Intelligent-ReaxFF: Evaluating the reactive force field parameters with machine learning

Machine learning has been widely used in quantum chemistries, such as data mining in quantum mechanics calculation and representations of potential energy surface by neural networks. In this study, we report our efforts on the optimization of the ReaxFF parameters with machine learning frameworks. Although deep neural network potentials like High-Dimensional Neural Network Potentials (HDNNP) have achieved much success in applications such as materials modeling, factors like the memory usage, training time, and accuracies are still problems when the training data set is big. On the other hand, classical potentials like ReaxFF and REBO does not have these problems, and a combination of two is an ideal solution. Machine learning has generated techniques such as automatic differentiation and backpropagation, with which we can optimize deep neural networks or complexed interatomic potentials like ReaxFF. With the TensorFlow coding platform, we have constructed an Intelligent ReaxFF (I-ReaxFF) model with terms of matrix (or tensor) operations that can optimize ReaxFF parameters automatically with gradient-based optimizers like adaptive moment solver (Adam) and back-propagations. As inherited from TensorFlow, one significant feature of our code is the GPU acceleration. The training speed can be five times faster with GPU acceleration than pure CPU calculation. Another feather is that it can directly use the ab initio molecular dynamics trajectories with surrounding periodic images as training data, therefore, allowing the data set can be prepared with ease.