Active Learning Under Safety and Manufacturability Constraints: Candidate Selection, Feasible Regions and Stopping Rules

Active learning can concentrate experiments where new data are most informative, but unconstrained selection is inappropriate for photopolymer development because candidate formulations can be unstable, unmixable, uncurable, unsafe or incompatible with equipment. This review defines a constrained active-learning architecture with separate models for scientific response, feasibility and operational risk. It distinguishes uncertainty sampling from objective optimisation, retains failed experiments as boundary information, and introduces stopping rules based on decision resolution and value of information. Published materials-discovery and adaptive-manufacturing studies demonstrate the power of uncertainty-guided experimentation, while safety-focused SDL literature emphasises the need for bounded autonomy. The photopolymer implementation proposed here requires pre-authorised search domains, hard interlocks and independent confirmation. ER-407.