📝 SummarXiv

Abstract

We report computational uncertainties in Boltzmann Transport Equation (BTE)-based lattice thermal conductivity prediction of 50 diverse semiconductors from the use of different BTE solvers (ShengBTE, Phono3Py, and in-house code) and interatomic forces. The interatomic forces are obtained either using the density functional theory (DFT) as implemented in packages Quantum Espresso and VASP employing commonly used exchange correlation functionals (PBE, LDA, PBEsol, and rSCAN) or using the pre-trained foundational machine learning forcefields trained on two different material datasets. We find that the considered BTE solvers introduce minimal uncertainties and, using the same interatomic force constants, all solvers result in an excellent agreement with each other, with a mean absolute percentage error (MAPE) of only 1%. While this error increases to around 10% with the use of different DFT packages, the error is still small and can be reduced further with the use of stringent planewave energy cutoffs. On the other hand, the differences in thermal conductivity due to the use of different exchange correlation functionals are large, with a MAPE of more than 20%. The currently available pre-trained foundational ML models predict the right trend for thermal conductivity, but the associated errors are high, limiting their applications for coarse screening of materials.