Abstract:Machine learning interatomic potentials (MLIPs) have substantially advanced atomistic simulations in materials science and chemistry by providing a compelling balance between accuracy and computational efficiency. While leading MLIPs rely on representations of atomic environments using spherical tensors, Cartesian representations offer potential advantages in simplicity and efficiency. In this work, we introduce Cartesian Atomic Moment Potentials (CAMP), an approach equivalent to models based on spherical tensors but operating entirely in the Cartesian space. CAMP constructs atomic moment tensors from neighboring atoms and combines these through tensor products to incorporate higher body-order interactions, which can provide a complete description of local atomic environments. By integrating these into a graph neural network (GNN) framework, CAMP enables physically-motivated and systematically improvable potentials. It requires minimal hyperparameter tuning that simplifies the training process. The model demonstrates excellent performance across diverse systems, including periodic structures, small organic molecules, and two-dimensional materials. It achieves accuracy, efficiency, and stability in molecular dynamics simulations surpassing or comparable to current leading models. By combining the strengths of Cartesian representations with the expressiveness of GNNs, CAMP provides a powerful tool for atomistic simulations to accelerate materials understanding and discovery.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to develop a new type of machine - learning interatomic potential (MLIP) to improve the accuracy and computational efficiency of atomic simulations in the fields of materials science and chemistry. Specifically, the authors propose Cartesian Atomic Moment Potentials (CAMP), a method that operates entirely in Cartesian space and aims to replace existing models that rely on spherical - tensor representations.
### Main Problems
1. **Improving the Accuracy and Efficiency of Atomic Simulations**
- Traditional machine - learning interatomic potentials (MLIPs) mainly rely on spherical - tensor representations when describing atomic environments. Although these representation methods are effective, they may be less intuitive and more complex in some cases.
- CAMP simplifies the representation method and improves computational efficiency by using atomic moment tensors in the Cartesian coordinate system to describe atomic environments.
2. **Systematic Improvement and Physical Motivation**
- Existing MLIPs perform well in molecular dynamics (MD) simulations, but there is still room for improvement. The design goal of CAMP is to provide a potentially function that can be systematically improved and physically motivated.
- CAMP effectively captures higher - order many - body interactions by introducing hyper - moments, that is, the self - interactions between atomic moment tensors, thus providing a more complete description of the local atomic environment.
3. **Reducing Hyper - parameter Tuning**
- Many existing MLIPs require a large amount of hyper - parameter tuning, which increases the complexity of the training process. CAMP simplifies the training process by reducing the need for hyper - parameter tuning through simplifying the feature representation and model architecture.
4. **Wide Applicability**
- CAMP is designed to be applicable to various material systems, including periodic structures, small organic molecules, and two - dimensional materials. Through extensive benchmark tests, the performance of CAMP in different material systems has been verified.
### Solutions
- **Cartesian Atomic Moment Tensor**: CAMP uses atomic moment tensors in the Cartesian coordinate system to describe atomic environments and combines these moment tensors through tensor products to include higher - order many - body interactions.
- **Graph Neural Network (GNN) Framework**: CAMP integrates these atomic moment tensors into the GNN framework, enabling the model to be iteratively updated and optimized, and finally forming a stable MLIP.
- **Message - passing Mechanism**: Through the message - passing mechanism, CAMP can efficiently propagate information and update atomic features at different levels, thus ensuring the accuracy and stability of the model.
### Results
- **Performance Evaluation**: CAMP performs excellently on multiple material systems, including periodic structures (such as LiPS inorganic crystals), small organic molecules, and two - dimensional materials (such as graphene). It shows higher accuracy and efficiency in molecular dynamics simulations, surpassing or being comparable to the currently leading models.
- **Long - term Stability and Diffusion Coefficient**: CAMP shows good long - term stability in molecular dynamics simulations and can accurately predict the diffusion coefficient of lithium ions in LiPS, which is consistent with the results of ab initio molecular dynamics (AIMD).
In conclusion, this paper solves the complexity and limitations of existing MLIPs in describing atomic environments by proposing CAMP, providing a new method that is more concise, efficient, and physically motivated, and significantly improving the accuracy and computational efficiency of atomic simulations.