Leveraging Low-Fidelity Data to Improve Machine Learning of Sparse High-Fidelity Thermal Conductivity Data via Transfer Learning

TL;DR

This paper uses transfer learning to enhance machine learning models for predicting semiconductor thermal conductivity by combining small high-fidelity datasets with large low-fidelity datasets, leading to improved accuracy and efficient screening.

Contribution

It introduces a transfer learning approach that effectively leverages low-fidelity empirical data to improve high-fidelity thermal conductivity predictions in semiconductors.

Findings

Transfer learning improves model R2 by up to 23%.

Average factor difference reduces by up to 30%.

Several high-conductivity candidates identified and verified.

Abstract

Lattice thermal conductivity (TC) of semiconductors is crucial for various applications, ranging from microelectronics to thermoelectrics. Data-driven approach can potentially establish the critical composition-property relationship needed for fast screening of candidates with desirable TC, but the small number of available data remains the main challenge. TC can be efficiently calculated using empirical models, but they have inferior accuracy compared to the more resource-demanding first-principles calculations. Here, we demonstrate the use of transfer learning (TL) to improve the machine learning models trained on small but high-fidelity TC data from experiments and first-principles calculations, by leveraging a large but low-fidelity data generated from empirical TC models, where the trainings on high- and low-fidelity TC data are treated as different but related tasks. TL improves the model accuracy by as much as 23% in R2 and reduces the average factor difference by as much as 30%. Using the TL model, a large semiconductor database is screened, and several candidates with room temperature TC > 350 W/mK are identified and further verified using first-principles simulations. This study demonstrates that TL can leverage big low-fidelity data as a proxy task to improve models for the target task with high-fidelity but small data. Such a capability of TL may have important implications to materials informatics in general.