A transfer-learning-enhanced reduced-order surrogate for rapid thermal property extraction in patterned FDTR

Patterned-transducer frequency-domain thermoreflectance (FDTR) enhances sensitivity to low-thermal-conductivity materials by suppressing lateral heat spreading within the metal transducer. However, the finite patterned geometry is not directly described by conventional analytical models formulated for laterally unbounded transducers, making iterative parameter fitting based on finite-element simulations computationally prohibitive. Here, we develop a reduced-order surrogate model, enhanced by transfer learning, that directly predicts the FDTR phase response for rapid inverse thermal-property extraction. Within the original training domain, the surrogate achieves a mean phase root mean square error of 0.19° relative to the full-order comsol simulations and accelerates forward prediction by 534-fold, reducing a representative three-parameter fit from more than 5 h to approximately 1 min. Measurements of Al/SiO2 samples yield consistent silica thermal conductivities using conventional and patterned transducers, validating the surrogate-based fitting framework. For anisotropic Z-cut LiNbO3, the method yields radial and cross-plane thermal conductivities of 3.35 ± 0.19 and 3.80±0.75W/(mK), respectively, in close agreement with independent time-domain thermoreflectance measurements on the same material. Transfer learning extends the surrogate to broader parameter domains using only 1000 new simulations, compared with 6000 used to develop the original model, while outperforming direct training under the same target-domain data budget. This framework enables rapid inverse thermal characterization for complex measurement geometries where repeated high-fidelity forward simulations would otherwise be computationally prohibitive.