Research Digest — 2026-05-25¶
ML Interatomic Potentials & Workflows¶
1. TriForces: Augmenting Atomistic GNNs for Transferable Representations¶
Source: arXiv:2605.20581 (Accepted at ICML 2026) · 📅 2026-05-20 · ↗ Open paper
Presents TriForces, a model-agnostic three-stream framework that separates composition and structure information in atomistic GNNs, combined with self-supervised learning to preserve transferable representations. On the OMat24 benchmark in the limited-data regime, TriForces reduces energy MAE by 57% at 20K samples and improves force MAE across all sample sizes compared to baselines. It also enables efficient similar-structure retrieval through its learned latent space, without requiring DFT labels during pre-training.
Relevance to DENG.Group
Directly relevant to Yanhao Deng's MLIP development work. The improved transferability of TriForces addresses a critical challenge in deploying MLIPs for battery materials, where training data is often limited. The model-agnostic design means it can be applied on top of the group's existing MACE or Allegro models for halide and oxide electrolytes, potentially improving accuracy on out-of-distribution configurations encountered at grain boundaries and interfaces.
2. Dataset-Aware Entropy-Maximized Active Learning for Machine-Learned Interatomic Potentials¶
Source: arXiv:2605.20384 · 📅 2026-05-19 · ↗ Open paper
Proposes an active learning framework that combines local entropy-driven molecular dynamics with global dataset-aware filtering to efficiently select training structures for MLIPs. The method uses a log-determinant of fingerprint covariance to select only configurations providing genuinely new information, employing dual covariance modes for ordered and disordered phases. Demonstrated on carbon, silicon, and NaCl systems, entropy-driven sampling achieves 3–10× lower energy MAE compared to random MD sampling at matched training set sizes of 100–800 structures.
Relevance to DENG.Group
Highly relevant to Yanhao Deng's MLIP training pipeline. The dramatic improvement in data efficiency means the group could achieve high-quality potentials for complex solid electrolyte systems (halides, sulfides) with far fewer expensive DFT calculations. The dual covariance modes for ordered and disordered phases are particularly valuable for modeling doped halide electrolytes where partial occupancies and configurational disorder are common.
3. Lang2MLIP: End-to-End Language-to-Machine Learning Interatomic Potential Development with Autonomous Agentic Workflows¶
Source: arXiv:2605.14527 · 📅 2026-05-14 · ↗ Open paper
Introduces Lang2MLIP, a multi-agent framework that takes natural-language input and formulates end-to-end MLIP development as a sequential decision-making problem solved by large language models. At each step, a decision-making agent observes the current dataset, model, and evaluation results, then automatically selects an appropriate action to improve the model—removing the need for a predefined pipeline. The framework is evaluated on a solid electrolyte interphase (SEI) system with multiple components and interfaces.
Relevance to DENG.Group
Directly relevant to Yanhao Deng's computational workflow. Lang2MLIP's application to SEI systems with multiple components and interfaces closely mirrors the Deng group's challenge of building MLIPs for complex electrode–electrolyte interfaces. The agentic self-correction capability, where the system revisits earlier stages when failures arise, could automate the iterative training loop that is currently a major bottleneck in the group's MLIP development for halide and sulfide electrolyte systems.
4. Upscaling DFT-Trained Machine-Learning Interatomic Potential toward Quantum Monte Carlo Accuracy: Sulfur-Vacancy Migration in Monolayer MoS₂ as a Testbed¶
Source: arXiv:2605.22601 · 📅 2026-05-21 · ↗ Open paper
Demonstrates a multi-fidelity approach to train MLIPs at quantum Monte Carlo (QMC) accuracy by fine-tuning only the readout layers of a DFT-trained MACE model using a limited dataset of QMC energies alongside DFT forces. Applied to sulfur vacancy migration in monolayer MoS₂, the QMC-fine-tuned potential significantly improves both energetics and atomic forces over the DFT baseline. The method opens the door to large-scale near-QMC-quality simulations that would be impossible with brute-force QMC.
Relevance to DENG.Group
Relevant to Yanhao Deng's MLIP accuracy goals. The multi-fidelity approach of fine-tuning only readout layers with limited high-level data is a practical strategy the group could adopt to improve the accuracy of their battery material MLIPs beyond DFT level, particularly for critical configurations like transition states for Li⁺ migration. The methodology could be applied to upgrade existing DFT-trained potentials for halide electrolytes using higher-level calculations on a small, targeted subset of configurations.
Solid Electrolytes & Ionic Conductors¶
5. Ordered–Disordered Ionic Cocrystalline Solid-State Electrolytes for Rapid Ion Migration in Sodium Metal Batteries¶
Source: Journal of the American Chemical Society (10.1021/jacs.6c01095) · 📅 2026-05-22 · ↗ Open paper
Reports an ionic cocrystalline solid-state electrolyte, NaClO₄(SN)₃, featuring a unique ordered–disordered hybrid lattice with an ordered Na⁺-coordination backbone and orientationally disordered succinonitrile molecules serving as ionic pathways. The electrolyte achieves 0.94 mS cm⁻¹ ionic conductivity at 25 °C with a low activation energy of 0.26 eV, an electrochemical stability window beyond 4.6 V vs Na/Na⁺, and a melting point of 36.2 °C enabling in situ melting infiltration into electrodes for conformal interfaces.
Relevance to DENG.Group
Highly relevant to the Deng group's expanding work on sodium solid-state batteries. The ordered–disordered cocrystal engineering strategy provides a new design principle that Mengke Li and Yan Li could explore computationally. The in situ melting infiltration approach addresses the interfacial contact challenge that is critical for solid-state battery performance. The group's computational tools could be used to screen for alternative cocrystal compositions with similar or improved Na⁺ conductivity.
6. Phosphonium Poly(Ionic Liquid) Electrolytes for Fast Lithium-Ion Conduction¶
Source: Journal of the American Chemical Society (10.1021/jacs.6c02428) · 📅 2026-05-22 · ↗ Open paper
Introduces a new class of solid polymer electrolytes based on phosphonium poly(ionic liquid)s that outperform their ammonium counterparts. The poly(DADMP)FSI:LiFSI 1:1.5 composition achieves 1.5 × 10⁻³ S cm⁻¹ at 80 °C and 1.5 × 10⁻⁴ S cm⁻¹ at 30 °C, with a high lithium transference number of 0.6–0.7 and electrochemical stability beyond 4.5 V. MD simulations reveal that the larger, more flexible phosphonium cation enables weaker cation–anion interactions and faster ion transport compared to ammonium analogues.
Relevance to DENG.Group
Relevant to the group's interest in solid polymer electrolytes. The phosphonium poly(IL) design principle could inspire computational screening of alternative cation chemistries for polymer electrolytes. The MD simulation methodology used to understand ion coordination environments and transport mechanisms parallels the group's computational approach. The finding that single-anion (FSI) systems outperform mixed-anion (FSI/TFSI) systems in phosphonium backbones provides a useful reference for the group's work on anion effects in electrolyte design.
Defects & Interfaces¶
7. AI-Screened Small-Molecule Templating Effect Enabling 2D Architectures for Dendrite-Free Lithium Metal Batteries¶
Source: Matter (10.1016/j.matt.2026.00079) · 📅 2026-05-22 · ↗ Open paper
Develops an AI-assisted screening workflow that identified sucrose and citric acid as a synergistic molecular couple guiding 2D crystallization of Li₆.₂₅Al₀.₂₅La₃Zr₂O₁₂ (LALZO) nanosheets for composite polymer electrolytes. The resulting brick-and-mortar-like architecture establishes continuous Li⁺ transport highways while providing robust physical barriers against dendrite growth. The LNSs@PEO composite enables stable cycling over 3,000 h in lithium metal cells, with full cells retaining 96.7% capacity after 300 cycles.
Relevance to DENG.Group
Directly relevant to Cheng Peng's interface and dendrite studies. The AI-driven molecular screening approach for designing 2D ceramic fillers could be adapted for computational screening of templating molecules for halide electrolyte synthesis. The brick-and-mortar architecture provides a concrete mechanical model that could be simulated using the group's computational tools to understand how 2D fillers deflect dendrite growth. The 3,000-hour cycling stability demonstrates the practical impact of combining continuous Li⁺ pathways with mechanical reinforcement.
8. Saddle-Node Bifurcation in Interfacial Morphology Selects Battery Degradation Phase¶
Source: arXiv:2605.10252 · 📅 2026-05-11 · ↗ Open paper
Proposes a minimal nonlinear closure ODE for the dynamic active-area factor of a battery interface that exhibits a saddle-node bifurcation separating a smooth passivating phase from a morphologically unstable phase. Maps four canonical anode configurations (graphite, silicon composite, lithium metal, and anode-free Li/Cu) onto the closure using long-cycle experimental data, finding that the anode-free configuration sits within 5% of the critical threshold. Derives three falsifiable predictions (critical current density, critical temperature shift, and critical-slowing-down exponent) consistent with available data.
Relevance to DENG.Group
Relevant to the group's phase-field and dendrite growth modeling work. The saddle-node bifurcation framework provides a new analytical tool for understanding the sharp transition between stable and unstable deposition in solid-state batteries. The finding that anode-free configurations sit near the critical threshold has direct implications for the group's work on lithium metal and anode-free solid-state battery designs. The analytical predictions could be validated against the group's phase-field simulations of dendrite growth in solid electrolytes.