| Abstract Scope |
AI–driven experimentation can fulfill ICME’s promise only when algorithms receive clear signals about what progress looks like, and those signals are encoded as rigorously engineered rewards. In this talk I will trace the emerging landscape of reward engineering for materials discovery, beginning with performance-oriented metrics that convert materials property/performance targets into actionable optimization objectives. I will then pivot to knowledge-oriented rewards—quantifiers of uncertainty reduction, entropy minimization and design-space coverage—that prevent premature convergence and turn exploration itself into a measurable benefit. Building on these foundations, I will introduce meta-reward learning, where a higher-level agent dynamically reshapes the reward landscape in response to evolving data, continuously re-balancing exploration and exploitation. Finally, I will discuss how these ideas extend to the multi-agent, self-driving laboratories now emerging: planners, simulators and robotic experimenters negotiate shared or role-specific incentives so that individual actions collectively advance both material performance and scientific understanding, enabling truly autonomous, knowledge-centric ICME. |