Fast approximations by machine learning: predicting the energy of dimers using convolutional neural networks

Abstract

We introduce fast approximations by machine learning (FAML) to compute the energy of molecular systems. FAML can be six times faster than a traditional quantum chemistry approach for molecular geometry optimisation, at least for a simple dimer. Hardware accelerators for machine learning (ML) can further improve FAML’s performance. Since the quantum chemistry calculations show poor algorithmic scaling, faster methods that produce a similar level of accuracy to the more rigorous level of quantum theory are important. As a FAML proof-of-concept, we use a convolutional neural network (CNN) to make energy predictions on the F2 molecular dimer system. Training data for the CNN is computed using a quantum chemistry application (i.e., GAMESS) and represented as an image. Using fivefold cross-validation, we find that the predictions made by the CNN provide a good prediction to the theoretical calculations in a fraction of the time.