Machine learning force field using decomposed atomic energies from ab initio calculations

The development of machine learning (ML) force field from ab initio calculation is currently an intensely studied topic. In this talk, we will present our results using both neural network model and Gaussian process regression to represent such force fields. Examples of one element systems (e.g., Si) and two element systems (e.g., Fe-H) will be represented. We show that the ML force field can reproduce well the density functional theory (DFT) ab initio results. Large scale crystal growth can be simulated using the ML force field. In our ML force field development, we have used not only the atomic forces, but also a special decomposed DFT energy on each atom to carry out the training. We deployed a systematic feature function set to describe the atomic environment of a central atom. We will also present the comparison between the neural network model and the Gaussian process regression model.