Accuracy and Computational Cost of Standard and Sequential Time-Window Physics-Informed Neural Networks for Transient Heat Conduction with a Moving Gaussian Heat Source

Background The thermal field generated by a moving laser heat source evolves continuously as the laser scans across a material surface. Physics-Informed Neural Networks (PINNs) offer an alternative for modeling transient heat transfer, but quantitative comparisons between full-domain and sequential time-window training remain limited. Methods This study compares Standard PINN and Sequential Time-Window PINN for two-dimensional transient heat conduction with a moving Gaussian heat source at scanning speeds of 1, 2, and 3 mms−1 over 10 s. Reference solutions were obtained using the forward-time central-space (FTCS) method on a 500×500 grid after grid-convergence and stability checks. Standard PINN was trained over the full temporal domain, whereas Sequential Time-Window PINN used ten consecutive 1 s windows with parameter transfer between windows. Results Standard PINN achieved a maximum-temperature mean absolute error (MAE) of 0.461–0.619 °C and a mean absolute percentage error (MAPE) of 0.210−0.322% . Sequential Time-Window PINN yielded a maximum-temperature MAE of 19.248−25.581°C . Its absolute errors ranged from 103.05 to 158.16 °C at 1 s and decreased to 0.98−6.64°C at 10 s. Sequential Time-Window PINN trained 6.80−7.89 times faster than Standard PINN. However, both PINN approaches required more computational time than FTCS. Conclusions Standard PINN provided higher predictive accuracy, whereas Sequential Time-Window PINN reduced training time but produced larger errors, particularly during the early simulation. Differences in training-point counts, collocation-point sampling, and temporal strategy prevent attributing the runtime reduction solely to time-windowing. Further controlled experiments are needed to isolate the effects of these factors on accuracy and computational cost.