Denoise Pretraining on Nonequilibrium Molecules for Accurate and Transferable Neural Potentials.

Area of Science:

• Computational chemistry
• Machine learning for molecular modeling

Background:

• Equivariant graph neural networks (GNNs) offer fast surrogate models for expensive quantum mechanics (QM) calculations.
• Developing accurate and transferable GNN potential models is hindered by limited QM data, especially for complex systems.

Purpose of the Study:

• To introduce a novel denoising pretraining strategy for enhancing GNN-based molecular potential predictions.
• To improve the accuracy and transferability of GNN potential models using limited computational resources.

Main Methods:

• Perturbing atomic coordinates of nonequilibrium molecular conformations with random noise.
• Pretraining GNNs to denoise these perturbed conformations, thereby recovering original coordinates.
• Evaluating performance across various molecular systems and GNN architectures.

Main Results:

• Pretraining significantly boosts the accuracy of neural potentials.
• The denoising pretraining approach is model-agnostic, enhancing both invariant and equivariant GNNs.
• Models pretrained on small molecules show remarkable transferability to diverse systems (different elements, charged, bio-, and large molecules).

Conclusions:

• Denoising pretraining is an effective strategy for improving GNN potential accuracy and generalizability.
• This method addresses data limitations in computational chemistry, enabling more efficient molecular modeling.
• The approach holds significant potential for developing robust neural potentials for complex molecular systems.