2017 Annual Meeting
(747c) Efficient Phase Diagram Sampling By Active Learning
We propose and implement an active learning-based efficient phase diagram sampling method, which can save upwards of an order of magnitude of state point exploration time. This adaptive sampling method is inspired by advances in machine learning research. It works by adaptively calculating the potentially most informative state points by considering their proximity to explored state points and the estimated phase boundary location. It incorporates little computational overhead and is easily generalizable to cases of high state space dimensionality. Batch sampling is also supported to take advantage of modern parallel test settings, such as supercomputing clusters. We demonstrate its use, and compare its results to traditional fixed-grid sampling. This proposed method can assist our research community by saving time and computational resources.