Predicting interfacial thermal resistance by machine learning

Lots of factors affect the interfacial thermal resistance (ITR), making ITR prediction a high-dimensional mathematical problem. Machine learning is a cost-effective method to address this high-dimensional problem. The ITR predictive model is based on experimental data. The physical, chemical, and material properties of ITR are categorized into three sets of descriptors: property descriptors, compound descriptors, and process descriptors. There are three algorithms used for the models, support vector machines (SVMs), Gaussian Regression Processes (GRPs) and Regression tree ensembles of LSBoost (simplified as LSBoost). Those descriptors assist all the models in reducing the mismatch between predicted and experimental values and reaching high predictive performance of 96%. Over 80 thousand material systems composed of 293 kinds of materials (including elements or binary compounds) were input for prediction. Among the top-100 high-ITR predictions by the three different algorithms, 25 material systems are repeatedly predicted by at least two algorithms. One of the 25 material systems, Bi/Si, was processed and achieved the ultra-low thermal conductivity in our previous work. We believe that the high-ITR material systems from the prediction can be the potential candidates for thermoelectric application. This study proposed a strategy of material systems exploration for thermal management by means of machine learning.