Accelerating Materials Discovery With Active Thermodynamic Surface Shape Mapping

Automated materials discovery often involves learning the locations of shape features on thermodynamic surfaces that are costly to observe. For example, the phase diagram of a novel material may be estimated by locating ridges on a simulated heat capacity surface. Traditional acquisition functions used in active learning are sample-efficient to learn the locations of optima or level sets, but they may not be efficient for learning arbitrary shape features, such as heat capacity ridges. We propose a new acquisition function, Expected Shape feature Dispersion Improvement (ESDI) to learn arbitrary shape features on surfaces. We evaluate ESDI on a phase diagram learning task. ESDI consistently outperforms three competitor approaches across multiple metrics. These initial results position ESDI as a promising tool to accelerate materials design through faster surface learning.