Augmenting Molecular Graphs with Geometries via Machine Learning Interatomic Potentials.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Accurate 3D molecular geometries are crucial for predicting molecular properties.
• Traditional methods like density functional theory (DFT) for obtaining these geometries are computationally expensive.
• Machine learning interatomic potential (MLIP) models offer a potential alternative for efficient geometry prediction.

Purpose of the Study:

• To investigate the use of MLIP foundation models for obtaining accurate 3D molecular geometries.
• To demonstrate the effectiveness of MLIP models in reducing the need for expensive DFT calculations.
• To show how MLIP-generated geometries can enhance downstream molecular property predictions.

Main Methods:

• Curated a large-scale dataset of 3.5 million molecules and 300 million snapshots for molecular relaxation.
• Trained MLIP foundation models using supervised learning to predict energy and forces from 3D molecular structures.
• Employed MLIP models for geometry optimization to obtain low-energy 3D geometries and introduced geometry fine-tuning to mitigate biases.

Main Results:

• Developed MLIP foundation models capable of generating accurate 3D molecular geometries.
• Demonstrated that MLIP models can be used explicitly for geometry optimization or implicitly through fine-tuning.
• Showcased that MLIP-generated relaxed geometries improve downstream molecular property predictions.

Conclusions:

• MLIP foundation models trained on relaxation data provide valuable molecular geometries.
• These MLIP-derived geometries can significantly benefit downstream property prediction tasks.
• MLIP models present a computationally efficient approach for obtaining essential 3D molecular structures.