MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures

🌈 Abstract

The paper presents MatterSim, a deep learning model for accurately predicting materials properties across a wide range of chemical compositions, temperatures, and pressures. Key highlights:

• MatterSim leverages deep graph neural networks, active learning, and large-scale first-principles computations to explore the vast materials space and achieve up to 10-fold increase in prediction accuracy compared to previous best-in-class models.

• MatterSim serves as a universal machine learning force field, enabling efficient and precise simulations of materials' ground-state energetics, lattice dynamics, mechanical and thermodynamic properties under realistic conditions.

• The model can accurately predict temperature- and pressure-dependent free energies of a wide range of inorganic solids, opening opportunities for fast and accurate phase diagram predictions.

• MatterSim supports continuous learning and customization, achieving high data efficiency with up to 97% reduction in data requirements for fine-tuning to desired levels of theory or direct structure-to-property predictions.

🙋 Q&A

[01] Learning the materials space under first-principles supervision

1. How does MatterSim explore the extensive materials space? MatterSim employs an active learning approach that integrates a deep graph neural network, a materials explorer, a first-principles supervisor, and an ensemble uncertainty monitor. The materials explorer gathers structures from both equilibrium and off-equilibrium conditions, covering a wide range of temperatures and pressures. The first-principles supervisor provides supervision signals to train the deep learning model, which then guides the materials explorer to sample the most uncertain regions of the materials space.

2. How does the dataset generated by MatterSim differ from existing materials databases? Compared to existing databases, the dataset generated by MatterSim has significantly better coverage of the chemical and structural space. It contains more distinct atomic environments across the periodic table, including better representation of noble gas elements. The dataset also includes a wider range of temperatures (0-5000K) and pressures (0-1000 GPa), going beyond the typical relaxed structures near local energy minima.

3. What are the key architectural features of the models used in MatterSim? MatterSim utilizes two primary model architectures - M3GNet and Graphormer. M3GNet is an efficient graph neural network model with high data efficiency, while Graphormer is a more complex transformer-based model with better scalability and accuracy, but higher computational cost. The choice between the two models depends on the specific task requirements in terms of speed and accuracy.

[02] MatterSim as a zero-shot atomistic emulator

1. How does MatterSim perform as a universal machine learning force field? MatterSim demonstrates remarkable zero-shot capabilities in predicting energies, forces, and stresses of materials across the periodic table under a wide range of temperatures (0-5000K) and pressures (0-1000 GPa). It outperforms previous best-in-class force fields by an order of magnitude in accuracy, especially on datasets sampled from high-temperature and high-pressure conditions.

2. How does MatterSim enable materials discovery? By leveraging MatterSim's accurate prediction of ground-state energies, the paper demonstrates the model's capability in high-throughput screening of new materials. Using random structure search, MatterSim identified 16,399 stable structures, with 1,974 of them being newly discovered, on the combined energy convex hull formed by existing databases and the newly generated structures.

3. What are the key capabilities of MatterSim in predicting lattice dynamics, mechanical properties, and thermodynamics? MatterSim achieves high accuracy in predicting phonon spectra, bulk modulus, and temperature/pressure-dependent free energies of a wide range of inorganic solids. The model's predictions are comparable to first-principles methods and experimental measurements, enabling fast and accurate construction of phase diagrams for materials.

[03] MatterSim as an active learner and with arbitrary level of theory

1. How does MatterSim leverage active learning to enhance its performance on complex systems? MatterSim provides uncertainty quantification through an ensemble of models, allowing it to identify regions of high uncertainty in the simulation trajectory. By selectively incorporating a small fraction of the high-uncertainty data points as additional training, MatterSim can reach the same level of accuracy as a model trained from scratch using the full dataset, demonstrating high data efficiency.

2. How can MatterSim be fine-tuned to achieve higher levels of theory? The paper demonstrates that by fine-tuning MatterSim with only 30 configurations computed at the rev-PBE0-D3 level of theory, the model can accurately reproduce the structural and dynamical properties of liquid water, matching the performance of a model trained from scratch using 900 configurations. This showcases MatterSim's ability to be efficiently customized to higher levels of theory beyond the PBE functional used in the initial training.