The work analyzes data from the Pacific Northwest National Laboratory's CICERO workflow for autonomous selective precipitation. Active learning is employed to select experiments, aiming to connect laboratory results with product requirements and process costs. A conditional retrospective benchmark using fitted models and recycled neodymium-iron-boron (NdFeB) magnet records shows active learning identifies the best recorded result with fewer experiments than nonadaptive space filling. Adaptive policies reach the recorded enrichment maximum by 16 to 24 wells, compared to 48 wells for nonadaptive methods.
Two-stage reconstruction ties two adaptive alternatives at 16 wells. Conditional analyses of recycled samarium-cobalt (SmCo) magnets reveal a tradeoff between purity and yield, dependent on assumptions about the starting amount. Rankings for produced water from oil and gas extraction depend on phase and dilution assumptions requiring confirmation.
The proposed method selects batches based on their expected reduction in downstream Bayes risk. Exploratory simulations show a hybrid filtering approach has lower estimated loss than a joint search. Differences involving the synthetic two-stage policy are small relative to estimation uncertainty. A pre-registered test is proposed under a shared loss and logging standard, requiring clarified measurements and records, a defined process decision and relevant outputs, credible economic inputs, and validation at the intended scale.
Source: https://arxiv.org/abs/2609.09413