Transformer enables ion transport behavior evolution and conductivity regulation for solid electrolyte

Ab initio molecular dynamics (AIMD) is an important technique for studying ion transport within solid electrolyte and interface effects between electrode and electrolyte, which is particularly critical for the rational design of new energy materials. However, AIMD is limited by the high-cost density functional theory (DFT) solution process and is unable to reach the time scale of the entire dynamic simulation, resulting in time-consuming AIMD calculations and a considerable scarcity of AIMD-based conductor data. Here, we propose a sequence relational large model based on transformer (T-AIMD) to infer ion diffusion from mean square displacement sequence data and hybrid multi-source material descriptor. T-AIMD successfully learns the whole long-range atomic diffusion to predict the ionic conductivity (σ) of any ion in any crystal structure to find fast-ion conductors, thus reducing the cost of AIMD simulation by a factor of 100. Using T-AIMD, we built the largest database of mixed ion conductors, and the σ of representative solid electrolytes has been successfully validated in previous battery experiments. Further, the manufactured solid-state battery with the predicted promising electrolyte exhibits almost no obvious capacity decay after 50 cycles with a high initial specific capacity of 1270 mAh g−1, which is promising to help devices work in extreme environments while guaranteeing battery life. By speeding up the prediction time of AIMD, the proposed T-AIMD opens the door for scientists to explore the atomic and molecular behaviors of other molecules/materials on long time scales, and will ultimately benefit the exploration of other key scientific questions in the energy field.