Machine learning force fields (MLFFs) have the unique ability to combine the accuracy of quantum mechanical methods such as density functional theory (DFT) with a high degree of computational efficiency. This unparalleled combination allows for the extension of length and time scales far beyond what is feasible with DFT while making little to no sacrifices pertaining to predictive power.
In this project we will develop machine learning force fields for defects in semiconductors and isolators, which are key for applications ranging from solar cells to single-photon emitters in quantum information science. The high computational cost of DFT and similar first-principles methods severely limits the system sizes and time scales we can probe, typically leading to artifacts which are difficult to control.
To facilitate this, GPUs are key. State-of-the art MLFF implementations make use of GPU acceleration for both training and inference, allowing us to increase both the predictive power of the models and decrease their uncertainties. The faster training and inference that GPU acceleration affords us allows for the training of several models, allowing us to find better hyperparameter sets. Having several models for a given physical system also allows us to make use of ensemble techniques, allowing us to perform active learning and get better estimates of model performance. It also allows us to do bootstrap aggregation, further improving model performance.