Recently, there has been a growing interest on AI technology application to material science. In this presentation, we introduce our recent activities to accelerate scientific computations which can emulates wet-lab experiments. The first one is to accelerate ab-initio MD simulations via machine learning interatomic potential (MLIP) models; We propose a new uncertainty-based active learning approach that enables to train MLIP model with the minimal amount of the first-principles computation, such as DFT. The second one is to accelerate Hamiltonian Monte-Carlo simulations. We proposed a self-tuning method to identify the most efficient and accurate hyper-parameters for the simulation, which enables us to accelerate 25% speedup with requiring 100 times faster tuning time. Finally, we also introduce our latest physics-informed weakly supervised learning approach for MLIP models.
Makoto Takamoto