Abstract:A Model capable of handling various elemental species and substances is essential for discovering new materials in the vast phase and compound space. Message-passing neural networks (MPNNs) are promising as such models, in which various vector operations model the atomic interaction with its neighbors. However, conventional MPNNs tend to overlook the importance of physicochemical information for each node atom, relying solely on the geometric features of the material graph. We propose the new three-body MPNN architecture with a message-passing layer that utilizes optimized basis functions based on the electronic structure of the node elemental species. This enables conveying the message that includes physical information and better represents the interaction for each elemental species. Inspired by the LCAO (linear combination of atomic orbitals) method, a classical method for calculating the orbital interactions, the linear combination of atomic basis constructed based on the wave function of hydrogen-like atoms is used for the present message-passing. Our model achieved higher prediction accuracy with smaller parameters than the state-of-the-art models on the challenging crystalline dataset containing many elemental species, including heavy elements. Ablation studies have also shown that the proposed method is effective in improving accuracy. Our implementation is available online.

What problem does this paper attempt to address?

The paper aims to address the problem of predicting new material properties in materials science, particularly the challenges faced when dealing with large-scale crystal structures containing multiple elements. Specifically, the paper proposes a new architecture called LCAONet, which is a method based on Message Passing Neural Networks (MPNN) to better simulate interactions between atoms. Traditional MPNNs often rely solely on the geometric features of the material graph when modeling atomic interactions, neglecting the physicochemical information of each node atom. LCAONet addresses this issue by introducing a new message-passing layer that uses optimized basis functions based on the electronic structure of node elements to transmit information. This allows the network to convey messages containing physical information and better represent interactions between different elements. LCAONet is inspired by the Linear Combination of Atomic Orbitals (LCAO) method, a classical computational method for orbital interactions. Building on this, the paper proposes a new set of radial basis functions constructed based on hydrogen-like atomic wave functions to more accurately characterize atomic interactions. Additionally, the number and type of basis functions are optimized according to the electronic structure of the node atoms, thereby improving the parameter efficiency of the model and introducing useful inductive biases. To validate the effectiveness of LCAONet, the authors conducted experiments on datasets of periodic systems with multiple element types, including heavy elements (atomic number up to 83). The experimental results show that LCAONet exhibits higher accuracy in predicting formation energies compared to current state-of-the-art models, especially when dealing with heavy elements and hypothetical structures. Furthermore, LCAONet achieves better performance with significantly fewer parameters than other models, indicating that its designed basis functions and message-passing mechanism have advantages in parameter efficiency. In summary, this study improves the MPNN model by combining physicochemical information and message-passing techniques, enabling it to more effectively predict the quantum mechanical properties of complex crystal systems.

LCAONet: Message-passing with physically optimized atomic basis functions

Long-Short-Range Message-Passing: A Physics-Informed Framework to Capture Non-Local Interaction for Scalable Molecular Dynamics Simulation

Long-Short-Range Message-Passing: A Fragmentation-Based Framework to Capture Non-Local Atomistic Interactions

ABT-MPNN: an atom-bond transformer-based message-passing neural network for molecular property prediction

Global Neural Network Potential with Explicit Many-Body Functions for Improved Descriptions of Complex Potential Energy Surface.

Ewald-based Long-Range Message Passing for Molecular Graphs

A neural network approach to running high-precision atomic computations

Including physics-informed atomization constraints in neural networks for reactive chemistry

Accuracy Assessment of Atomistic Neural Network Potentials: The Impact of Cutoff Radius and Message Passing

Accurate and scalable multi-element graph neural network force field and molecular dynamics with direct force architecture

Many-Body Function Corrected Neural Network with Atomic Attention (Mbnn-Att) for Molecular Property Prediction.

AIMNet2: A Neural Network Potential to Meet your Neutral, Charged, Organic, and Elemental-Organic Needs

Stratified construction of neural network based interatomic models for multicomponent materials

Efficient Integration of Molecular Representation and Message-Passing Neural Networks for Predicting Small Molecule Drug-like Properties

A Theoretical Formulation of Many-body Message Passing Neural Networks

A Message Passing Neural Network Surrogate Model for Bond-Associated Peridynamic Material Correspondence Formulation

Improving the Performance of Long-Range-Corrected Exchange-Correlation Functional with an Embedded Neural Network

Neural Atoms: Propagating Long-range Interaction in Molecular Graphs through Efficient Communication Channel

A2I Transformer: Permutation-equivariant attention network for pairwise and many-body interactions with minimal featurization

A Physics-Aware Neural Network for Protein-Ligand Interactions with Quantum Chemical Accuracy

Cartesian atomic cluster expansion for machine learning interatomic potentials