Efficient Approximations of Complete Interatomic Potentials for Crystal Property Prediction (2306.10045v9)

Abstract: We study property prediction for crystal materials. A crystal structure consists of a minimal unit cell that is repeated infinitely in 3D space. How to accurately represent such repetitive structures in machine learning models remains unresolved. Current methods construct graphs by establishing edges only between nearby nodes, thereby failing to faithfully capture infinite repeating patterns and distant interatomic interactions. In this work, we propose several innovations to overcome these limitations. First, we propose to model physics-principled interatomic potentials directly instead of only using distances as in many existing methods. These potentials include the Coulomb potential, London dispersion potential, and Pauli repulsion potential. Second, we model the complete set of potentials among all atoms, instead of only between nearby atoms as in existing methods. This is enabled by our approximations of infinite potential summations, where we extend the Ewald summation for several potential series approximations with provable error bounds. Finally, we propose to incorporate our computations of complete interatomic potentials into message passing neural networks for representation learning. We perform experiments on the JARVIS and Materials Project benchmarks for evaluation. Results show that the use of interatomic potentials and complete interatomic potentials leads to consistent performance improvements with reasonable computational costs. Our code is publicly available as part of the AIRS library (https://github.com/divelab/AIRS/tree/main/OpenMat/PotNet).

• The paper presents PotNet, a graph-based deep learning model that integrates complete interatomic interactions for improved crystal property prediction.
• It leverages physics-based potentials, such as Coulomb and London dispersion, computed with Ewald summation to model infinite periodic structures.
• Experimental evaluations on JARVIS and Materials Project datasets demonstrate enhanced accuracy in formation energy and band gap predictions with modest computational overhead.

Efficient Approximations of Complete Interatomic Potentials for Crystal Property Prediction

The paper "Efficient Approximations of Complete Interatomic Potentials for Crystal Property Prediction" addresses the intricate challenge of modeling crystal structures for predicting their properties using machine learning techniques. It introduces a novel graph-based deep learning model capable of capturing complete interatomic interactions in crystal lattices.

Summary of Contributions

The paper highlights key limitations in existing approaches which typically rely on radius graphs, capturing only localized atomic interactions and ignoring the infinite periodic nature of crystal structures. This paper proposes a solution, named PotNet, which innovatively uses interatomic potential functions instead of mere atomic distances in graph neural networks (GNNs) to better represent crystal structures.

Innovations in Interatomic Potentials

1. Physics-Based Potentials: The model explicitly uses physical interatomic potentials, including Coulomb, London dispersion, and Pauli repulsion potentials, instead of using expansions based on distances alone. This leverages well-established principles from physics to enhance the accuracy of crystal property prediction.
2. Complete Interatomic Interactions: PotNet models interactions among all atoms rather than restricting them to nearby connections. This is achieved through mathematical approximations that account for the repeated crystal structure, employing Ewald summation techniques with provable error bounds.
3. Integration with GNNs: The model incorporates complete interatomic potentials into the message-passing framework of GNNs. This allows PotNet to encode 3D geometric information effectively, enabling comprehensive interaction modeling within the crystal structure.

Experimental Results

The implementation of PotNet is evaluated against benchmark datasets, namely the JARVIS and Materials Project datasets, showing consistent improvements over traditional methods in several property prediction tasks. Notably, PotNet excelled in tasks such as formation energy and band gap predictions, validating its enhanced modeling capabilities. Moreover, the paper demonstrated that PotNet achieves these improvements with reasonable computational overhead, maintaining efficiency despite the added complexity of modeling infinite potential interactions.

Implications and Future Directions

The proposed PotNet establishes a more accurate method for crystal property prediction by explicitly modeling interatomic potentials and leveraging complete potential information. This innovation may have significant practical implications in fields such as materials science, chemistry, and condensed matter physics, where precise prediction of material properties is crucial.

The potential applications of PotNet extend to the development of novel materials by allowing for accurate simulations and predictions of material properties. Future research could explore integrating PotNet with broader datasets and extending its capabilities to handle many-body interactions, possibly enabling even more precise material simulations.

The paper also opens avenues for further improvement in computational efficiency, possibly through refining the approximations used in summing infinite potentials or by integrating more advanced deep learning architectures that could allow for faster and more accurate predictions. The introduction of detailed infinite-range interactions marks a significant step forward in the field of crystal structure analysis and machine learning, providing a robust foundation for future advancements.