Lithography Solvent Discovery using Neuro-Symbolic Search Agent

Scientific formulation discovery often unfolds under imperfect evaluators: wet-lab validation is expensive, design spaces are combinatorial, and available proxy scores only partially reflect downstream experimental performance. We study this challenge in lithography solvent design, a mixed discrete--continuous problem that requires selecting solvent components and optimizing their mixing ratios under explicit physicochemical constraints. We propose, a diversity-aware neuro-symbolic search framework for solvent discovery under evaluator uncertainty. LLMs act as chemistry-informed hypothesis generators over discrete formulation topologies, while a differentiable physics-informed module refines continuous mixture ratios and enforces feasibility. To reduce premature collapse toward evaluator-preferred patterns, combines sibling-aware local diversification with memory-driven global planning. Experiments show that maintains full compliance with the explicit physicochemical constraints adopted in our framework and improves exploration diversity over score-centric baselines. Preliminary lithography tests further suggest that representative candidates discovered through diverse search can exhibit favorable qualitative pattern definition under the tested conditions, even when they are not top-ranked by the proxy evaluator. These results highlight the value of diversity-aware search for scientific discovery when available evaluators are informative but incomplete.