Dimensionality-guided sparse sampling for time-domain thermoreflectance: high-throughput thermal mapping of multilayer devices

Thermal characterization of multilayer microelectronic devices requires efficient measurement strategies that balance accuracy against investigation resources. Time-domain thermoreflectance (TDTR) offers non-contact thermal property measurement, yet conventional approaches rely on densely sampled data and prior knowledge of material properties. We address these limitations through a data-driven framework combining dimensionality analysis with measurement optimization. Using an autoencoder, we reveal that across the broad range of bilayer systems in the diffusive regime, TDTR signals exhibit a low dimensionality: three degrees of freedom for single-frequency measurements, four for multifrequency measurements. Guided by this limit, Bayesian optimization identifies a nine-point sparse sampling configuration that captures information equivalent to 192-point scans. A neural network trained on this configuration enables simultaneous extraction of film thickness, thermal conductivity, and interfacial conductance where sensitivity permits. We validate the framework on SiO 2 /Si specimens and demonstrate its utility by mapping multilayer ceramic capacitors. Our approach achieves thermal property mapping without prior knowledge of the sample structure, transforming TDTR into a high-throughput image modality.