Source-linked AI summary
A Hierarchical Synergistic Deep Learning Framework Integrating Composition, Structure, and Ionic Transport for Solid-State Electrolyte Discovery
Hongwei Du, Dingyang Lv, Baole Wei, Yongheng Li, Feng Yu, Ziheng Lu, Siqi Shi, Hong Wang
TL;DR
Reliable multi-objective screening of solid-state electrolytes is limited by distribution mismatch, heterogeneous property datasets, and scarce kinetic transport data. The paper addresses these challenges with a hierarchical framework that assigns specialized compositional, structural, and transport models to sequential stages, identifying 97 high-performance candidates from 30.36 million Alex/ICSD-derived materials. Its transport analysis highlights Li+ jump-network connectivity and oxide-framework rigidity as key factors within the studied systems.
Problem
Training-data distribution mismatch, cross-property dataset heterogeneity, and scarce kinetic labels complicate reliable multi-constraint SSE screening.
Method
A hierarchical workflow assigns L-G-DCNN, multi-fidelity DenseGNN, MatterSim, and DeePMD to sequential compositional, structural, and transport-screening tasks.
Results
97 high-performance candidates with room-temperature conductivities of 0.109 to 59.0 mS/cm were identified from the 30.36-million Alex/ICSD-derived candidate space.
Takeaways & Limitations
Li+ jump-network connectivity is identified as more important for room-temperature ionic conductivity than the quantity of geometric Li sites, while oxide rigidity may limit transport.
Takeaways & Limitations
Prototype-based elemental substitution limits discovery of entirely novel structures, while idealized calculations omit several real battery operating factors and can diverge from experimental performance.
Abstract
from arXiv · showhide
Inorganic solid-state electrolytes must combine high room-temperature ionic conductivity, a wide electrochemical window, excellent electronic insulation, and favorable mechanical compliance. Single models struggle to support reliable multi-objective screening across vast chemical spaces because of training-data distribution mismatch, cross-property dataset heterogeneity, and scarce kinetic transport data. To overcome these limitations, we develop a hierarchical synergistic deep-learning framework that sequentially coordinates efficiency, accuracy, and reliability through four complementary modules. The in-house-developed L-G-DCNN and a multi-fidelity implementation built on DenseGNN serve as compositional and structural experts for thermodynamic coarse screening and multi-property evaluation, respectively; MatterSim and system-specific DeePMD models provide transport pre-assessment and kinetic validation. Systematic benchmarks show that each module outperforms mainstream counterparts in its task, while retrospective validation establishes dual closed-loop verification of module-level accuracy and end-to-end workflow reliability. Applied to 30,364,908 Alex/ICSD-derived candidates, the framework identifies 97 high-performance candidates with room-temperature ionic conductivities of 0.109--59.0 mS/cm, including 94 halides, one borohydride, and two oxides. Consistency with independent experimental data confirms that 76 of the 94 halides fall within reported high-conductivity structural regions. Analysis reveals that Li$^{+}$ jump-network connectivity, rather than the number of geometric Li sites, is the core determinant of room-temperature ionic conductivity. Li-defect engineering effectively enhances oxide transport, whereas the inherent rigidity of the O$^{2-}$ framework suggests a potential upper limit on oxide electrolyte performance.
1 Introduction
Reliable screening of solid-state electrolytes requires balancing computational efficiency, multi-property accuracy, and transport-validation reliability across chemically diverse, data-limited spaces. The proposed workflow assigns specialized models to sequential compositional, structural, and dynamic screening stages.
- SSE screening must jointly target ionic conductivity, electrochemical stability, electronic insulation, and mechanical compliance.
- Training-data distribution mismatch, uneven cross-property datasets, and scarce kinetic labels hinder reliable multi-constraint screening.Existing structure models often train on stable or metastable structures, while kinetic labels are scarce and strongly time dependent.
- The framework uses L-G-DCNN for low-cost compositional coarse filtering before structure-sensitive multi-property evaluation.The staged design assigns distinct representations and task-specific models to successive screening stages.
- DenseGNN combines multi-fidelity data to evaluate convex hull energy, band gap, and elastic moduli, while MatterSim and DeePMD assess and validate ionic transport.Transport evaluation progresses from MatterSim mobility assessment to high-precision DeePMD kinetic validation.
- Systematic benchmarks and retrospective validation assess module-level accuracy and end-to-end workflow reliability before large-scale deployment.The framework is benchmarked across compositional, structural, and dynamic data domains, with classic SSEs used for retrospective validation.
- Transport analysis identifies Li+ jump-network connectivity, rather than geometric Li-site count, as the core determinant of room-temperature ionic conductivity.Li-defect engineering can enhance oxide transport, but rigid oxide anion frameworks may impose a potential upper limit.
2 Results
The study validates a hierarchical screening framework across compositional, structural, and transport tasks before applying it to large candidate spaces. The workflow identifies high-conductivity candidates, agrees with experimental halide trends, and links transport to connected Li+ jump networks and oxide-specific trade-offs.
- Systematic Validation: Systematic benchmarks validate dedicated models for compositional screening, structural multi-property prediction, and ionic transport evaluation.The framework uses task-specific models rather than a single universal model, with retrospective validation against established experimental trends.
- Systematic Validation: False-positive ionic-conductivity decisions were markedly reduced relative to a single-model M3GNet workflow in a 170,470-compound MP benchmark.The same-protocol comparison evaluated agreement with DFT reference decisions at critical screening nodes.
- Systematic Validation: 156 common SSEs reproduced established chemical and structural trends, including dominant monoclinic, orthorhombic, and triclinic crystal systems and mostly sub-0.1 eV/atom convex-hull energies.Relaxed screening criteria were used to examine representative sulfide, halide, and oxide families and compare conductivity with stability, electronic, and electrochemical properties.
- Extended Screening: 97 final candidates span 0.109 to 59.0 mS/cm, comprising 94 halides, 2 oxides, and 1 closo-type borohydride.MatterSim first identified 124 candidates above 0.1 mS/cm, followed by long-time DeePMD validation of 97 candidates.
- Experimental Consistency: 76 of 94 halide candidates fall within experimentally verified hcp-O or ccp-M high-conductivity regions, including 18 of the top 20 candidates by conductivity.Additionally, 73 of 94 candidates fall within measured conductivity ranges for corresponding experimental structure regions.
- Transport Mechanisms and Oxide Design: Connected Li+ jump networks, rather than geometric Li-site quantity, distinguish long-range transport, while Li-defect engineering improves oxide conductivity but oxide-framework rigidity may limit performance.Li3BO3 lacks a connected jump network despite dense Li sites; Li-defect modification of derivatives can produce large conductivity gains with trade-offs among stability, window, and band gap.
3 Discussion
The framework addresses three challenges in multi-constraint solid-state electrolyte screening and identifies 97 high-performance candidates from a 30.36-million-candidate space. The authors also define limitations and future priorities involving structural novelty, realistic operating conditions, adaptive routing, and broader conductor classes.
- 97 high-performance candidates were identified from the 30.36-million Alex/ICSD-derived candidate space, with conductivities of 0.109 to 59.0 mS/cm.
- The framework addresses training-data distribution mismatch, cross-property dataset heterogeneity, and kinetic data scarcity through specialized models operating within validated performance domains.
- Structure generation based mainly on elemental substitution of existing prototypes limits discovery of entirely novel structural motifs.
- Single-crystal 0 K calculations omit finite-temperature effects, interface reactions, grain boundaries, impurities, and microstructural defects, contributing to discrepancies with experimental performance.
- Future development prioritizes innovative anion frameworks, continuous migration pathways, multi-scale realistic-condition modeling, adaptive routing, and extension to Na+/Mg2+ conductors and protective layers.
4 Data and Methods
The workflow combines composition, structure, and transport models with staged screening, molecular dynamics, and DFT-referenced validation. Training and validation choices address metastability, dataset heterogeneity, finite-size effects, and trajectory reliability.
- Candidate generation: 81 elements were used for candidate generation after excluding chemically inert, radioactive, scarce, redox-active, or electronically problematic elements.The exclusions included noble gases, actinides, promethium, late lanthanides, several scarce elements, and transition metals with accessible redox couples.
- Model training: L-G-DCNN training included OQMD and Materials Project formation energies spanning -15 to 10.0 eV·atom-1, including unrelaxed and metastable structures.Including such structures was intended to reduce false negatives for metastable candidates.
- Stability evaluation: MatterSim-based electrochemical stability calculations parsed candidate structures, retrieved Materials Project energies and corrections, and relaxed structures with RelaxCalc.The workflow extracted chemical systems before retrieving energy entries and applying structure optimization.
- Transport evaluation: Transport evaluation used target-temperature molecular dynamics, periodic-boundary unwrapping, Li self-MSD analysis, Nernst–Einstein conductivity, and Arrhenius fitting.MatterSim supported high-throughput pre-evaluation, while DeePMD was used for final-candidate transport calculations.
- Transport reliability: At least 1200-atom supercells were used for room-temperature runs, while trajectory quality was assessed with fixed-window, block, and trajectory diagnostics.Formal size-convergence studies and independent replicate trajectories were not available for every candidate.
- Potential validation: DeePMD was compared with VASP-PBE labels for 916 structures spanning ambient, strained, and high-temperature configurations.The benchmark covered 441 room-temperature trajectory frames, 379 small-strain structures, and 96 high-temperature frames.
5 Data Availability
The article and Supporting Information contain the study’s data, while code, processed datasets, candidate structures, screening outputs, validation data, and reproducibility scripts are publicly available in the MatCascade repository.
- Repository contents: The MatCascade repository publicly provides submission-version code, processed datasets, candidate lists, crystal structures, screening summaries, validation data, cross-validation tables, and reproducibility scripts.The raw long-time transport data are not included because of their size.