Development of Multimodal Machine Learning Potentials: Toward a Physics-Aware Artificial Intelligence.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine learning interatomic potentials (MLIPs) bridge the accuracy gap between quantum mechanics and classical force fields.
• Developing transferable MLIPs is crucial for diverse chemical tasks.

Purpose of the Study:

• To present out-of-the-box approaches for developing transferable MLIPs.
• To introduce and discuss the capabilities of ANAKIN-ME (ANI), AIMNet, and ML-EHM models.

Main Methods:

• Utilized vast datasets for training the ANI model with Justin Smith Symmetry Functions.
• Developed AIMNet inspired by quantum theory of atoms in molecules, encoding long-range interactions and element representations.
• Introduced ML-EHM, a physics-aware model combining ML with the extended Hückel method.

Main Results:

• ANI demonstrates knowledge transferability and flexibility through large-scale training.
• AIMNet-ME extends applicability to open-shell systems, accurately predicting molecular energy dependence on charge and spin.
• ML-EHM correctly describes physical phenomena like orbital crossing, showcasing improved generalization through physics-informed constraints.

Conclusions:

• Advanced methods like active learning, transfer learning, and multitask learning can optimize MLIP training data selection.
• Incorporating physical constraints and symmetries significantly enhances ML model generalization for interatomic potentials.