ANI-1: an extensible neural network potential with DFT accuracy at force field computational cost

Area of Science:

• Computational chemistry
• Machine learning in science
• Deep learning for molecular modeling

Background:

• Deep learning excels in pattern recognition tasks like image and speech processing.
• Accurate molecular energy prediction is crucial for simulating chemical reactions and material properties.
• Existing methods for molecular energy calculations can be computationally expensive.

Purpose of the Study:

• To introduce Accurate NeurAl networK engINe for Molecular Energies (ANI), a novel deep neural network potential for organic molecules.
• To develop a transferable and accurate model for predicting molecular energies using quantum mechanical data.
• To enable accelerated and physically relevant sampling of molecular potential surfaces.

Main Methods:

• Utilized deep neural networks trained on quantum mechanical (QM) density functional theory (DFT) calculations.
• Developed a molecular representation using atomic environment vectors (AEV) based on modified Behler-Parrinello symmetry functions.
• Introduced a Normal Mode Sampling (NMS) method for generating molecular conformations.
• Trained the ANI-1 potential on a subset of GDB databases for molecules with up to 8 heavy atoms (H, C, N, O).

Main Results:

• ANI-1 demonstrates chemical accuracy comparable to DFT calculations.
• The model was validated on larger molecular systems (up to 54 atoms) than those in the training set.
• The AEV molecular representation allows training on both configurational and conformational space.

Conclusions:

• ANI provides an accurate and transferable potential for organic molecules.
• The developed method significantly accelerates molecular simulations while maintaining physical relevance.
• ANI represents a breakthrough in applying deep learning to molecular energy prediction.