Probabilistic prediction of material stability: integrating convex hulls into active learning

Area of Science:

• Materials Science
• Computational Chemistry
• Thermodynamics

Background:

• Active learning accelerates exploration of complex chemical spaces.
• Traditional active learning struggles with global properties like convex hulls for phase diagrams.
• Material stability depends on competing phases, not just individual energies.

Purpose of the Study:

• Develop a novel active learning algorithm for efficient convex hull determination.
• Address the global nature of phase diagram prediction in materials science.
• Integrate uncertainty quantification into materials discovery workflows.

Main Methods:

• Introduced convex hull-aware active learning (CAL), a Bayesian algorithm.
• CAL prioritizes compositions near or on the convex hull.
• Quantifies uncertainty in convex hull predictions and subsequent thermodynamic properties.

Main Results:

• CAL significantly reduces the number of observations needed to predict convex hulls.
• Identifies irrelevant compositions quickly, focusing experiments.
• Provides robust uncertainty quantification for stability and chemical potential predictions.

Conclusions:

• CAL enhances search efficiency in materials science by focusing on critical data points.
• Enables reliable thermodynamic predictions within uncertainty-based workflows.
• Facilitates faster discovery of stable materials and phase diagrams.