| Abstract Scope |
Reinforcement learning (RL) methods offer a novel approach to identifying unique materials for the energy sector. This work presents an RL-based strategy for alloy design that addresses limitations of current approaches. While Bayesian optimization dominates recent alloy design research, it is inherently slow and constrained by its underlying model, particularly struggling with high-dimensional optimization. Our RL-based approach transforms single-point optimization into sequential decision-making, enabling identification and evaluation of multiple solutions for each design challenge. The sequential nature of RL enhances explainability through visualization of element selection patterns across episodes. We introduce a customized alloy design environment and benchmark our RL strategy against state-of-the-art optimization methods. Experimental results demonstrate the effectiveness of RL-based predictions for our energy sector use cases. This framework provides a more flexible and interpretable alternative to traditional optimization approaches, with particular advantages in exploring complex, high-dimensional alloy design spaces. |