Abstract
Patterned-transducer frequency-domain thermoreflectance (FDTR) improves sensitivity to low-thermal-conductivity materials by suppressing lateral heat spreading in the metal transducer, yet the loss of lateral semi-infinite symmetry invalidates analytical models and necessitates expensive finite-element simulations, making iterative fitting computationally prohibitive. This work develops a transfer-learning-enhanced reduced-order surrogate that directly predicts the phase response for rapid inverse fitting. The proposed approach achieves a 534× speedup over COMSOL with sub-degree accuracy (mean RMSE 0.19°), reducing a representative three-parameter fit from over 5 hours to approximately one minute. Experimental validation on Al/SiO2 samples demonstrates consistent extraction of silica thermal conductivities across conventional and patterned configurations, confirming metrological reliability. Transfer learning further extends the surrogate to expanded parameter domains with limited additional high-fidelity data, cutting training data requirements by over 80% while preserving prediction accuracy. The proposed framework provides a practical route for high-throughput thermal characterization in complex geometries where conventional numerical fitting is infeasible.

This publication has 20 references indexed in Scilit: