<br/>In the last decade, machine-learning has made substantial contributions to molecular simulation. For example, the development of DFT-trained neural networks to create atomistic potentials that can be used in molecular dynamics or Monte Carlo simulations.  For performance reasons, the codes for this work has conventionally been written in compiled languages such as Fortran and C++. While nearly all the progress in this field has been made with this approach, we believe it is currently hindering more rapid progress in the field. At the core of these methods, one uses the Cartesian coordinates, atomic symbols, and unit cell of an atomic geometry as the input to these models. Then, a set of features are derived from those inputs to characterize the local environment of each atom, and a machine-learned model is trained to output the energy for a set of these feature vectors. The challenge is deriving the expressions for the forces and stress, which depend on derivatives of the model and features with respect to the positions and strain tensor. Current machine-learning frameworks such as TensorFlow provide automatic differentiation, which could alleviate the difficulty in deriving and implementing these derivatives. This will enable one to readily develop new feature vectors, and to try novel model architectures much more rapidly, and should lead to overall better machine-learned atomistic potentials. In this talk, we will present our work in implementing these ideas in TensorFlow, along with lessons we have learned about it.

Machine-learning frameworks in molecular simulation

255th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.