PhysThermix: Physics-Informed Surrogate for Thermal Prediction of 2.5D Chiplet Architecture

Fast thermal prediction enables chiplet layout exploration, but high-fidelity numerical simulation is too slow for design-space search, and data-driven surrogates do not explicitly enforce heat-transfer physics. We introduce PhysThermix, an extension of ThermLeT that adds spatial thermal conductivity and heat transfer coefficient maps as inputs, together with residual losses derived from a depth-averaged steady heat equation and adiabatic lateral boundaries. We evaluate on 150 held-out power maps spanning three chiplet architectures against finite-element ground truth. PhysThermix reduces MAE by 45% (2.59 to 1.42 °C) and RMSE by 39% (3.13 to 1.90 °C) relative to ThermLeT. It improves on 148 of 150 maps and lowers worst-case error from 6.98 to 4.12 °C peak temperature. With ablations study we observe that combination of property maps and physics loss attains the best result. The largest improvements occur in inter-die spreading regions and near package boundaries, where lateral conduction rather than local power density sets the field.