The SEI is a passive layer, typically several hundred angstroms thick, that forms spontaneously when the electrolyte decomposes on contact with lithium metal, the most reducing electrode known. Ideally, it should block electrons while letting lithium ions pass freely, but in practice it is a heterogeneous, mosaic-like bi-layer: an inorganic-rich inner layer hugging the electrode and a porous organic outer layer facing the electrolyte. Microscopic studies have shown that the inorganic sublayer is not a continuous film but a dense packing of crystalline grains separated by numerous grain boundaries, and it is precisely this grain-boundary-rich architecture that shapes how ions travel. Nonuniform transport through these grains and boundaries concentrates lithium deposition in localized spots, which amplifies with every charge-discharge cycle until dendrites nucleate, grow, and penetrate.

Earlier computational work mapped lithium-ion transport only in a handful of representative SEI components, mostly lithium fluoride and lithium oxide, using density functional theory with nudged elastic band, or NEB, calculations to trace minimum-energy pathways. Those studies revealed that grain boundaries create open diffusion channels that can speed ion transport, but they covered few pathways and few surface orientations. The Oklahoma team instead built a systematic dataset spanning eight major inorganic SEI phases: lithium fluoride, chloride, bromide, and iodide; lithium oxide; lithium sulfide; lithium nitride; and lithium carbonate. For each, they generated NEB diffusion data across low- and medium-energy crystal surfaces and across both homogeneous grain boundaries and heterogeneous interfaces between different phases, producing a unified design space for SEI ion migration.

The computational machinery behind the dataset was substantial. Bulk structures were drawn from the Materials Project repository and optimized with the Vienna Ab-initio Simulation Package using the Perdew-Burke-Ernzerhof functional, a 500 eV plane-wave cutoff, and forces converged below 0.1 eV per angstrom. Grain slabs were cut from surfaces with energies below 250 meV per square angstrom, since such low-energy facets are the ones most likely to appear in real grains. Homogeneous grain boundaries were constructed using coincidence site lattice theory with sigma values of 3, 5, 7, 9, 11, and 13, and boundaries exceeding 1,000 millijoules per square meter in energy were discarded as physically implausible. Heterogeneous interfaces were assembled with the Zur-McGill lattice-matching algorithm, keeping only commensurate models with average lattice mismatch below five percent. Diffusion was modeled as vacancy-mediated hopping, the mechanism prior work identified as most favorable in SEI compounds, with climbing-image NEB calculations locating the saddle points along each pathway.

The resulting dataset comprised 280 grain-diffusion NEB calculations plus 198 homogeneous grain-boundary and 104 heterogeneous-interface cases, each NEB run containing seven intermediate structures along the migration path, for a total of 2,114 data points. That volume of quantum-mechanical data is exactly what makes machine learning worthwhile: exhaustive NEB screening of every conceivable SEI chemistry, surface orientation, and boundary configuration would be computationally prohibitive, so a predictive model trained on a representative sample can extend the reach of first-principles accuracy to configurations never explicitly calculated.

The modeling architecture is where the study departs from conventional approaches. Each NEB image, whether initial, intermediate, or final, was converted into a periodic crystal graph in which atoms are nodes and interatomic interactions within a cutoff radius are edges, encoded by the directional message-passing network DimeNet++. A pretrained graph variational autoencoder, adapted from the crystal diffusion VAE of Xie and colleagues, then learned latent representations of entire NEB trajectories, encoding all seven structures of a diffusion event jointly rather than treating each configuration in isolation. The autoencoder’s decoder is denoising-based, recovering structures from Gaussian-perturbed coordinates and interpolated atom-type distributions, which strengthens the latent space. These latent embeddings were concatenated with composition-based descriptors from the Matminer library and structural descriptors such as exposed surface, surface or grain-boundary energy, and a saddle-point indicator, then processed by a multi-layer Transformer encoder whose self-attention captures energy curvature and collective structural evolution along the pathway.

The framework’s predictive performance was strong in both transport regimes. For grain diffusion, the best model, combining original descriptors with GVAE latent embeddings, achieved a test-set R-squared of 0.93; for grain-boundary diffusion, the corresponding model reached 0.94. In both cases, the model that fused raw descriptors with latent embeddings outperformed versions trained on either alone, indicating that the autoencoder’s latent space carries physically meaningful information about configurational relationships among diffusion pathways that fixed feature vectors miss. Predicted minimum-energy paths reproduced the saddle points of independently calculated NEB profiles across all eight SEI components, and a held-out validation set never touched during training or testing confirmed the model’s reliability.

Feature-importance analysis, performed by permutation, revealed what actually governs migration. Composition-derived descriptors such as melting temperature, space-group number, and ground-state volume ranked high statistically, but the authors interpret these as proxies for bonding character rather than direct causal variables. The physically direct signals were structural: the saddle-point indicator carried the largest importance score among structural features, followed by grain-boundary energy, boundary type, and coincidence site lattice value. Lower-sigma boundaries, which preserve better atomic registry between grains, supported smoother, lower-barrier pathways, while exposed surface orientation, particularly the (210) and (111) planes, shaped the local coordination geometry and bottleneck topology that determine how easily a lithium ion can hop.

The transport trends that emerged are strikingly chemistry-specific. Latent-space sampling showed lithium halides defining the lowest and most uniform barrier landscape, with mean migration barriers of just 0.12 eV for lithium chloride and 0.11 eV for lithium iodide, and narrow standard deviations below 0.10 eV. Lithium nitride and lithium carbonate, by contrast, exhibited much broader and higher barrier distributions, reaching means of 0.77 and 0.86 eV respectively, and heterogeneous interfaces were systematically more resistive than homogeneous ones, with halide-carbonate combinations exceeding 1.50 eV. Even the most favorable homogeneous boundaries, such as LiF-LiF and LiI-LiI, had mean barriers of 0.62 and 0.71 eV, higher than the best intra-grain values, confirming that grain boundaries act as bottlenecks even when their parent phases are good conductors.

The authors are candid about limitations: the analysis covers only vacancy-mediated diffusion, only crystalline inorganic phases, and static zero-temperature energetics, omitting the organic species, amorphous regions, and dynamic disorder of real SEI layers. Even so, the framework delivers something the field has lacked: a quantitative structure-transport map showing that SEI performance must be engineered through phase selection, surface orientation, grain-boundary character, and interface stability together, not through chemistry alone. Because localized low-energy pathways concentrate ionic flux and thereby promote nonuniform lithium deposition, controlling the microstructural distribution of grains and boundaries may prove just as important as choosing the right electrolyte additives in the quest to keep lithium-metal batteries dendrite-free.

Subject of Research: Machine learning prediction of lithium-ion transport barriers in the inorganic solid electrolyte interphase of lithium-metal batteries

Article Title: Deep learning assisted modeling of Li-ion transport in SEI: a graph neural network based study

Article References: Kulathuvayal, A. S., & Su, Y. (2026). Deep learning assisted modeling of Li-ion transport in SEI: a graph neural network based study. Frontiers in Batteries and Electrochemistry, 5, Article 1849012. https://doi.org/10.3389/fbael.2026.1849012

Image Credits: AI Generated

DOI: 10.3389/fbael.2026.1849012

Keywords: solid electrolyte interphase, lithium-metal batteries, graph neural networks, density functional theory, nudged elastic band, migration barrier, grain boundaries, lithium dendrites, variational autoencoder, machine learning, ion diffusion, battery materials