AI Cuts Rare-Earth Recycling Work in Half

This paper analyzes archived records from Pacific Northwest National Laboratory's CICERO workflow for autonomous selective precipitation — not new lab experiments. In a conditional retrospective benchmark with fitted models and recycled NdFeB magnet records, adaptive active-learning policies reached the best recorded enrichment (the selected rare-earth-to-iron ratio relative to the feed) in 16 to 24 wells, versus 48 for nonadaptive space filling. A two-stage reconstruction tied two adaptive alternatives at 16 wells. Conditional analyses of recycled SmCo magnets showed a Round 2 tradeoff between purity and nominal yield, and rankings for produced water from oil and gas extraction depended on phase and dilution assumptions that require confirmation. The authors propose choosing batches by their expected reduction in downstream Bayes risk; in exploratory simulations, a hybrid that filters candidates had lower estimated loss than the implemented joint search across routes and conditions, though differences involving the synthetic two-stage policy were small relative to estimation uncertainty. They outline a pre-registered prospective test requiring clarified measurements and records, a defined process decision with relevant outputs, credible economic inputs, and validation at the intended scale.