A Three-Stage, Expert-in-the-Loop AI Pipeline for the Discovery of a Cobalt- and Nickel-Free Rock-Salt Li-ion Cathode

Cobalt- and nickel-free rock-salt Li-M-O oxides are one of the most promising high-capacity Li-ion cathode families, but they remain a hard target for automated discovery: the transport-relevant channel information that governs Li$^+$ intercalation is not captured by formation-energy predictors alone, and the cation-disordered rock-salt (DRX) chemistry that unlocks their capacity is not represented in the ordered crystallographic entries that public databases and current generative crystal models produce. We report an end-to-end pipeline that operates on this representation gap by searching over \emph{ordered} rock-salt-prototype Li-M-O entries and then handing the shortlist to a domain expert whose reading of the DRX literature drives the final composition choice and the synthesis route. The pipeline has three stages: (i) a proposal stage that gathers candidates from public materials databases (Materials Project, OQMD, NOMAD, COD) and, as a complementary source, from structure-based generative crystal models (CCDCGAN and FTCP), in every case with ordered site occupancy as stored upstream; (ii) an automated multi-property analysis stage that scores each ordered candidate on ionic-transport channel connectivity (CAVD), migration barriers (BVSE), elemental cost proxied by crustal abundance, and machine-learning predictions of specific capacity, gravimetric energy density, average voltage and formation energy per atom, complemented by a synthesisability prior; and (iii) an expert-driven ranking stage in which the principal investigator of an inorganic-chemistry group at an academic partner defined a flexible weighted objective over the property table, browsed interactive 3D structural views of the top-ranked candidates, and selected the composition. In this campaign the academic partner, informed by the DRX literature for this Li-M-O window, then synthesised the selected composition via high-energy ball milling into a disordered rock-salt phase, characterised it structurally, and integrated it into a coin-cell format, where the un-doped baseline (denoted V1 in our internal notation) delivered a first-discharge specific capacity close to $290$ mAh/g at $\sim 3$ V vs.\ Li/Li$^+$, above the $250$ mAh/g KPI set at project inception. The candidate reported here came from the database branch of stage (i); the generative branch was tested in parallel but was, at the time of this work, not yet reliable enough to contribute to the final shortlist. We describe the pipeline, the coin-cell validation, and the design choices that made the transfer from the table to the wet lab tractable. Follow-up work on doping refinement and industrial scale-up is under active patent examination and is deliberately kept outside the scope of this manuscript.