Research Digest — 2026-03-23¶

Papers¶

1. An Analytical Model of Alkali Metal Dendrite Growth in Ceramic Solid Electrolytes based on Griffith's Theory¶

Source: arXiv (cond-mat.mtrl-sci), Mar 20, 2026 · 📅 2026-03-20 · ↗ Open paper

Dendrite propagation in ceramic SEs is controlled by a competition between mechanical fracture energy (Griffith's theory) and Joule heating from current detouring around the dendrite. The critical current density follows \(J_{\mathrm{crit}} \propto c_{\max}^{3/2}\), where \(c_{\max}\) is the longest pre-existing interfacial defect length. Analytical fracture mechanics coupled with electrostatic energy balance. Derives a closed-form scaling law for critical current density in ceramic SEs based on minimal power dissipation, and predicts that \(J_{\mathrm{crit}}\) scattering must follow a Weibull distribution — analogous to tensile strength statistics in brittle ceramics.

2. Prediction and Experimental Verification of Electrolyte Solvation Structure from an OMol25-Trained Interatomic Potential¶

Source: arXiv (physics.chem-ph), Mar 20, 2026 · 📅 2026-03-20 · ↗ Open paper

The OMol25-trained Universal Model of Atoms (UMA-OMol) predicts experimental densities and X-ray structure factors in Na-ion electrolytes with substantially better agreement than materials-only models, revealing systematic trends in solvation structure as a function of cation/anion identity, concentration, and solvent topology. MLIP-driven molecular dynamics with experimental validation (X-ray structure factors, density measurements). Comparison of OMol25-trained vs materials-trained MLIPs across diverse Na-ion electrolyte compositions. First systematic, experimentally validated comparison of molecular vs materials training data for battery electrolyte MLIPs. Shows that OMol25 captures ion-solvent interactions and contact ion pair formation that materials-trained models completely miss.

3. Multi-fidelity Machine Learning Interatomic Potentials for Charged Point Defects¶

Source: arXiv (cond-mat.mtrl-sci), Mar 5, 2026 · 📅 2026-03-05 · ↗ Open paper

Foundation MLIPs fail at describing charged point defects because defect environments involve coordination and electron counts far from training data. A global defect charge embedding combined with multi-fidelity training (semi-local + hybrid functional data) yields defect-capable force fields that predict charge-transition levels in quantitative agreement with direct DFT. Global defect charge embeddings in MLIP architecture. Multi-fidelity training combining PBE-level data (abundant, cheap) with hybrid functional data (scarce, accurate) for charged defects in \(\ce{Sb2Se3}\). First explicit demonstration that foundation MLIPs fail for charged defect physics, with a concrete solution: charge-state-dependent global embeddings plus multi-fidelity training. The resulting model predicts charge-transition levels at a fraction of DFT cost.

4. Proof-Carrying Materials: Falsifiable Safety Certificates for Machine-Learned Interatomic Potentials¶

Source: arXiv (cond-mat.mtrl-sci), Mar 12, 2026 · 📅 2026-03-12 · ↗ Open paper

Proof-Carrying Materials (PCM) provides falsifiable safety certificates for MLIP predictions through adversarial falsification, bootstrap envelope refinement with 95% confidence intervals, and formal Lean 4 certification. Auditing CHGNet, TensorNet, and MACE reveals architecture-specific blind spots with near-zero pairwise error correlations. Three-stage framework: adversarial falsification across compositional space, bootstrap envelope refinement, and Lean 4 formal verification. Validated on 25,000-material benchmark and independent Quantum ESPRESSO calculations. First formal reliability guarantee framework for MLIP-based materials screening. Shows that MACE, CHGNet, and TensorNet have completely different blind spots (r ≤ 0.13 error correlation), meaning ensembles are essential — a single model is never sufficient.

5. Scaling Machine Learning Interatomic Potentials with Mixtures of Experts¶

Source: arXiv (physics.chem-ph), Mar 9, 2026 · 📅 2026-03-09 · ↗ Open paper

Sparse MoE with element-wise routing and shared experts yields substantial performance gains over monolithic models. Nonlinear MoE outperforms linear MoE, and the routing patterns reveal chemically interpretable expert specialization — different elements are handled by different expert subnetworks. Systematic development of MoE and MoLE architectures for MLIPs with different routing strategies (element-wise, configuration-level, global). Benchmarked across three major MLIP datasets. First systematic study of MoE for MLIPs showing that element-wise routing with sparse activation consistently outperforms other strategies. The chemically interpretable routing patterns are novel — the model effectively learns a periodic-table-aware decomposition of chemical space.

6. A Recipe for Scalable Attention-Based MLIPs: Unlocking Long-Range Accuracy with All-to-All Node Attention¶

Source: arXiv (cs.LG), Mar 6, 2026 · 📅 2026-03-06 · ↗ Open paper

AllScAIP uses an all-to-all node attention component to capture long-range interactions in a fully data-driven manner. As data and model size scale, physics-based inductive biases become less important — or even counterproductive — while all-to-all attention remains critical for accuracy on electrolyte-scale systems. Attention-based, energy-conserving MLIP architecture trained on O(100 million) samples. Benchmarked on OMol25, OMat24, and OC20M. Includes stable long-timescale MD validation against experimental observables. Shows that physics-based inductive biases (equivariance, message passing locality) improve sample efficiency in small-data regimes but reverse at scale, while all-to-all attention consistently matters. This challenges the prevailing assumption that equivariant architectures are always superior.