AI-guided de novo design of flotation collectors under small-data conditions: Application to bastnaesite flotation

The discovery of flotation collectors remains constrained by limited experimental data and continued reliance on empirical modification of known molecular structures. An AI-guided de novo molecular design framework was developed using only 36 reference collector structures and six bastnaesite-fluorite separation experiments. The framework integrates mechanism-informed molecular generation, structural filtering, small-data prediction, and uncertainty-aware ranking. Candidate structures were generated and screened computationally, and 4-( N , N -dipropylamino)benzohydroxamic acid (DPABHA) was selected for synthesis and experimental validation. The framework generates 121 candidate collectors, of which 90 satisfy the predefined screening criteria, and DPABHA ranks seventh among the retained candidates. Under the optimized mixed-mineral flotation conditions of pH=8.0 and a DPABHA dosage of 2.0 × 10 −5 mol/L, DPABHA achieves a bastnaesite recovery of 92.43%, limits fluorite recovery to 14.75%, and produces a concentrate with a rare-earth oxide (REO) grade of 61.59 wt%. Interfacial analyses and theoretical calculations consistently show preferential adsorption and stronger hydrophobic modification of DPABHA on bastnaesite than on fluorite. These findings demonstrate that mechanism-informed small-data learning provides a practical route for de novo flotation collector discovery under severe data scarcity.