EXTENDED ABSTRACT: Machine learning (ML) based regression techniques have become a prominent tool to construct accurate interatomic potentials for materials modeling and simulations. This allows us to explore the configuration space much faster and more thoroughly than before, and enables theoretical studies that we could only have dreamed of 10 years ago. However, construction of suitable training structures is usually a fairly time-consuming trial and error process based on intuition. The thus obtained training datasets are normally huge and might contain unnecessary structures outside the phase space of interest, thus reducing the accuracy of the generated ML potential. In this talk, I will present the efficient on-the-fly active learning method based on Bayesian linear regression, through which the training structures are automatically generated during first-principles (FP) molecular dynamics (MD) simulations without the need of human intervention, while retaining near FP accuracy. To speed up the simulations, we develop the moment tensor ML potential originally developed by Shapeev through optimizing the basis sets using the generic algorism and simulated annealing. The optimized moment tensor ML potential shows a speedup by almost a factor of two for a comparable accuracy. Based on this, we further develop a code that integrates the active-learning training structures construction, ML model regression and validation. Using the developed code, I will present a few selected applications in alloys and ab initio modelings of anodic dissolution and cathodic process for
hydrogen evolution reaction.