March 23, 2026¶

Papers¶

1. Dendrite Growth in Ceramic Electrolytes¶

Full title: An Analytical Model of Alkali Metal Dendrite Growth in Ceramic Solid Electrolytes based on Griffith's Theory

Authors: Ansgar Lowack¹
Link: arXiv:2603.20113
Source: arXiv (cond-mat.mtrl-sci), Mar 20, 2026
Why it matters: Directly relevant to Shoutong Jin's phase field dendrite work. Provides an analytical framework for understanding critical current density in ceramic solid electrolytes, complementing the group's computational phase field approach with a mechanics-based analytical model.
Key idea: 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.
Method: Analytical fracture mechanics coupled with electrostatic energy balance.
What is actually new: 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.
Potential reuse: Shoutong Jin can incorporate the \(J_{\mathrm{crit}} \propto c_{\max}^{3/2}\) scaling into phase field simulations to set physically motivated boundary conditions. The Weibull distribution prediction offers a testable hypothesis: measure \(J_{\mathrm{crit}}\) across many samples with controlled defect populations and check for Weibull scaling.

Scores: Relevance 4 | Novelty 3 | Usefulness 3 | Total: 10

Topics: Phase Field / Dendrites, Solid Electrolytes

2. MLIP Validation for Electrolyte Solvation¶

Full title: Prediction and Experimental Verification of Electrolyte Solvation Structure from an OMol25-Trained Interatomic Potential

Authors: Nitesh Kumar¹ (SLAC National Accelerator Laboratory), Jianwei Lai¹, Casey S. Mezerkor¹, Jiaqi Wang¹, Kamila M. Wiaderek², J. David Bazak¹, Samuel M. Blau¹ (SLAC National Accelerator Laboratory), Ethan J. Crumlin¹ (SLAC National Accelerator Laboratory)
Link: arXiv:2603.20183
Source: arXiv (physics.chem-ph), Mar 20, 2026
Why it matters: Demonstrates that MLIPs trained on molecular datasets (OMol25) dramatically outperform materials-trained MLIPs for liquid electrolyte structure prediction. This has implications for Yanhao Deng's MLIP work — the choice of training dataset matters as much as architecture.
Key idea: 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.
Method: 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.
What is actually new: 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.
Potential reuse: Yanhao should test whether halide electrolyte MLIPs benefit from fine-tuning with molecular DFT data alongside bulk crystal data. The solvation structure analysis pipeline (CIPs, ion correlations) is directly transferable to investigating \(\ce{Li+}\) environments in polymer electrolytes (Naibing Wu).

Scores: Relevance 4 | Novelty 4 | Usefulness 4 | Total: 12

Topics: ML Interatomic Potentials, Solid Electrolytes

3. Multi-Fidelity MLIPs for Charged Defects¶

Full title: Multi-fidelity Machine Learning Interatomic Potentials for Charged Point Defects

Authors: Xinwei Wang¹ (Imperial College London), Irea Mosquera-Lois², Aron Walsh¹ (Imperial College London)
Link: arXiv:2603.05238
Source: arXiv (cond-mat.mtrl-sci), Mar 5, 2026
Why it matters: Directly addresses a major gap in Jerry's group: current MLIPs fail for charged defects — exactly what Yan Li and Mengke Li study in halide electrolytes. This paper introduces global defect charge embeddings and a multi-fidelity approach that could transform defect MLIP accuracy.
Key idea: 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.
Method: 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 \(\ce{Sb2Se3}\).
What is actually new: 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.
Potential reuse: This is a must-try for halide electrolyte defect work. Yan Li and Mengke Li should apply the multi-fidelity approach to \(\ce{Li3YCl6}\) charged vacancies and antisite defects. The defect charge embedding concept is architecture-agnostic and could be combined with MACE or CHGNet.

Scores: Relevance 5 | Novelty 4 | Usefulness 4 | Total: 13

Topics: ML Interatomic Potentials, Defects & Interfaces

4. Safety Certificates for MLIP Screening¶

Full title: Proof-Carrying Materials: Falsifiable Safety Certificates for Machine-Learned Interatomic Potentials

Authors: Abhinaba Basu¹, Pavan Chakraborty¹
Link: arXiv:2603.12183
Source: arXiv (cond-mat.mtrl-sci), Mar 12, 2026
Why it matters: A single MLIP used as a stability filter misses 93% of DFT-stable materials (recall 0.07). This is a wake-up call for the group's high-throughput screening workflows — any results obtained from single-MLIP screening need formal reliability guarantees.
Key idea: 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.
Method: 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.
What is actually new: 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.
Potential reuse: Jerry should adopt PCM-style adversarial validation for any halide electrolyte screening campaigns. The risk model (AUC-ROC = 0.938) that predicts failures on unseen materials and transfers across architectures is directly usable as a pre-screening reliability filter.

Scores: Relevance 3 | Novelty 5 | Usefulness 4 | Total: 12

Topics: ML Interatomic Potentials, Tools & Workflows

5. Mixture of Experts for MLIPs¶

Full title: Scaling Machine Learning Interatomic Potentials with Mixtures of Experts

Authors: Yuzhi Liu¹ (Princeton University), Duo Zhang¹, Anyang Peng¹, Weinan E¹ (Princeton University), Linfeng Zhang², Han Wang¹ (Princeton University)
Link: arXiv:2603.07977
Source: arXiv (physics.chem-ph), Mar 9, 2026
Why it matters: From Weinan E's group (pioneer of DeePMD). Introduces Mixture-of-Experts (MoE) for MLIPs — a scalable architecture that achieves state-of-the-art accuracy on OMol25, OMat24, and OC20M by learning element-specific expert specialization aligned with periodic-table trends.
Key idea: 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.
Method: Systematic development of MoE and MoLE architectures for MLIPs with different routing strategies (element-wise, configuration-level, global). Benchmarked across three major MLIP datasets.
What is actually new: 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.
Potential reuse: Yanhao Deng should benchmark MoE-MLIP against MACE for halide electrolyte systems. If MoE handles halide elements more effectively through specialized experts, it could improve accuracy for \(\ce{Li3YCl6}\) and related compositions without increasing inference cost.

Scores: Relevance 3 | Novelty 4 | Usefulness 3 | Total: 10

Topics: ML Interatomic Potentials

6. Scalable Attention for Long-Range MLIPs¶

Full title: A Recipe for Scalable Attention-Based MLIPs: Unlocking Long-Range Accuracy with All-to-All Node Attention

Authors: Eric Qu¹, Brandon M. Wood¹, Aditi S. Krishnapriyan², Zachary W. Ulissi¹ (Carnegie Mellon University)
Link: arXiv:2603.06567
Source: arXiv (cs.LG), Mar 6, 2026
Why it matters: Complements the long-range MLIP theme from the March 21 digest (EquiEwald, self-consistent electrostatic MLIPs). AllScAIP takes a data-driven approach to long-range interactions via all-to-all attention, rather than explicit physics-based corrections.
Key idea: 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.
Method: 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.
What is actually new: 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.
Potential reuse: The scaling findings are relevant for Jerry's group when deciding how much DFT training data to generate for halide electrolyte MLIPs. If the group can produce >10M training samples, all-to-all attention may outperform equivariant MACE.

Scores: Relevance 3 | Novelty 4 | Usefulness 3 | Total: 10

Topics: ML Interatomic Potentials

Synthesis¶

Emerging patterns:

Long-range interactions in MLIPs are being tackled from three orthogonal directions. AllScAIP uses data-driven all-to-all attention, the March 21 EquiEwald adds reciprocal-space message passing, and self-consistent electrostatic MLIPs take a charge-density coarse-graining approach. The field is converging on a clear problem but diverging on solutions.
MLIP reliability is under scrutiny. Proof-Carrying Materials reveals that single-MLIP screening has a 93% miss rate for stable materials, while OMol25 MLIP validation shows that training data choice matters as much as architecture. Together these papers make a strong case: MLIP users must validate against experiment or DFT, not trust model predictions blindly.
Defect physics remains the Achilles' heel of foundation MLIPs. The multi-fidelity defect MLIP explicitly demonstrates that current MLIPs fail for charged point defects and provides a concrete fix (charge embeddings + multi-fidelity training). This directly impacts halide electrolyte research where defect properties govern ionic conductivity.

Gaps and limitations:

None of the long-range MLIP papers benchmark on ionic conductors or charged defects. AllScAIP tests on molecular systems and OC20M; EquiEwald tests on water and small clusters. The community desperately needs a benchmark suite for solid electrolyte ion transport.
The dendrite analytical model Paper 1 lacks coupling to electrochemistry — it treats the SE as a passive mechanical barrier without considering electrochemical reactions at the SE/electrode interface.

Contradictions or unexplored directions:

AllScAIP claims physics-based inductive biases reverse at scale, contradicting the core premise behind equivariant architectures (MACE, EquiEwald). This needs reconciliation — is the reversal dataset-dependent, or does it hold for solid-state systems?

Research Ideas¶

Idea 1: Multi-Fidelity MLIP for Halide Defect Formation Energies¶

Based on: Multi-Fidelity MLIPs for Charged Defects, Proof-Carrying Materials
Core hypothesis: Combining defect charge embeddings with multi-fidelity training (PBE + hybrid functional data) will yield MLIPs that predict charged vacancy and antisite formation energies in \(\ce{Li3YCl6}\) within 50 meV of direct hybrid DFT, while standard foundation MLIPs (MACE, CHGNet) will deviate by >200 meV.
Why non-obvious: Current MLIPs are validated on bulk properties, not defect physics. The multi-fidelity approach assumes that hybrid functional accuracy is needed only at defect sites, not everywhere — an untested assumption for halides where defect states may delocalize.
Minimal validation plan: (1) Generate PBE-level defect configurations for \(\ce{Li}\) vacancy, \(\ce{Y}\) vacancy, and \(\ce{Li-Y}\) antisite in \(\ce{Li3YCl6}\) using existing DFT workflow. (2) Add hybrid functional single-point calculations for a subset. (3) Train MACE with and without charge embeddings and multi-fidelity data. (4) Compare formation energies and migration barriers against full hybrid DFT.

Idea 2: Weibull Distribution of Critical Current Density¶

Based on: Dendrite Growth in Ceramic Electrolytes
Core hypothesis: The critical current density for Li penetration through \(\ce{Li3YCl6}\) pellets follows a Weibull distribution whose shape parameter depends on the processing-induced defect population (sintering temperature, pressure).
Why non-obvious: Most experimental studies report single \(J_{\mathrm{crit}}\) values and attribute sample-to-sample variation to noise. The analytical model predicts this variation is intrinsic and informative — it encodes the defect size distribution.
Minimal validation plan: (1) Collaborate with an experimental group to measure \(J_{\mathrm{crit}}\) on ≥20 identically processed \(\ce{Li3YCl6}\) pellets. (2) Fit Weibull distributions and extract shape/scale parameters. (3) Correlate with microstructural characterization (grain size, porosity) from SEM. (4) Validate the \(J_{\mathrm{crit}} \propto c_{\max}^{3/2}\) scaling.

Idea 3: Cross-Architecture Ensemble for Halide Screening¶

Based on: Proof-Carrying Materials, MLIP Validation for Electrolyte Solvation
Core hypothesis: An ensemble of MACE, CHGNet, and a molecular-trained MLIP (e.g., UMA-OMol fine-tuned on halide data) will discover more stable halide compositions than any single model alone, because their error correlations are near-zero (r ≤ 0.13).
Why non-obvious: PCM shows that different architectures have completely different blind spots. For halide screening, this means a composition that MACE flags as unstable might be flagged as stable by CHGNet or UMA-OMol — and vice versa. An ensemble voting scheme should reduce false negatives.
Minimal validation plan: (1) Generate a candidate library of 500 halide compositions (varying cation/anion chemistry). (2) Run stability screening with MACE, CHGNet, and a fine-tuned UMA-OMol. (3) Apply PCM's adversarial falsification to flag uncertain predictions. (4) Validate the top 20 ensemble-disagreement candidates with DFT relaxation.

ML Interatomic Potentials Solid Electrolytes Defects & Interfaces Phase Field / Dendrites Tools & Workflows