Choosing a recovery process for scale‑up requires linking laboratory results with product specifications, process costs, and scale effects. We examined records from the Pacific Northwest National Laboratory’s CICERO workflow that supports autonomous selective precipitation. Active learning leverages existing results to select the next experiments. In a conditional retrospective benchmark using fitted models and recycled neodymium‑iron‑boron (NdFeB) magnet data, active learning identified the best recorded outcome with fewer experiments than a non‑adaptive space‑filling design. Enrichment is defined as the rare‑earth‑to‑iron ratio relative to the feed. Adaptive policies reach the recorded enrichment maximum within 16–24 wells, whereas the non‑adaptive approach needs about 48. Our two‑stage reconstruction ties the two adaptive alternatives at 16 wells. Conditional analysis of recycled samarium‑cobalt (SmCo) magnets reveals a Round 2 trade‑off between purity and nominal yield; recovery fraction is computed from an assumed starting amount, and NdFeB Round 1 routes differ in enrichment. Rankings for produced water from oil‑and‑gas extraction depend on phase and dilution assumptions and therefore require validation. We propose selecting batches by their expected reduction in downstream Bayes risk: the minimum expected loss among available process decisions under current beliefs. Exploratory simulations show that a hybrid that first filters candidates yields lower estimated loss than a joint search across all routes and conditions. Differences involving the synthetic two‑stage policy are small relative to estimation uncertainty. Finally, we outline a pre‑registered prospective test under a shared loss and logging standard, requiring clarified measurements, a defined process decision and outputs, credible economic inputs, and validation at the intended scale.
Review