Quantitative prediction of grain boundary thermal conductivities from local atomic environments

Quantifying the dependence of thermal conductivity on grain boundary (GB) structure is critical for controlling nanoscale thermal transport in many technologically important materials. A major obstacle to determining such a relationship is the lack of a robust and physically intuitive structure descriptor capable of distinguishing between disparate GB structures. We demonstrate that a microscopic structure metric, the local distortion factor, correlates well with atomically decomposed thermal conductivities obtained from perturbed molecular dynamics for a wide variety of MgO GBs. Based on this correlation, a model for accurately predicting thermal conductivity of GBs is constructed using machine learning techniques. The model reveals that small distortions to local atomic environments are sufficient to reduce overall thermal conductivity dramatically. The method developed should enable more precise design of next-generation thermal materials as it allows GB structures exhibiting the desired thermal transport behaviour to be identified with small computational overhead.