On-the-fly training of polynomial machine learning potentials in computing lattice thermal conductivity (2401.17531v3)

Abstract: The application of first-principles calculations for predicting lattice thermal conductivity (LTC) in crystalline materials, in conjunction with the linearized phonon Boltzmann equation, has gained increasing popularity. In this calculation, the determination of force constants through first-principles calculations is critical for accurate LTC predictions. For material exploration, performing first-principles LTC calculations in a high-throughput manner is now expected, although it requires significant computational resources. To reduce computational demands, we integrated polynomial machine learning potentials on-the-fly during the first-principles LTC calculations. This paper presents a systematic approach to first-principles LTC calculations. We designed and optimized an efficient workflow that integrates multiple modular software packages. We applied this approach to calculate LTCs for 103 compounds of the wurtzite, zincblende, and rocksalt types to evaluate the performance of the polynomial machine learning potentials in LTC calculations. We demonstrate a significant reduction in the computational resources required for the LTC predictions.

• The paper demonstrates a novel workflow integrating polynomial machine learning potentials into LTC computations, significantly reducing computational costs.
• The methodology trains MLPs on a small, representative dataset to accurately reproduce force constants for supercells using symmetry reductions.
• Results on 103 binary compounds validate the approach, showing that as few as 20 training supercells yield reliable lattice thermal conductivity predictions.

On-the-fly Training of Polynomial Machine Learning Potentials in Computing Lattice Thermal Conductivity

Introduction

The paper presents a novel methodology for reducing the computational demands associated with first-principles calculations of lattice thermal conductivity (LTC) in crystalline materials. The approach is based on integrating polynomial machine learning potentials (MLPs) on-the-fly within the first-principles LTC computational framework. This work focuses on systematically designing a workflow that leverages these MLPs to enhance efficiency and decrease resource usage while maintaining computational accuracy for LTC predictions.

Methodology

The traditional approach to computing LTC involves calculating supercell force constants through first-principles computations, which are computationally expensive. Here, the authors propose an alternative strategy:

1. Initial Setup: Generation of displacement-force datasets from supercells using first-principles calculations.
2. Training MLPs: Polynomial MLPs are trained with a small subset of displacement-force data derived from these first-principles calculations.
3. Efficient Data Generation: The trained MLPs are used to compute forces on a larger scale displacement-force dataset, which is then employed to calculate supercell force constants.
4. LTC Computation: Uses the derived supercell force constants to calculate the LTC values, involving solving the phonon Boltzmann equation using the phono3py code.

Computational Workflow

The computational workflow is designed to encapsulate the above methodology and is implemented in a modular fashion. Key steps include:

• Data Preparation: Generation of supercells with random atomic displacements and calculation of forces from first-principles methods.
• MLP Training: Use polynomial MLPs to efficiently approximate the energy surfaces represented by crystal potentials.
• High-throughput LTC Calculations: Utilize the symmetry-reduced force constant calculations for computational efficiency, leveraging the symfc and spglib codes for symmetry handling.

The proposed workflow integrates different software packages and ensures high modularity, allowing for straightforward user deployment and efficient execution.

Results and Discussion

The efficiency of the proposed strategy is demonstrated by computing LTCs for 103 binary compounds with different crystal structures (wurtzite, zincblende, and rocksalt). Key findings include:

• Performance of Polynomial MLPs: The polynomial MLP approach significantly reduces computational demands compared to the conventional methods, especially at the stage of force constants calculation.
• Accuracy: LTC values obtained with MLPs are consistent with those calculated using high-fidelity datasets via first-principles methods.
• Efficiency: Empirical evidence suggests that a reduced number of training supercells (as few as 20) is sufficient for accurate LTC predictions, showcasing the efficiency gains of this approach.

Implications and Future Directions

This research provides a promising direction for large-scale high-throughput LTC calculations. The integration of polynomial MLPs minimizes computational overhead while offering practical accuracy, suggesting its utility for rapid screening in materials discovery.

Future work could involve automating dataset preparation and exploring more complex compounds to extend the scope of application. Moreover, the development and dissemination of high-quality datasets could further enhance accessibility and applicability in various research settings.

Conclusion

The study introduces a systematic approach that efficiently incorporates polynomial MLPs into LTC calculations, achieving significant computational savings while ensuring high accuracy. This work paves the way for scalable, high-throughput computational materials characterization, emphasizing the synergy between machine learning models and first-principles simulations.