PRISM: Unifying multiphysics optothermal transport and identifiability-guided inference for nanoscale thermal metrology

Abstract

Quantifying thermal transport across micro- and nanoscale heterostructures is critical for solid-state devices and relies on fitting heat-transfer models to thermal-wave spectra. When probing scales approach optical absorption depths, electron-phonon equilibration lengths, or phonon mean free paths, classical surface-heating Fourier models can match measured phase spectra while returning biased, measurement-dependent parameters. Here we introduce PRISM (Physics-Resolved Inference and Screening Meta-framework), an agentic framework coupling multilayer optical transfer-matrix electrodynamics, two-temperature electron-phonon relaxation, and nonlocal Guyer-Krumhansl phonon hydrodynamics within a unified admittance formalism, paired with singular-value-decomposition identifiability screening and knowledge-guided routing. Demonstrated with frequency-domain thermoreflectance across stacks relevant to microelectronic contacts and dielectric barriers, PRISM corrects a 24-28% interfacial conductance underestimation in Au/Si, recovers intrinsic silica conductivity in Au/Ti/SiO$_2$, and eliminates a 40% cross-plane conductivity underestimation in Au/MoS$_2$. In bimetallic Au/Ti/Si stacks, where all three non-classical mechanisms act simultaneously, nonlocal modeling reconciles the $116$-$137$ W m$^{-1}$ K$^{-1}$ local-fit scatter across spot sizes and frequency windows into a configuration-independent intrinsic silicon conductivity of $141$ W m$^{-1}$ K$^{-1}$. For ion-irradiated and graded media with continuous conductivity profiles, PRISM can autonomously construct and verify a deep operator network solver to reconstruct steep damage gradients. By uniting non-equilibrium transport physics and identifiability screening with autonomous model synthesis, PRISM establishes a generalizable paradigm readily extensible to broader optothermal and thermal-wave metrology modalities.