US researchers used generative artificial intelligence to screen thousands of candidate crystal structures for multivalent-ion battery electrodes, then built a separate AI system to propose lab synthesis routes for the untested materials. Researchers at the New Jersey Institute of Technology (NJIT) and Rensselaer Polytechnic Institute (RPI) identified five previously unreported oxide compositions that could serve as electrode hosts for multivalent-ion batteries. The team, led by Joy Datta under adviser Dibakar Datta at NJIT, working with Nikhil Koratkar at RPI, combined a crystal diffusion variational autoencoder (CDVAE) with a fine-tuned large language model to generate roughly 20,000 candidate transition metal oxide structures. The models were trained on more than 44,000 known inorganic crystal structures from the Materials Project database. The researchers filtered the generated structures using formation energy, thermodynamic stability and electronic band gap criteria, narrowing the field to 55 compositions. Of those, the CDVAE model produced the five structures, featuring open-tunnel frameworks designed to accommodate magnesium, calcium, aluminum or zinc ions. One candidate, Ca4In2O2, showed an energy above the convex hull of 0.36 eV/atom, a level the researchers describe as metastable rather than fully stable. A phonon dispersion analysis showed no unstable vibrational modes across its entire Brillouin zone, indicating it could potentially be synthesized under non-equilibrium conditions. The team also developed a retrieval-augmented generation system that searches scientific