Skip to content

March 20, 2026

Papers

1. MLIP Fine-Tuning Bias

Full title: Bias in Universal MLIPs and its Effects on Fine-Tuning

  • Authors: Nicolas Wong¹, Julia H. Yang² [et al.]
  • Link: arXiv:2603.10159
  • Source: arXiv (cond-mat.mtrl-sci), Mar 10, 2026
  • Why it matters: Directly relevant to Yanhao Deng's MLIP work. If you're fine-tuning CHGNet/MACE for halide electrolytes, this paper identifies a concrete failure mode — "naive fine-tuning" generates constrained datasets that don't represent real MD dynamics, causing extrapolation failures.
  • Key idea: Universal MLIPs carry systematic biases that persist into fine-tuned models. Periodic fine-tuning (single trajectory, retrain at intervals) dramatically outperforms naive multi-trajectory fine-tuning.
  • Method: Compared naive vs periodic fine-tuning strategies on uMLIPs, analyzed bias through PCA and Q-residual analysis as epistemic uncertainty proxy.
  • What is actually new: Systematic demonstration that how you generate fine-tuning data matters as much as how much. Periodic retraining from evolving trajectories captures dynamics that static datasets miss. Q-residuals as uncertainty proxy for large simulations.
  • Potential reuse: Adopt periodic fine-tuning protocol for halide MLIP development (Yanhao). Use Q-residual analysis to detect when fine-tuned models are extrapolating unsafely during long MD runs. Directly applicable to the \(\ce{Li3YCl_{6-x}Br_x}\) or other halide MLIP workflows.

Scores: Relevance 5 | Novelty 4 | Usefulness 5 | Total: 14

2. PFP v8 with r2SCAN

Full title: PFP v8: Universal MLIP with r2SCAN Functional

  • Authors: Chikashi Shinagawa¹, So Takamoto¹, Daiki Shintani¹, …, Ju Li² [et al.]
  • Link: arXiv:2603.11063
  • Source: arXiv (physics.chem-ph / cond-mat.mtrl-sci), Mar 9 (v2 Mar 18), 2026
  • Why it matters: Most MLIPs inherit PBE-level errors (~130 K melting point error). PFP v8 trains on r2SCAN meta-GGA — if it works, this is a step-change in accuracy for thermodynamic predictions relevant to electrolyte stability screening.
  • Key idea: Training uMLIPs on r2SCAN instead of PBE reduces systematic functional error. PFP v8 halves melting point prediction error vs PBE-trained models.
  • Method: Trained Matlantis PFP v8 on large r2SCAN dataset across crystals, molecules, and surfaces. Benchmarked zero-shot predictions against experiment and high-level theory.
  • What is actually new: First large-scale demonstration that uMLIPs can systematically outperform PBE-DFT against experiment without domain-specific fine-tuning. ~130 K average melting point error (vs ~260 K for PBE-trained).
  • Potential reuse: Test PFP v8 for halide electrolyte phase stability predictions (Yan Li / Mengke Li) — r2SCAN should give more accurate formation energies and defect chemistry. Benchmark against VASP r2SCAN for \(\ce{Li3YCl6}\) before committing to MLIP-only workflows.

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

3. Fine-Tuned CHGNet for Halide Electrolytes

Full title: Predicting Crystal Structures and Ionic Conductivities in Li₃YCl₆₋ₓBrₓ Using Fine-Tuned CHGNet

  • Authors: Jonas Böhm¹, Aurélie Champagne¹ [et al.]
  • Link: arXiv:2510.09861 (v3 Mar 2, 2026)
  • Source: arXiv (cond-mat.mtrl-sci), updated Mar 2, 2026
  • Why it matters: This is your system — \(\ce{Li3YCl_{6-x}Br_x}\) halide electrolytes. Directly relevant to Yan Li and Mengke Li's work on halide electrolyte transport and degradation.
  • Key idea: Fine-tuned CHGNet on disordered halide structures achieves near-DFT accuracy for total energies and Li-ion dynamics at 10⁴× lower cost. Composition-dependent conductivity trends revealed via ML-MD.
  • Method: Systematic enumeration of ordered structural models from experimentally refined disordered \(\ce{Li3YCl6}\) / \(\ce{Li3YBr6}\), iterative fine-tuning of CHGNet with MD + DFT, NVE/NVT MD for ionic conductivity.
  • What is actually new: Concrete workflow for handling site disorder in halide electrolytes with MLIPs. Demonstrated that Br substitution systematically affects phase stability and conductivity in this ternary system.
  • Potential reuse: The fine-tuning strategy is directly adoptable. Compare their CHGNet approach with MACE for the same system (Yanhao). Their structural enumeration protocol for disordered halides can feed into Mengke Li's transport studies. Potential collaboration opportunity — check if their ordered models capture the same physics as your DFT calculations.

Scores: Relevance 5 | Novelty 3 | Usefulness 5 | Total: 13

4. Latent Space MLIP Design

Full title: Latent Space Design of Interatomic Potentials

  • Authors: Susan R. Atlas¹
  • Link: arXiv:2603.05655
  • Source: arXiv (physics.chem-ph), Mar 5, 2026
  • Why it matters: Proposes a fundamentally different approach to MLIPs — physics-based latent space from DFT constraints rather than purely data-driven. Relevant to Yanhao's work if standard graph-network MLIPs hit accuracy limits for complex electrolyte systems.
  • Key idea: Construct latent space patterns using DFT theorems and analytic constraints, linking electronic and atomic length scales through electron density. Connects ground, excited, and charge-transfer states.
  • Method: Theoretical framework mapping DFT constraints to latent space components, building on ensemble charge-transfer potential formalism.
  • What is actually new: Moves MLIP design from "fit large datasets" to "encode known physics as latent structure." Could address the combinatoric complexity and undiscovered bonding motif problems in current MLIPs.
  • Potential reuse: If the framework becomes implementable, it could improve MLIP accuracy for systems with mixed bonding character (e.g., electrolyte/electrode interfaces where charge transfer matters — Umang's project). Worth monitoring but likely not immediately applicable.

Scores: Relevance 3 | Novelty 5 | Usefulness 2 | Total: 10

5. Bayesian Polymer Electrolyte Discovery

Full title: Discovery of Polymer Electrolytes with Bayesian Optimization and High-Throughput MD

  • Authors: Antonia S. Kuhn¹, Jurğis Ruža¹, KyuJung Jun¹, Pablo Leon¹, Rafael Gómez-Bombarelli¹ (MIT)
  • Link: arXiv:2602.17595
  • Source: arXiv (cond-mat.mtrl-sci), Feb 19, 2026
  • Why it matters: Directly relevant to Naibing Wu's solid polymer electrolyte project. They screened 1.7M polymer candidates and found architectures that outperform PEO/LiTFSI.
  • Key idea: Bayesian optimization-guided high-throughput MD screening of 1.7M hypothetical polymer electrolytes. Branched architectures and ketone functional groups enhance ion-hopping mechanisms.
  • Method: Classical MD simulations (1.7M chemical space), warm-started batch Bayesian optimization, evaluated 767 homopolymers iteratively. Open-source framework provided.
  • What is actually new: The scale (1.7M candidates) and the mechanistic finding that branched + ketone architectures beat PEO. Also provides Li vs Na cation transport comparison in the same framework.
  • Potential reuse: Naibing should look at the top candidates and test with atomistic MD or MLIP-driven MD for more accurate transport predictions. Their open-source framework could be adapted for composite polymer electrolytes. The Li/Na comparison is relevant if Naibing wants to extend to Na systems.

Scores: Relevance 5 | Novelty 3 | Usefulness 5 | Total: 13

6. Autoregressive Lattice Models

Full title: Scaling Autoregressive Models for Lattice Thermodynamics

  • Authors: Xiaochen Du¹, Juno Nam¹, Sulin Liu¹, Rafael Gómez-Bombarelli¹ (MIT)
  • Link: arXiv:2603.14695
  • Source: arXiv (cond-mat.stat-mech), Mar 16, 2026
  • Why it matters: Lattice site disorder (cation/anion mixing, vacancy ordering) is central to solid electrolyte properties. This offers a new way to sample thermodynamically relevant disordered configurations.
  • Key idea: Any-order autoregressive models with marginalization can directly learn thermodynamic distributions on crystal lattices, avoiding critical slowing down of MCMC. Scales from small to larger supercells.
  • Method: Any-order ARMs + marginalization models (MAMs) with Transformer architectures and lattice-aware positional encodings. Validated on Ising model and CuAu alloys.
  • What is actually new: Enables sampling of thermodynamic distributions on lattices without expensive MCMC. Any-order generation means you can condition on known subset of sites (useful for interfaces).
  • Potential reuse: Generate thermodynamically realistic disordered configurations for halide electrolytes (e.g., \(\ce{Li3YCl_{6-x}Br_x}\) site disorder) without exhaustive enumeration. Could replace the enumeration approach in Paper 3 above. Also relevant for Cheng Peng's grain boundary work — generating realistic GB structures with proper site disorder.

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

7. Surrogate Phase Field Models

Full title: Adaptive Uncertainty-Guided Surrogates for Phase Field Dendritic Solidification

  • Authors: Eider Garate-Perez¹, Kerman López de Calle-Etxabe¹, Oihana Garcia¹, Borja Calvo¹, Meritxell Gómez-Omella¹, Jon Lambarri¹ (UPV/EHU)
  • Link: arXiv:2603.00093
  • Source: arXiv (physics.comp-ph), Feb 17, 2026
  • Why it matters: Relevant to Shoutong Jin's phase field dendrite growth work. If his simulations are computationally expensive (multi-physics coupling), surrogate models could accelerate parametric studies.
  • Key idea: ML surrogates (XGBoost + CNN) with uncertainty-driven adaptive sampling to approximate phase field dendritic solidification with far fewer full simulations.
  • Method: Monte Carlo dropout (CNN) and bagging (XGBoost) for uncertainty estimation. Adaptive sampling within hyperspheres around high-uncertainty regions. Compared against OLHS-PSO.
  • What is actually new: Self-supervised strategy for temporal instance selection in spatio-temporal surrogate modeling. Quantifies CO₂ emissions of compute alongside accuracy — a useful framing.
  • Potential reuse: Shoutong could train surrogates for his multi-physics (mechanical-thermal-electrochemical) dendrite model to rapidly explore parameter spaces (applied pressure, temperature, SEI properties). The adaptive sampling idea is generalizable to any expensive PDE solver.

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

8. 2026 Solid Electrolyte Roadmap

Full title: 2026 Roadmap on Next-Generation Solid Electrolytes

  • Authors: Hui-Chia Yu¹, Bernardo Orvananos², Scott Cronin¹, Martin Bazant³ (MIT), Scott Barnett⁴, K. Thornton⁵
  • Link: ChemRxiv (Mar 16, 2026)
  • Source: ChemRxiv
  • Why it matters: Major community roadmap — useful for identifying where the field is heading and positioning your group's contributions.
  • Key idea: Comprehensive roadmap covering sulfide, oxide, polymer, halide electrolytes with emphasis on ML, scalable processing, and high-throughput synthesis as key enablers for the next decade.
  • Method: Expert review / community consensus
  • What is actually new: Identifies ML-driven discovery, multiscale modeling, and interface engineering as the three pillars for solid electrolyte advancement. Highlights gaps in scalable processing and interfacial stability.
  • Potential reuse: Use as a reference for grant proposals and review papers. The identified gaps (scalable processing, air stability, interface engineering) can guide project direction. Position your group's halide MLIP + interface work within this roadmap.

Scores: Relevance 4 | Novelty 2 | Usefulness 4 | Total: 10

Synthesis

Emerging patterns:

  1. Fine-tuning universal MLIPs is now the default strategy for domain-specific solid electrolyte work (Paper 1, Paper 2, Paper 3). The field is converging on: start from pretrained uMLIP → fine-tune with DFT → run long MD. But Paper 1 shows the method of fine-tuning matters critically.

  2. Meta-GGA functionals (r2SCAN) are entering the MLIP training pipeline (Paper 2). This challenges the assumption that PBE-level accuracy is sufficient for thermodynamic screening. If r2SCAN-trained models become standard, everyone's DFT benchmarks need updating.

  3. Bayesian optimization + high-throughput MD is replacing brute-force screening for polymer electrolytes (Paper 5). Similar approaches are likely coming for inorganic electrolytes.

  4. Generative models for lattice disorder (Paper 6) are emerging as an alternative to enumeration + DFT for handling site-disordered solid electrolytes.

Gaps and limitations:

  • Interface modeling is lagging. Most MLIP work focuses on bulk properties. The electrolyte/electrode interface (Umang's project) remains underserved by MLIPs — charge transfer, electrochemical potentials, and interfacial reactions are poorly captured.
  • Grain boundary effects are almost absent from recent MLIP papers. Paper 6 could help generate realistic GB structures, but no one is combining MLIPs with systematic GB property prediction in solid electrolytes.
  • Uncertainty quantification for MLIP-MD is nascent. Paper 1 introduces Q-residuals, but systematic UQ for long-timescale ion transport predictions is missing.

Contradictions and unexplored directions:

  • Paper 2 claims r2SCAN-based uMLIPs beat PBE-based ones without fine-tuning, while Papers 1 and 3 show fine-tuning is still essential for domain-specific accuracy. These aren't contradictory but suggest a hierarchy: better base functional → better zero-shot → but fine-tuning still needed for transport properties. The unexplored question: does r2SCAN fine-tuning converge faster/fewer DFT points needed?

Research Ideas

Idea 1: Periodic Fine-Tuning for Halides

  • Based on: Paper 1, Paper 3
  • Core hypothesis: Periodic fine-tuning of MACE on a single evolving \(\ce{Li3YCl_{6-x}Br_x}\) MD trajectory will converge to DFT-level transport predictions with fewer DFT calculations than naive multi-structure fine-tuning.
  • Why non-obvious: Most groups fine-tune on diverse static structures. This paper suggests less diverse but dynamically consistent data is better — counterintuitive for conventional ML thinking.
  • Minimal validation plan: (1) Fine-tune MACE on \(\ce{Li3YCl6}\) using naive vs periodic protocol with the same DFT budget. (2) Compare predicted ionic conductivities at 300-600 K against DFT-NEB barriers and experimental data.

Idea 2: Generative Disorder Models

  • Based on: Paper 3, Paper 6
  • Core hypothesis: Any-order autoregressive lattice models can generate thermodynamically realistic \(\ce{Li3YCl_{6-x}Br_x}\) configurations that better represent the true disordered state than enumeration-based approaches.
  • Why non-obvious: Enumeration assumes all symmetrically distinct configurations are equally probable. In reality, Boltzmann-weighted disorder matters — some configurations dominate at finite temperature, and these determine transport properties.
  • Minimal validation plan: (1) Train an any-order ARM on \(\ce{Li3YCl6}\) DFT energies for all enumerated orderings of Li/Y sites. (2) Sample thermodynamic ensemble at 300-600 K and compare pair distribution functions / XRD patterns against experiment. (3) Run MLIP-MD on generated structures and compare conductivity with Paper 3's enumeration-based results.

Idea 3: Surrogate-Accelerated Phase Field

  • Based on: Paper 7
  • Core hypothesis: An uncertainty-guided ML surrogate of Shoutong's multi-physics phase field model can identify optimal (pressure, temperature, SEI stiffness) combinations that suppress dendrite penetration with 10× fewer full simulations.
  • Why non-obvious: Phase field parametric studies are normally limited by compute budget to a handful of parameter combinations. Active learning with surrogates enables systematic exploration of high-dimensional parameter spaces.
  • Minimal validation plan: (1) Generate training set from ~50-100 phase field simulations across the (P, T, SEI) parameter space. (2) Train CNN surrogate with uncertainty estimation. (3) Validate on held-out full simulations and compare accuracy vs compute savings.

Idea 4: r2SCAN vs PBE Fine-Tuning

  • Based on: Paper 1, Paper 2
  • Core hypothesis: Fine-tuning an r2SCAN-pretrained uMLIP (PFP v8) requires fewer DFT single-points to achieve target accuracy for halide electrolyte defect formation energies than fine-tuning a PBE-pretrained model.
  • Why non-obvious: r2SCAN better describes the PES → the pretrained model starts closer to the truth → fewer correction points needed. This would mean higher upfront training cost but lower per-project fine-tuning cost.
  • Minimal validation plan: (1) Benchmark PFP v8 (r2SCAN) vs CHGNet/MACE (PBE) on \(\ce{Li3YCl6}\) defect formation energies (vacancies, antisites). (2) Fine-tune both with identical DFT budgets (N=50, 100, 200 points) and compare convergence of defect energy errors.

Idea 5: GB Structure-Property Maps

  • Based on: Paper 6, Paper 8
  • Core hypothesis: Generative polycrystalline models (PolyCrysDiff-style) combined with MLIPs can map grain boundary structure → ionic conductivity for halide electrolytes, identifying which GB types are conductive vs blocking.
  • Why non-obvious: The 2026 roadmap (Paper 8) identifies grain boundary effects as a major gap. Most GB work in solid electrolytes studies a few hand-picked GBs. A data-driven approach could reveal unexpected correlations between GB structure and transport.
  • Minimal validation plan: (1) Generate 50-100 Σ3, Σ5, Σ7 GBs for \(\ce{Li3YCl6}\) using CSL rotation. (2) Compute Li migration barriers across GBs with DFT-NEB (or fine-tuned MLIP). (3) Correlate GB structural descriptors (free volume, misorientation angle, excess energy) with conductivity reduction factor.