Accelerating Phonon Thermal Conductivity Prediction by an Order of Magnitude Through Machine Learning-Assisted Extraction of Anharmonic Force Constants (2409.00360v1)

Abstract: The calculation of material phonon thermal conductivity from density functional theory calculations requires computationally expensive evaluation of anharmonic interatomic force constants and has remained a computational bottleneck in the high-throughput discovery of materials. In this work, we present a machine learning-assisted approach for the extraction of anharmonic force constants through local learning of the potential energy surface. We demonstrate our approach on a diverse collection of 220 ternary materials for which the total computational time for anharmonic force constants evaluation is reduced by more than an order of magnitude from 480,000 cpu-hours to less than 12,000 cpu-hours while preserving the thermal conductivity prediction accuracy to within 10%. Our approach removes a major hurdle in computational thermal conductivity evaluation and will pave the way forward for the high-throughput discovery of materials.

• The paper introduces an ML-assisted method that extracts anharmonic force constants, reducing DFT computation time by over 90%.
• The technique employs a Gaussian Approximation Potential with SOAP descriptors to accurately model atomic interactions in complex materials.
• Benchmarking on 220 materials and a Type-I clathrate demonstrates that the approach maintains thermal conductivity prediction errors within 10% of traditional methods.

Accelerating Phonon Thermal Conductivity Prediction through Machine Learning

This paper presents a ML-assisted approach to significantly reduce the computational time required to predict phonon thermal conductivity while maintaining accuracy. The authors introduce a method that utilizes ML to extract anharmonic force constants (IFCs) from density functional theory (DFT) calculations in a more efficient manner than traditional approaches.

Introduction and Motivation

Current methods to determine phonon thermal conductivity ($\kappa$) are computationally intensive due to the need for evaluating anharmonic IFCs using DFT. This process represents a bottleneck in high-throughput material discovery aimed at optimizing $\kappa$ for applications such as thermoelectrics and thermal barrier coatings. Recent ML approaches for direct $\kappa$ prediction suffer from accuracy limitations due to a lack of high-quality datasets. This paper proposes an alternative strategy that combines DFT with ML to accelerate the extraction of IFCs, thereby facilitating faster and scalable predictions of $\kappa$.

Methodology Overview

The phonon contribution to thermal conductivity, computed via the Boltzmann Transport Equation (BTE), requires accurate IFCs. The traditional method involves time-consuming DFT calculations to obtain these constants using the finite-difference approach. The proposed ML technique focuses on learning the local potential energy surface efficiently, using a Gaussian Approximation Potential (GAP) model with two-body, three-body, and SOAP descriptors. This approach allows for the accurate representation of atomic interactions with a significantly reduced number of DFT calculations.

Results and Performance

Efficiency and Accuracy

The ML-assisted approach was evaluated on 220 diverse ternary materials, demonstrating a reduction in computation time from 480,000 cpu-hours to less than 12,000 cpu-hours, with accuracy preserved to within 10% of traditional DFT-driven methods. The study also highlights improvements in $\kappa$ prediction when the ML model is trained on thermally populated configurations, resulting in a significant reduction in mean absolute percentage error (MAPE) for IFC extraction.

Case Study: Type-I Clathrate

For the complex BGG type-I clathrate, the ML-assisted approach reduced computational time from 47,000 to 5,000 cpu-hours while maintaining $\kappa$ prediction accuracy to within 4%. This demonstrates the method's potential for applications in materials with intricate crystal structures.

Higher-Order Thermal Transport Predictions

The study extends to exploring higher-order thermal transport phenomena, confirming the applicability of the ML approach in scenarios requiring iterative BTE solutions, multi-phonon interactions, and coherence effects in heat transfer. The MAPE for these advanced scenarios remained below 10%, underscoring the robustness of the methodology.

Comparison with Other Approaches

The paper contrasts the proposed ML strategy with other ML models and regularization techniques, such as hiPhive and MLIP, showing superior performance in terms of accuracy and computational efficiency. The locally trained SOAP-GAP model demonstrates the lowest error margins and best captures the local potential energy landscapes required for accurate thermal conductivity predictions.

Conclusion

The introduction of an ML-driven framework for anharmonic IFC extraction marks significant progress in reducing the computational overhead associated with phonon thermal conductivity predictions. By preserving prediction accuracy and drastically cutting down computational demands, this method paves the way for enhanced high-throughput material discovery processes, enabling the rapid exploration and optimization of material properties for thermal applications. Future developments will further refine harmonic IFC extraction processes, promising comprehensive solutions in computational materials science.