Purpose This study addresses a key limitation in physics-informed neural networks (PINNs), namely the reliance on manually selected or heuristically tuned loss weights governing the balance between data fidelity, physics residuals and boundary constraints. Improper weighting often leads to instability, poor reproducibility and sensitivity to user expertise, particularly in manufacturing-oriented thermal modelling and digital twin applications. Design/methodology/approach A genetic algorithm (GA)-based meta-optimization framework is proposed to automate the selection of normalized loss weights in PINNs. The GA operates as an offline optimization layer, while the inner PINN enforces the governing partial differential equations. A composite fitness function evaluates candidate weight configurations using physics consistency, data agreement and boundary-condition satisfaction. The framework is validated using a two-dimensional transient heat-conduction problem with a moving Gaussian heat source representative of laser-based manufacturing processes under noisy data conditions. Comparative analysis is performed against fixed-weight and adaptive-weight PINN strategies. Findings The proposed GA-optimized PINN achieves an approximately 39% reduction in governing-equation residual root-mean-square error (RMSE) compared with fixed-weight training while maintaining comparable prediction accuracy and stable convergence. The framework produces lower PDE residual distributions and improved robustness under noisy conditions. Compared with adaptive weighting approaches, the GA-based framework demonstrates improved global loss balancing and stronger physics consistency without degrading predictive performance. Originality/value This work introduces a global meta-optimization strategy for PINN loss balancing using a genetic algorithm, enabling systematic and interpretable selection of normalized loss weights for reliable physics-informed surrogate modelling in manufacturing digital twin applications Graphical abstract A diagram illustrating a GA-Optimized Physics-Informed Neural Network for Manufacturing Digital Twin. A diagram of a GA-Optimized Physics-Informed Neural Network for Manufacturing Digital Twin. Panel A: Physical Manufacturing Process. A laser beam heats a melt pool on a physical manufacturing asset. An infrared camera and thermocouple stream sensor data to a system for streaming data assimilation. Panel B: GA-Optimized PINN Framework. A genetic algorithm meta-optimization process involving initialization, selection, crossover, and mutation optimizes loss weights. These optimized loss weights are used in a Physics-Informed Neural Network surrogate model, which includes data loss, physics residual loss, and boundary loss components. Panel C: Digital Twin Outputs. The outputs include full-field temperature prediction, monitoring over time, anomaly detection, and control feedback.
Aswin Karkadakattil (Wed,) studied this question.