Multi-objective Optimization for Materials Discovery via Adaptive Design.

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning

Background:

• Materials discovery often involves optimizing multiple competing properties simultaneously.
• Identifying new materials requires efficient experimental or computational strategies to improve upon existing data points on the Pareto front (PF).

Purpose of the Study:

• To evaluate the effectiveness of optimal learning strategies for identifying materials with targeted properties.
• To compare the performance of different optimal learning designs (Maximin, Centroid) against baseline methods for Pareto front optimization.

Main Methods:

• Applied optimal learning concepts and methods to three distinct materials datasets: experimental shape memory alloys, computational M2AX phases, and computational piezoelectric compounds.
• Compared Maximin and Centroid design strategies, based on value of information criteria, with random selection, pure exploitation, and pure exploration.

Main Results:

• Maximin and Centroid strategies demonstrated superior efficiency in identifying Pareto front points compared to random selection, exploitation, or exploration.
• The Maximin algorithm consistently outperformed other methods across all tested datasets, regardless of size or origin.
• Maximin's robustness was particularly evident with less accurate machine learning models, indicating forgiveness towards surrogate model inaccuracies.

Conclusions:

• Optimal learning, specifically the Maximin strategy, is a highly effective approach for accelerating materials discovery and design.
• This method efficiently guides the search for materials with improved, often competing, properties, reducing the number of necessary experiments or calculations.
• The Maximin algorithm offers a reliable and forgiving design strategy for Pareto front optimization in materials science.