Reinforcement Learning for Cathode Material Design Through Sequential Decision-Making Frameworks

The cathode material design is a persistent challenge in the development of next-generation rechargeable batteries. The cathode performance is critically influenced by certain key parameters, i.e., composition, crystal structures, ion transport, and degradation behaviour. Moreover, techno-economic and sustainable considerations also play a pivotal role in the viable cathode material design. In recent years, the integration of static machine learning models with conventional experimental techniques has significantly enhanced the cathode material design. However, the sequential nature of cathode discovery has not been fully captured by these techniques as they do not update their decision strategy based on prior outcomes. In this review, reinforcement learning (RL) as a decision making technique for cathode material design has been evaluated. Firstly, cathode design space, including major cathode families, optimisation objectives, and key material variables have been explored. Afterwards, cathode discovery has been presented in terms of RL states, actions, rewards, policies, environments, constraints, and feedback. The key focus of this review is to analyse how RL can support the composition selection, dopant, and crystal structure optimisation. The review also discusses the current limitations of RL based cathode design including data scarcity, dataset bias, limited cathode specific benchmarks, reward function design, physical validity, and experimental validations. The future directions have been proposed for physics informed and experimentally validated RL infrastructure that integrates the density functional theory, molecular dynamics, artificial intelligence and human expertise.