Thermo-Constrained Physics Informed Neural Network Based Optimization of Mechanical Performance in Recycled PLA Additive Manufacturing

• A thermo-constrained physics-informed neural network is developed for recycled PLA additive manufacturing. • Physical laws governing crystallinity, degradation, and geometry are embedded directly into the learning framework. • The proposed model achieves high prediction accuracy with R² ≈ 0.96, outperforming conventional machine-learning models. • Thermo-constrained Bayesian optimization identifies Pareto-optimal trade-offs between strength, stiffness, and ductility. • Optimized recycled PLA components exhibit near-virgin mechanical performance with improved resistance to moisture and recycling degradation. Recycled polylactic acid (PLA) has emerged as a sustainable feedstock for material extrusion–based additive manufacturing; however, repeated recycling introduces thermal degradation, moisture sensitivity, and microstructural inconsistencies that significantly deteriorate mechanical performance. Existing data-driven prediction and optimization approaches largely rely on black-box machine learning models, which often fail to capture thermomechanical consistency, leading to poor generalization and physically unrealistic optimization outcomes. To address these limitations, this study aims to enhance the mechanical performance and reliability of recycled PLA components through a physics-informed and thermo-constrained computational framework. The primary objective is to establish robust correlations between processing parameters, thermal history, and mechanical properties, while simultaneously optimizing strength–ductility trade-offs under realistic degradation constraints. A novel Thermo-Constrained Physics-Informed Neural Network (TC-PINN) coupled with Bayesian Optimization (BO) is developed, embedding thermomechanical principles such as crystallinity evolution, annealing effects, and degradation trends directly into the learning process. The framework integrates experimental data, physics-based constraints, and probabilistic optimization to predict ultimate tensile strength, Young’s modulus, elongation at break, and crystallinity with high fidelity. The proposed method employs Python-based scientific computing tools, including numerical solvers, neural network libraries, and optimization modules, to ensure reproducibility and scalability. Compared to data-driven models such as BPNN (R² = 0.86), RF (0.71), and SVR (0.44), the proposed thermo-constrained PINN integrated with Bayesian optimization achieves superior prediction accuracy (R² ≈ 0.96) and improved mechanical property retention under recycling and moisture effects. Pareto-front analysis highlights achievable trade-offs between strength and ductility, supporting application-driven decision-making. Overall, the study confirms that physics-guided artificial intelligence provides a reliable pathway for overcoming material degradation challenges in recycled polymer additive manufacturing. The proposed framework offers a scalable, interpretable, and industry-relevant solution for sustainable high-performance 3D printing.