soil-covering plate angle had the greatest effect on soil compaction. After rounding the multi-objective optimization results to match the practical adjustment precision of the machine, the optimal parameters were determined as a soil-covering plate angle of 25°, a ridging blade installation angle of 30°, and a rotational speed of 410 r/min. Under these conditions, the predicted qualification rates of ridge dimensions and soil compaction were 91.37% and 95.34%, respectively. Field tests using the optimal parameter combination yielded corresponding qualification rates of 88.6% and 91.4%, with relative errors of 3.0% and 4.1% compared with the predicted values. The results provide a basis for coordinated matching between soil supply and shaping compaction during potato ridging in hilly heavy clay soils.

Abstract: In this work, we investigate the atomic-scale mechanical behavior of key metals (Al, Au, Cu, and Ni) used in drilling alloy and structural components for smart mining systems. Molecular dynamics (MD) simulations were performed to explore their stress-strain behavior under uniaxial compression from 300 K to 1000 K. The results display a linear decrease in uniaxial compressive strength (UCS) and Young’s modulus with increasing temperature due to lattice expansion and weakened interatomic potential derived from the Embedded Atom Method (EAM) formalism. Among all the metal studied, Ni exhibits the highest mechanical performance, maintaining a compressive strength of 7.40 GPa at 1000 K, which makes it a promising candidate for high-performance drilling alloys operating in severe thermal environments. Cu shows moderate thermal stability with a UCS of 5.54 GPa at 1000 K, whereas Au demonstrates the lowest thermal resistance and the greatest softening tendency, reaching only 2.21 GPa at the same temperature. Common neighbor analysis (CNA) and Radial Distribution Function (RDF) reveal that Ni preserves stronger atomic ordering because of its higher cohesive energy, while structural degradation is associated with the reduction of perfect Face Centered Cubic (FCC) configurations and the formation of defects and dislocations under compression. In comparison, Al shows stable defect recovery on heating, giving better resistance to thermal fatigue in automated mining sensor housings. These atomistic results that are temperature dependent are valuable input parameters for predictive algorithms and digital twins in advanced mining engineering.

Abstract: Concurrent faults in gears, bearings and shafts remain among the hardest cases for vibration-based condition monitoring of industrial gearboxes, because their signatures overlap, modulate one another and shift with speed and load. The 2009 PHM Society gearbox has become the standard public benchmark for this multi-component problem, yet published results on it are scattered across input types, network designs and validation schemes, which makes headline accuracies hard to compare. This article presents a critical review of deep-learning diagnosis on this benchmark and the wider gearbox literature, organised around three design choices: signal representation, model family and evaluation protocol. To ground the discussion, we re-analyse a controlled experiment in which three convolutional architectures were trained on 10,240 images per representation (continuous wavelet transform scalograms, short-time Fourier transform spectrograms and wavelet synchrosqueezed images) covering eight spur-gearbox health states. A compact five-stage batch-normalised network reached 99.90%, 99.85% and 99.66% test accuracy with wavelet, Synchrosqueezed and spectrogram inputs, respectively, but the margins among the best models correspond to only a few of the 2,048 test images. On a separate held-out validation subset, accuracy fell to 94.53%, 92.03% and 77.34%, making the representation effect far more visible. Comparison with published capsule-attention, architecture-search and transformer studies shows that reported accuracies on the same benchmark range from below 80% to above 99.5%, mainly because of differences in splitting protocol. We conclude with a reporting checklist and research agenda for leakage-aware, physics-informed and deployable gearbox diagnosis.

Abstract: Urban air quality monitoring networks are traditionally designed to capture mean concentration fields. However, public health and environmental regulations are increasingly focused on extreme pollution events (e.g., the 99th percentile, p99), which exhibit high spatial variability driven by turbulence. Placing sensors at locations of maximum mean concentration (C_mean) often fails to capture these rare, extreme peaks. This study proposes a unified empirical and optimization framework for the active design of sensor networks specifically tailored for extreme value prediction. Using the COST ES1006 Michelstadt wind tunnel dataset, we combine the empirical relation p99 = C_mean + Kcσ with spatial interpolation and two optimization algorithms: Greedy forward selection and a Genetic Algorithm (GA). The objective is to identify the minimal set of sensor locations that minimizes the Mean Absolute Error (MAE) of p99 predictions across all non-sensor locations. The results demonstrate that classical placement strategies (sensors at maximum C_mean) yield an MAE of 96.98 ppmV, highlighting their inability to capture spatial extremes. In contrast, the proposed optimization framework drastically reduces the prediction error. With only 10 sensors, the Greedy algorithm achieves an MAE of 18.16 ppmV, while the GA achieves 19.21 ppmV. Furthermore, a sensitivity analysis reveals a clear "elbow" in the performance curve: a network of 8 to 10 optimally placed sensors is sufficient to reconstruct the p99 field with an MAE of approximately 20 ppmV, beyond which diminishing returns are observed. The optimal sensor distribution reveals a balanced strategy, combining source-proximal, high-altitude, and peripheral locations. This methodology provides a robust, cost-effective tool for engineers and policymakers to design efficient urban air quality monitoring networks capable of accurately predicting worst-case pollution scenarios.