A Physics-Informed Adaptive Segmentation Fitting Method for Transient Thermal Test Analysis

Abstract Transient thermal testing is a crucial technique for characterizing the thermal properties of semiconductor devices, but measurement noise often compromises its accuracy. Conventional square-root fitting (SRF) methods emphasize early-time details at the expense of global fidelity, whereas exponential fitting (EF) methods capture global behavior but sacrifice early-time physical detail. In this paper, we propose a physics-informed adaptive segmentation (PIAS) approach. By analyzing the differential characteristics of the thermal response, our method automatically identifies the physical transition region where heat flow evolves from quasi-one-dimensional to three-dimensional diffusion, thereby constraining the segmentation point without relying on empirical judgment. A composite objective function—incorporating global fitting error and derivative continuity constraints—is globally optimized via particle swarm optimization (PSO). Simulation and experimental results demonstrate that PIAS overcomes the inherent limitations of traditional techniques. Under strong noise conditions (signal-to-noise ratio as low as 45 dB), it outperforms leading commercial software by reducing the noise-induced error in total thermal resistance from 4.48% to just 0.15%, while achieving subsecond computation per fit. PIAS enables high-precision, fully automated transient thermal analysis with significant engineering relevance.