Machine Learning based Model for Prediction of Lattice Thermal Conductivity with Unprecedented Accuracy

In this work, the high-throughput screening is coupled with the machine learning to develop a prediction model on a dataset, which is independent of any particular class and uses only four descriptors related directly to the physics of lattice thermal conductivity. By employing physically meaningful conditions, high-throughput screening was carried on the dataset. The screening resulted in several ultralow and ultrahigh lattice thermal conductivity compounds.

High-throughput screening still requires the explicit evaluation of lattice thermal conductivity, developmentof the prediction model using machine learning was carried out. The significant contribution of our work was the preparation of data having as much variability as possible. The dataset in our case covers elements from several groups of the periodic table, including 6 crystal systems, two orders of magnitude variation in mass and volume, three orders of magnitude variation in lattice thermal conductivity.For finding the descriptors, an extensive property map was generated, which includes harmonic and quasi-harmonic properties such as phonon dispersion and Grüneisen parameter. From this property map, simple four descriptors were identified related to the physics of lattice thermal conductivity, namely maximum phonon frequency, integrated Grüneisen parameter, average atomic mass, volume. Using these four descriptors, the machine learning model was developed, which predicts the log-scaled lattice thermal conductivity with root mean square error of 0.21 (Figure a).



Figure: Scatter plot for DFT calculated versus (a) ML-predicted κland (b) Slack model predicted κl

As the quasi-harmonic properties also appear in the Slack model, which is a very popular physics based model for the prediction of lattice thermal conductivity, it is employed on our dataset for the comparison. The Slack model severely overestimates the lattice thermal conductivity (Figure b), highlighting the exceptional performance of developed machine learning model.

Reference : 

Coupling the High-Throughput Property Map to Machine Learning for Predicting Lattice Thermal Conductivity Chem. Mater. 31, 14, 5145-5151, 2019

Website URL :