Model output is not the end point

A high-scoring candidate formula is valuable only if it is feasible to synthesize, reproducible to test, and acceptable at cost. AI in materials R&D is more like a collaborator who is good at narrowing down the search space.

Data chain determines the upper limit

Structure, workmanship, performance and failed experiments should all be documented. Compared with stacking more complex models, establishing a traceable and contextual data system often brings better long-term returns.