Lattice Thermal Conductivity: An Accelerated Discovery Guided by Machine Learning

In
the present work, we used machine learning (ML) techniques to
build a crystal-based model that can predict the lattice thermal conductivity
(LTC) of crystalline materials. To achieve this, first, LTCs of 119
compounds at various temperatures (100–1000 K) were obtained
based on density functional theory (DFT) and phonon calculations,
and then, these data were employed in the next learning process to
build a predictive model using various ML algorithms. The ML results
showed that the model built based on the random forest (RF) algorithm
with an R2 score of 0.957 was the most
accurate compared with the models built using other algorithms. Additionally,
the accuracy of this model was validated using new cases of four compounds,
which was not seen for the model before, where a good matching between
calculated and predicted LTCs of the new compounds was found. To find
candidates with ultralow LTCs (<1 W m–1 K–1) at room temperature, the model was used to screen
compounds (32116) in the Inorganic Crystal Structure Database. From
the screened compounds, Cs2SnI6 and SrS were
selected to validate the ML prediction using the counterpart theoretical
calculations (DFT and phonon), and it was found that the outcome behaviors
by both methods (ML prediction and DFT/phonon calculations) are fairly
consistent. Considering the type of employed feature, the prime novelty
in this work is that the built model can credibly predict the LTC–temperature
behaviors of new compounds that are constructed based on prototype
structures and chemical compositions, without the use of any DFT-relaxed
structure parameters. Accordingly, using the periodic table, prototype
structures, and the RF-based model, the LTC–temperature behavior
of a huge number of compounds can be predicated.