Machine Learning for Predicting Lattice Thermal Conductivity

This thesis presents a machine learning–accelerated approach for computing anharmonic interatomic force constants (IFCs), which are essential for predicting lattice thermal conductivity in crystalline materials. Traditionally, the calculation of these force constants requires a large number of computationally expensive single-point Density Functional Theory (DFT) evaluations. To address this computational bottleneck, Machine Learning Interatomic Potentials (MLIPs) are employed using the PANNA (Properties from Artificial Neural Network Architectures) framework. The layered chalcogenide GeS2 is used as a representative case study. During model development, an important limitation of standard MLIP training strategies is identified. Neural networks trained only on small structural perturbations tend to remain confined within a purely harmonic regime, preventing them from learning the highly distorted configurations that govern anharmonic lattice dynamics. This phenomenon is referred to as the harmonic trap. To overcome this limitation, a Query-by-Committee active learning workflow is implemented and combined with ab initio Molecular Dynamics (AIMD) data to expand the sampled configuration space. The resulting hybrid dataset enables the neural network potential to accurately represent both harmonic and anharmonic regions of the potential energy surface. As a result, the computational time required to evaluate 212 supercell configurations is successfully reduced from 55.80 node-hours of rigorous DFT calculations to just 7.03 seconds using the trained surrogate model. This corresponds to an extraordinary computational speedup of nearly four orders of magnitude. Furthermore, a systematic sensitivity analysis demonstrates that lattice thermal conductivity calculations are extremely sensitive to force prediction noise. To avoid artificial phonon scattering in the Boltzmann Transport Equation (BTE) solver, force prediction errors must remain below a critical threshold of ∼ 50µeV/Å. The proposed active learning framework successfully achieves this level of physical accuracy while maintaining substantial computational efficiency. Overall, this work establishes a robust methodology for accelerating anharmonic lattice dynamics calculations and enabling high-throughput screening of thermoelectric materials.