3rd TAGRA

Heat transfer problems in layered and multi-material systems involve discontinuous thermal conductivity, which makes the temperature gradient jump across material interfaces and challenges physics-informed neural networks (PINNs) represented by a single smooth network. This study implements a single-network PINN and a domain-decomposed extended PINN (XPINN) for transient heat conduction in a two-layer composite wall within a reproducible PyTorch-based Python framework, and compares both against an independent finite volume reference. The framework includes interface losses for temperature, flux and residual continuity, together with gradient-norm-based adaptive loss weighting. It is shown analytically that a loss function without interface conditions admits a spurious stationary solution that has zero loss and coincides with the initial profile, and the error measures of this solution reproduce the reported results of the single-network PINN. For a conductivity ratio of 0.2, the single-network PINN yields a relative L2 error of 19.9% and an interface flux mismatch of 2.00, whereas the XPINN reduces these values to 1.57% and 0.044. As the conductivity ratio decreases from 1 to 0.05, the error ratio between the two approaches increases from 1.0 to 8.3. Adaptive weighting reduced the flux mismatch by about 22% but did not improve the relative L2 error in the single-seed experiment. The results show that interface continuity conditions must be included explicitly in the loss function and that a low loss value alone is not a measure of accuracy. >> This study is supported by Ataturk University Coordinatorship of Scientific Research under project # FBA-2026-17685