Researchers Accurately Predict Phonon Properties in 2,429 Materials Using Universal ML Potentials

Predicting how heat moves through materials is crucial for designing everything from efficient electronics to advanced thermal insulators, and recent advances in machine learning offer promising new tools for this task. Md Zaibul Anam, Ogheneyoma Aghoghovbia, and Mohammed Al-Fahdi, all from the University of South Carolina, alongside Lingyu Kong and Victor Fung from Georgia Institute of Technology, and Ming Hu, present a comprehensive assessment of several cutting-edge machine learning models, known as large atomistic foundation models, for predicting phonon properties, the vibrations that govern heat transfer. Their work systematically benchmarks these models against a vast database of crystalline materials, revealing significant variations in performance and highlighting the complex relationship between predicting atomic forces and accurately forecasting thermal conductivity. This detailed comparison provides researchers with essential guidance for selecting the most appropriate machine learning tools for high-throughput materials discovery and the design of materials with tailored thermal properties.

Machine Learning Potentials for Thermal Conductivity Prediction

Researchers evaluated six machine learning potentials (MLPs), including EquiformerV2, MatterSim, MACE, and CHGNet, to determine their accuracy in predicting lattice thermal conductivity, a crucial property for designing materials for thermal management applications. The study utilized a dataset of 2,429 materials to comprehensively assess performance, comparing predictions against calculations based on density functional theory and available experimental data. They assessed the models’ ability to predict atomic forces and interatomic force constants, fundamental to understanding material vibrations and thermal transport. Results demonstrate that EquiformerV2 consistently achieved the highest accuracy across most metrics, making it the most reliable MLP for predicting lattice thermal conductivity. Other models within the EquiformerV2 family also performed well, generally outperforming MatterSim, MACE, and CHGNet. This research demonstrates that MLPs can significantly accelerate materials design by providing a fast and accurate method for predicting thermal conductivity, enabling high-throughput screening of materials with targeted thermal properties.

Predicting Phonon Properties with Machine Learning Potentials

Researchers systematically evaluated six universal machine learning interatomic potentials (uMLPs), including EquiformerV2, MatterSim, MACE, and CHGNet, to assess their ability to predict phonon properties across a diverse range of crystalline materials. Employing a benchmark dataset of 2,429 materials, they computed atomic forces within displaced supercells, deriving interatomic force constants (IFCs) crucial for understanding material vibrations. The team meticulously compared predictions from each uMLP against both density functional theory (DFT) calculations and experimental data, establishing a rigorous standard for evaluation. EquiformerV2, pretrained on a large dataset, demonstrated strong performance in predicting both atomic forces and third-order IFCs, indicating its ability to capture complex atomic interactions accurately. A fine-tuned version consistently outperformed other models in predicting second-order IFCs, lattice thermal conductivity, and other critical phonon properties. Interestingly, MatterSim, despite showing lower force accuracy, achieved intermediate IFC predictions, suggesting a complex relationship between force accuracy and phonon predictions.

EquiformerV2 Predicts Phonon Properties Accurately

Researchers are rapidly advancing universal machine learning potentials (uMLPs) to efficiently and accurately predict material properties. A recent study systematically benchmarks six prominent uMLPs, including EquiformerV2, MatterSim, MACE, and CHGNet, using a dataset of 2,429 crystalline materials. The investigation focused on calculating atomic forces, deriving interatomic force constants (IFCs), and ultimately predicting phonon properties, including lattice thermal conductivity, comparing results with both density functional theory (DFT) and experimental data. Results demonstrate that EquiformerV2 consistently achieved the highest accuracy across most metrics, making it a state-of-the-art approach for predicting phonon properties. The model’s ability to accurately predict atomic forces and third-order IFCs contributed to its superior performance. While MACE and CHGNet achieved comparable force prediction accuracy, notable discrepancies in IFC fitting led to poorer thermal conductivity predictions.

EquiformerV2 Excels at Phonon Property Prediction

This study presents a comprehensive benchmark of six universal machine learning potentials (uMLPs), including EquiformerV2, MatterSim, MACE, and CHGNet, for predicting phonon properties of crystalline materials. Researchers evaluated the models’ ability to predict atomic forces, interatomic force constants, and ultimately, lattice thermal conductivity, comparing results with both density functional theory calculations and experimental data across a large database of materials. The findings demonstrate that EquiformerV2 consistently outperforms the other models in predicting these phonon properties, including identifying dynamically stable materials and accurately predicting lattice thermal conductivity. The research highlights the critical role of the training dataset in determining the accuracy of uMLPs, with models trained on a large dataset generally performing better. MatterSim showed complex relationships between force accuracy and phonon predictions, achieving intermediate results despite lower force prediction accuracy.