PNNL researchers tested an algorithm that decides which chemistry experiment to try next when recovering rare-earth metals from recycled magnets, and it needed far fewer experiments than random testing.
The team ran a retrospective benchmark on CICERO, their autonomous workflow for selective precipitation, the wet-chemistry step that separates one metal from another in solution. Using records from recycled neodymium-iron-boron (NdFeB) magnets, the active-learning approach found the best previously recorded enrichment result, the ratio of rare-earth metal to iron in the output, in just 16 to 24 individual experiments, which the paper calls wells. A conventional approach that tests evenly across the whole space needed 48 experiments to match it, and a two-stage version of the algorithm tied the best adaptive result at 16 wells. Follow-up tests on recycled samarium-cobalt magnets and oil-and-gas wastewater samples used the same method but surfaced messier tradeoffs, including purity versus yield, that the authors say still need confirming.
Recycling rare earths from scrap magnets is one of the few near-term ways the US can loosen its dependence on Chinese-controlled rare-earth supply chains, and the bottleneck has never really been the chemistry. It is the slow, expensive process of tuning a recovery method before it can run at plant scale. Shaving lab experiments by half or more could shrink that runway and let labs screen more kinds of recycled feedstock without a matching jump in cost.
The authors are careful to call this a retrospective benchmark, not a proven process, and they are asking for a pre-registered prospective test before anyone treats the numbers as settled. That is a polite way of saying nobody has run this live on an actual production line yet.