Nick Terrel - Atomistic uncertainty driven data generation in ANI neural network potentials | SciPy

ANI neural network potentials can reproduce DFT energies and potential energy surfaces ~1 million times faster than traditional quantum mechanical methods

The system uses an ensemble of neural networks to predict atomic energies, with each atom type (C, H, etc.) having its own dedicated network

Model accuracy targets chemical accuracy of 1 kcal/mol, though achieving this consistently remains challenging

A key innovation is the Atomic Environment Vector (AEV) which captures local atomic environments within a cutoff radius to predict energies

The method scales approximately linearly with system size, making it suitable for large molecular systems (300-400 atoms)

Uncertainty quantification is done through QBC (Query by Committee) metric, helping identify when predictions may be unreliable

Active learning approaches are used to iteratively improve the model by identifying and adding high-uncertainty configurations

The model requires no explicit physics input beyond training data, learning representations directly from quantum mechanical calculations

Current challenges include:

• Accurate uncertainty quantification at the atomic level
• Sampling new chemical spaces efficiently
• Balancing accuracy vs computational cost
• Handling long-range interactions

The method has been successfully applied to reactive molecular dynamics simulations with systems of up to 22 million atoms