Research Digest — 2026-09-01¶
AI-Driven Materials Discovery¶
1. Wyckoff-Resolved Oxidation-State Atlas and Anion-Conditioned Priors for Materials Discovery¶
Source: arXiv (2608.28492) · 📅 2026-08-28 · ↗ Open paper
Introduces a Wyckoff-resolved oxidation-state atlas built from a May 2026 snapshot of 154,879 Materials Project structures, with a probability-ranked, charge-neutral assignment utility that works from compositions or Wyckoff grammars. Compared with a strict matched MP baseline, coverage increases by 18.7% (composition mode) and 28.0% (Wyckoff mode), and 99.98% of materials recovered only in Wyckoff mode have distinct formal oxidation states on different site tokens. The CSV/Python workflow provides a reproducible prior for structure decoration, generative crystal models, and symbolic Wyckoff-grammar discovery workflows.
Relevance to DENG.Group
Yanhao Deng and group-wide: charge-neutral oxidation-state priors are the quiet workhorse behind any generative or compositional solid-electrolyte screening pipeline (decorating candidate scaffolds before ML conductivity screens). This atlas plugs directly into the kind of large-scale discovery funnel the group runs for halide and chalcohalide electrolytes, where assigning Li/Cl/Sb oxidation states correctly across candidate structures is a recurring failure point.
ML Interatomic Potentials¶
2. uMOF: A Universal Database, Benchmark, and Machine Learning Interatomic Potentials for Metal-Organic Frameworks¶
Source: arXiv (2608.28100) · 📅 2026-08-28 · ↗ Open paper
Releases the largest r2SCAN-D4 DFT dataset for MOFs to date (85,524 configurations, 19,950 frameworks, 79 elements), a literature-mined benchmark of 3,986 verified experimental/computed property values extracted by an LLM pipeline, and two universal MLIPs (uMOF-MH, uMOF-POLAR) fine-tuned from MACE foundation models. On dynamics-sensitive properties like adsorption enthalpies the fine-tuned models cut error by >80% versus all baselines, reaching experimental uncertainty. Notably, a small fraction (~1.7%) of finite-temperature MD frames in the training mix proves decisive for MLIP stability.
Relevance to DENG.Group
Yanhao Deng: a methodological template for the group's fine-tuned-potential work — the finding that a ~2% admixture of MD snapshots stabilizes potentials, and the LLM-mined experimental benchmark for validation, transfer directly to MLIPs for halide/sulfide electrolytes where static snapshots alone give brittle dynamics. The dataset-level rigor (r2SCAN + explicit diversity) is also a useful bar for the group's own training-set curation.
Degradation & Physics-Informed Modeling¶
3. PiDDM: Physics-Informed Differentiable Degradation Modeling for Lithium-Ion Battery State-of-Health Prediction¶
Source: arXiv (2607.29095) · 📅 2026-07-31 · ↗ Open paper
Embeds Arrhenius-type degradation kinetics for SEI growth and loss of lithium inventory directly into the training loss of a differentiable capacity-fade model, evaluated on 55 cells across six cycling protocols. PiDDM achieves the lowest average SOH prediction error versus data-driven baselines and, in extrapolation tests on the final 10% of cycle life, captures accelerated end-of-life fade while avoiding the nonphysical capacity regeneration produced by MLP and baseline PINN models.
Relevance to DENG.Group
Shoutong Jin: the differentiable-physics strategy is complementary to his phase-field dendrite work — the same Arrhenius SEI/LLI kinetics he sees at dendrite/electrolyte interfaces become soft constraints here, a pattern worth borrowing for coupling atomistic interface physics to cell-scale predictions. Umang Agarwal: the treatment of SEI growth as the governing degradation channel connects his electrolyte/electrode interface kinetics to observable cell-level aging.
Community & Resources¶
4. An open access solid-state battery cell database to advance visualization, search, analysis, and AI data extraction¶
Source: Joule · 📅 2026-07-02 · ↗ Open paper
Presents an open-access database of solid-state battery cell test data aggregated from 33 journals, with tooling for visualization, search, structured analysis, and AI-based data extraction from the literature. The resource consolidates otherwise-scattered cell-level performance, cycling, and testing-condition data into a machine-readable format.
Relevance to DENG.Group
Group-wide: a ready source of experimental cell data (formats, stack pressures, cycling protocols, failure modes) for benchmarking the group's transport, interface, and dendrite simulations against real cells — useful for Yanhao and Naibing when anchoring MLIP-MD predictions to observed conductivities, and for Shoutong when parameterizing phase-field models. Catch-up pick: published early July but not previously covered in the digest.