A computational screen has narrowed a library of 30,364,908 generated structures to 97 predicted solid-state electrolyte candidates with room-temperature Li+ conductivities ranging from 0.109 to 59.0 mS/cm. The final set is heavily weighted toward halides: 94 candidates are halides, alongside two oxides and one closo-type borohydride. The results are computational estimates, not direct laboratory measurements.
A model for each stage
The study splits the job among specialist models. L-G-DCNN handles composition screening, multi-fidelity DenseGNN handles structure-sensitive static properties, MatterSim performs transport pre-assessment, and system-specific DeePMD carries out high-precision kinetic validation. The initial library was made by elemental-equivalent substitution into Alexandria and 2022 ICSD prototypes, producing 30,364,908 candidate structures.
At the composition stage, L-G-DCNN had the lowest reported mean absolute error, or average prediction miss, across the tested Matbench composition tasks. On MP data, its formation-enthalpy MAE was 0.078 eV/atom, and it outperformed Roost and CrabNet in that comparison. For structure-sensitive properties, DenseGNN had smaller reported errors than most mainstream structural models, while MODNet was described as comparable for dielectric-property prediction.
At 1100 K, MatterSim had the minimum reported diffusion error for sulfide, oxide and halide systems. At 300 K, it remained best for sulfide and halide systems and was comparable to SevenNet-MF-ompa for oxides. The comparison was made within the tested material families, temperatures and reference potentials.
Checks along the way
An end-to-end check used 170,470 MP compounds with DFT labels. Under the same protocol, the hierarchical workflow showed higher decision agreement with DFT and fewer ionic-conductivity false positives than a single-model M3GNet control. System-specific DeePMD transport potentials were compared with VASP-PBE energy, force and stress labels for 916 structures. CAVD+BVSE provided an independent static ranking-consistency check for 123 broad-pool structures and 96 applicable final candidates, rather than replacing finite-temperature molecular-dynamics validation.
At the late transport stage, MatterSim molecular dynamics at 300 K identified 124 candidates above 0.1 mS/cm. Long-time DeePMD simulations at 300 K, combined with trajectory-quality control, confirmed 97 final high-conductivity candidates.
The chemistry behind the numbers
Among the 94 halides in the final set, 76 fell within experimentally verified hcp-O or ccp-M high-conductivity regions, including 18 of the top 20 candidates by conductivity. Another 73 had calculated conductivities within corresponding measured ranges. That overlap supports consistency with reported trends, but it does not directly validate the predicted structures experimentally.
Trajectory analysis identified connectivity in Li+ jump networks, rather than the total number of geometric Li sites, as the reported structural correlate of room-temperature conductivity. In representative trajectories, network-contributing sites numbered 134 in Li6 CaCl8, 23 in Li2 B12 H12 and 5 in Li3 BO3. The study presents that as a correlation in modeled trajectories, not a universal causal rule.
An additional design test examined Li3 BO3 derivatives. Within a non-exhaustive set of 58 structures, 46 exceeded 0.1 mS/cm and 35 exceeded 1.0 mS/cm. The best cases were described as achieving more than four orders of magnitude improvement, but the design strategies involved trade-offs in band gap, electrochemical window and thermodynamic stability.
Still a computational result
The study remains a computational proposal. The halide comparison is a consistency check rather than direct experimental validation, and the reported conductivities are calculated values rather than measurements. Processed data, candidate lists and structures, validation data, BVSE tables and reproducibility scripts are publicly available. Raw long-timescale MD trajectories, DFT files, trained potentials and the complete 30,364,908-candidate library are not hosted in the repository. The front matter identifies the work as arXiv:2608.25592v1, dated 26 Aug 2026.
Paper data and sources
Original title: A Hierarchical Synergistic Deep Learning Framework Integrating Composition, Structure, and Ionic Transport for Solid-State Electrolyte Discovery
Authors: Hongwei Du, Dingyang Lv, Baole Wei et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text