MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures (2405.04967v2)

Abstract: Accurate and fast prediction of materials properties is central to the digital transformation of materials design. However, the vast design space and diverse operating conditions pose significant challenges for accurately modeling arbitrary material candidates and forecasting their properties. We present MatterSim, a deep learning model actively learned from large-scale first-principles computations, for efficient atomistic simulations at first-principles level and accurate prediction of broad material properties across the periodic table, spanning temperatures from 0 to 5000 K and pressures up to 1000 GPa. Out-of-the-box, the model serves as a machine learning force field, and shows remarkable capabilities not only in predicting ground-state material structures and energetics, but also in simulating their behavior under realistic temperatures and pressures, signifying an up to ten-fold enhancement in precision compared to the prior best-in-class. This enables MatterSim to compute materials' lattice dynamics, mechanical and thermodynamic properties, and beyond, to an accuracy comparable with first-principles methods. Specifically, MatterSim predicts Gibbs free energies for a wide range of inorganic solids with near-first-principles accuracy and achieves a 15 meV/atom resolution for temperatures up to 1000K compared with experiments. This opens an opportunity to predict experimental phase diagrams of materials at minimal computational cost. Moreover, MatterSim also serves as a platform for continuous learning and customization by integrating domain-specific data. The model can be fine-tuned for atomistic simulations at a desired level of theory or for direct structure-to-property predictions, achieving high data efficiency with a reduction in data requirements by up to 97%.

• The paper introduces MatterSim, a deep learning model that simulates atomic material properties with a ten-fold increase in prediction accuracy.
• It integrates graph neural networks with zero-shot and adaptive learning to forecast phonon frequencies, bulk modulus, and phase transitions precisely.
• Its performance on 17 million structures enables efficient materials screening and discovery, accelerating experimental validation and design.

MatterSim: A Comprehensive Deep Learning Approach for Atomistic Simulations

The presented paper introduces MatterSim, an ambitious deep learning model designed to forecast and simulate the properties of materials across the periodic table under a wide range of conditions, including various temperatures and pressures. Trained on extensive first-principles computational data, MatterSim serves as a formidable tool in materials discovery and design, embedding robust machine learning force fields (MLFF) capabilities.

Overview of MatterSim

MatterSim stands out by integrating state-of-the-art graph neural networks with innovative learning strategies. It uniquely combines zero-shot learning, adaptive learning, and customization, enabling simulations at conditions ranging from 0 to 5000 Kelvin and pressures up to 1000 GPa. The model’s remarkable proficiency is illustrated by its ten-fold precision improvement over predecessors when predicting material structures and energetics. This high level of accuracy is achieved through extensive training on 17 million labeled structures, surpassing conventional datasets in both chemical and configurational diversity.

Key Numerical Results and Predictions

MatterSim’s capabilities are exemplified through several rigorous validation tests and applications:

• Phonons and Mechanical Properties: MatterSim demonstrates an MAE of 0.87 THz in phonon frequency predictions and is effective in forecasting bulk modulus with an MAE of 2.47 GPa.
• Phase Diagrams: Its accuracy in predicting Gibbs free energy, with near-first-principles accuracy at 15 meV/atom at high temperatures, facilitates the computation of accurate phase diagrams, such as the B1-B2 phase transition in MgO.
• Materials Discovery: With a ten-fold improvement in energy prediction accuracy compared to existing models, MatterSim enables the efficient screening and discovery of new materials from over 80 million structures, identifying numerous novel stable compounds enriching the current materials database.

Implications and Future Directions

The implications of MatterSim are profound. Its ability to predict experimental phase diagrams and structural properties with minimal computational cost could accelerate the screening of new materials and the verification of their properties prior to experimental synthesis. Moreover, its adaptability through fine-tuning with minimal data from higher-level theoretical computations (e.g., hybrid functionals) highlights its potential to tailor predictions for specific applications.

The paper opens several avenues for future research. One critical area involves extending MatterSim's applicability to surface and interface modeling, which would require enriching the training dataset with surface-specific properties. The development of more sophisticated models that capture long-range interactions, particularly relevant for polymers and composite systems, is another potential next step. Furthermore, integrating semi-supervised or unsupervised pretraining could refine the model’s accuracy and generalizability further.

In conclusion, MatterSim embodies a remarkable stride forward in computational materials science. By bridging the gap between first-principles accuracy and scalability, it stands to significantly streamline the digital transformation of materials design and discovery, propelling experimental collaboration and innovation within this dynamic field.