Autonomous atomic Hamiltonian construction and active sampling of x-ray absorption spectroscopy by adversarial Bayesian optimization

X-ray absorption spectroscopy (XAS) is a well-established method for in-depth characterization of the electronic structure due to its sensitivity to the local coordination and electronic states of the active ions. In practice hundreds of energy points should be sampled during the XAS measurement, most of which are redundant and do not contain important information. In addition, it is also a tedious procedure to estimate reasonable parameters in the atomic Hamiltonian for mechanistic understanding. We implemented an Adversarial Bayesian optimization (ABO) algorithm comprising two coupled BOs to automatically fit the multiplet model Hamiltonian and meanwhile to sample effectively based on active learning. Taking NiO as an example, for simulated spectra which can be well fitted by the atomic model, we found that less than 30 sampling points are enough to obtain the complete XAS with the corresponding crystal field or charge transfer model, which can be selected based on intuitive hypothesis learning. Further application on the experimental spectra, it revealed that less than 80 sampling points can already give reasonable XAS and reliable atomic model parameters. Our ABO algorithm has a great potential for future application in automated physics-driven XAS analysis and active learning sampling.