A Structure-to-Density Model for Trajectory-Free Lithium Transport Screening

The time-averaged density of mobile lithium ion is where its transport becomes visible: it shows the sites lithium occupies, the pathways along which it moves, and, across temperatures, the effective barriers it must cross. Yet this density has only been available as a by-product of molecular dynamics, re-converged for every structure at every temperature. We present LiDen, to our knowledge the first model to predict this density directly from structure across temperature. From the lithium-free framework alone, it predicts the density from a single forward pass, by inferring a temperature-independent energy field and entropy field whose combination gives it, including at temperatures held out of training. On a benchmark of 4,186 lithium-bearing structures, LiDen predicts the density more accurately than histogramming a trajectory that costs over sixty times as much to generate, and a readout of the predicted field attains the lowest displacement error we are aware of on this benchmark ($0.335$ dex). The same field further yields the occupation sites, their connectivity and signatures of the transport mechanism: on the garnet LLZO it recovers the full crystallographic lithium sublattice, having never seen a lithium position, and assigns the site-to-site bridges migration barriers that bracket the reported activation energy.