Skip to content

March 21, 2026

Papers

1. An SO(3)-Equivariant Reciprocal-Space Neural Potential for Long-Range Interactions

  • Authors: Linfeng Zhang¹ (CUHK / Tencent AI Lab), Taoyong Cui¹, Dongzhan Zhou¹, Lei Bai¹, Sufei Zhang¹, Luca Rossi¹, Mao Su¹, Wanli Ouyang¹, …, Pheng-Ann Heng¹ (CUHK / Tencent AI Lab)
  • Link: arXiv:2603.18389
  • Source: arXiv (physics.chem-ph), Mar 19, 2026
  • Why it matters: Directly relevant to Yanhao Deng's MLIP work. Long-range electrostatics are critical for solid electrolytes (\(\ce{Li3YCl6-xBrx}\), argyrodites, NASICON) where \(\ce{Li+}\) migration depends on the electrostatic potential landscape. Current MLIPs (MACE, NequIP) use local environments only — this paper introduces a principled reciprocal-space extension that captures anisotropic multipolar correlations while preserving \(\ce{SO(3)}\) equivariance and energy-force consistency.
  • Key idea: EquiEwald performs equivariant message passing in reciprocal space using learned k-space filters, embedding an Ewald-inspired formulation inside an \(\ce{SO(3)}\)-equivariant neural network. This captures long-range electrostatics without sacrificing the physical consistency that makes modern equivariant MLIPs powerful.
  • Method: \(\ce{SO(3)}\)-equivariant architecture with reciprocal-space message passing, equivariant k-space filters, and equivariant inverse Fourier transform. Benchmarked on periodic and aperiodic systems against ab initio reference data.
  • What is actually new: First MLIP architecture to combine true reciprocal-space long-range treatment with \(\ce{SO(3)}\) equivariance in a unified framework. Previous long-range extensions either broke equivariance or sacrificed energy-force consistency. The Ewald-inspired reciprocal-space approach is physically principled rather than ad hoc.
  • Potential reuse: Test EquiEwald for halide electrolyte MLIPs (Yanhao) — if it captures long-range electrostatic effects better than local-only MACE, it could improve ionic conductivity predictions, especially for charged defects and interfaces where electrostatics dominate. Also relevant for Umang's electrolyte/electrode interface work where charge transfer and image charge effects matter.

Scores: 🔬 Relevance 4 | 💡 Novelty 5 | 🛠 Usefulness 4 | Total: 13

2. Design Space of Self-Consistent Electrostatic Machine Learning Interatomic Potentials

  • Authors: William J. Baldwin¹ (Cambridge), Ilyes Batatia² (Cambridge), Martin Vondrák¹ (Cambridge), Johannes T. Margraf³, …, Gábor Csányi¹ (Cambridge)
  • Link: arXiv:2603.14700
  • Source: arXiv (physics.chem-ph), Mar 16, 2026
  • Why it matters: From Gábor Csányi's group (creators of MACE). Directly relevant to Yanhao's work — provides a systematic framework for incorporating electrostatics into MLIPs using the MACE architecture. Addresses the same challenge as EquiEwald but from the charge-density coarse-graining perspective.
  • Key idea: Existing electrostatic MLIPs are coarse-grained approximations to DFT. By making this explicit, the paper identifies the full design space of self-consistent electrostatic MLIPs, reveals equivalences between previously proposed models, and identifies fundamental limitations. Implemented within MACE for controlled comparisons.
  • Method: Theoretical framework viewing electrostatic MLIPs as DFT coarse-graining. Shared charge density representation within MACE architecture. Benchmarked on metal-water interfaces (conducting vs insulating) and charged vacancies in \(\ce{SiO2}\).
  • What is actually new: A unifying theoretical framework that maps all existing electrostatic MLIP approaches onto a common design space. Demonstrates that current approaches fail on metal-water interfaces and charged vacancies — systems where self-consistency is essential. Shows that more expressive self-consistent models are needed beyond simple charge prediction.
  • Potential reuse: Use the framework to design electrostatic extensions for the halide MLIPs (Yanhao). The charged vacancy benchmark (\(\ce{SiO2}\)) is directly analogous to charged point defects in halide electrolytes (Yan Li / Mengke Li's work). The MACE implementation means this can be tested immediately without switching architectures. Combined with EquiEwald, there are now two complementary approaches to long-range MLIPs worth benchmarking.

Scores: 🔬 Relevance 4 | 💡 Novelty 4 | 🛠 Usefulness 4 | Total: 12

3. Revealing Hydroxide Ion Transport Mechanisms in Anion-Exchange Membranes from MLIP Simulations

  • Authors: Jonas Hänseroth¹ (TU Dresden), Muhammad Nawaz Qaisrani¹, Mostafa Moradi¹, Karl Skadell¹, …, Christian Dreßler¹ (TU Dresden)
  • Link: arXiv:2603.13705
  • Source: arXiv (cond-mat.mtrl-sci), Mar 14, 2026
  • Why it matters: Methodologically relevant to Naibing Wu's polymer electrolyte work and Mengke Li's ion transport studies. Demonstrates the fine-tuned MLIP workflow for ion transport in a complex heterogeneous membrane — a system analogous to polymer electrolytes where ion transport occurs through a nanostructured medium.
  • Key idea: Fine-tuned MLIPs enable large-scale MD simulations (\(>10\) ns, \(>10\) nm) of hydroxide ion transport in a commercial anion-exchange membrane. Water content controls the percolation transition from isolated clusters to connected hydrogen-bond networks that enable long-range proton transfer.
  • Method: Fine-tuned machine-learned interatomic potential for membrane systems. Large-scale MD simulations at varying water content. Computed diffusion coefficients and activation energies, validated against experiment.
  • What is actually new: First MLIP-based study of ion transport in commercial anion-exchange membranes at realistic length and time scales. Shows the water-content-driven percolation transition that governs transport, reproducing experimental diffusion trends.
  • Potential reuse: The fine-tuning workflow for membrane systems is directly transferable to Naibing's polymer electrolyte work. The structural transport analysis (percolation of hydrogen-bond networks) parallels what's needed for understanding \(\ce{Li+}\) transport in PEO-based and composite polymer electrolytes. Could inspire similar water/salt-content studies in polymer electrolytes.

Scores: 🔬 Relevance 3 | 💡 Novelty 4 | 🛠 Usefulness 3 | Total: 10

4. GPUMDkit: A User-Friendly Toolkit for GPUMD and NEP

  • Authors: Zihan Yan¹, Denan Li¹, Xin Wu¹, Zhoulin Liu¹, Chen Hua¹, Boyi Situ¹, Hao Yang¹, Shengjie Tang¹, Benrui Tang¹, Ziyang Wang¹, Shangzhao Yi¹, Huan Wang¹, Dian Huang¹, Ke Li¹, Qilin Guo¹, Zherui Chen¹, Ke Xu¹, Yanzhou Wang¹, Ziliang Wang¹, Gang Tang¹, Shi Liu¹, Zheyong Fan² (Xiamen University), …, Yizhou Zhu¹ (Xiamen University) [et al.]
  • Link: arXiv:2603.17367
  • Source: arXiv (cond-mat.mtrl-sci), Mar 18, 2026
  • Why it matters: Practical tool for the group. GPUMD/NEP is one of the fastest MLIP MD engines available — useful if the group wants to run large-scale MLIP-MD for ion transport (Mengke Li, Naibing Wu). GPUMDkit streamlines the entire workflow from training data preparation to trajectory analysis.
  • Key idea: Comprehensive toolkit that automates format conversion, structure sampling, active learning, property calculation, and data visualization for GPUMD and NEP, with both interactive and command-line interfaces.
  • Method: Modular Python toolkit wrapping GPUMD and NEP workflows. Supports format conversion, structure sampling, property calculation, and visualization.
  • What is actually new: Lowers the barrier to using GPUMD/NEP for MLIP development and MD simulation. The active learning integration and automated property calculation pipeline could save significant setup time.
  • Potential reuse: Tim (research engineer) could evaluate GPUMDkit as an alternative/complement to LAMMPS for MLIP-driven MD. NEP potentials can be trained on the same DFT datasets as MACE — benchmarking NEP vs MACE for halide electrolyte ionic conductivity predictions would be informative. The toolkit could accelerate the active learning workflow for Yanhao's MLIP development.

Scores: 🔬 Relevance 3 | 💡 Novelty 2 | 🛠 Usefulness 4 | Total: 9

Synthesis

Emerging patterns:

  1. Long-range electrostatics in MLIPs is having a moment. Two major papers this week (EquiEwald and self-consistent electrostatic MLIPs) address the same fundamental limitation from complementary angles — reciprocal-space message passing vs charge-density coarse-graining. Combined with yesterday's bias-in-fine-tuning and PFP v8 papers, the MLIP field is rapidly converging on models that can handle the electrostatic complexity of real materials.

  2. Fine-tuned MLIPs for ion transport in soft/complex media is becoming routine. The anion-exchange membrane paper follows the same playbook as yesterday's polymer electrolyte Bayesian screening: train/fine-tune an MLIP → run large-scale MD → extract transport mechanisms. This workflow is now proven across membranes, polymer electrolytes, and inorganic solid electrolytes.

Gaps and limitations:

  • Both electrostatic MLIP papers lack battery-relevant benchmarks. EquiEwald tests on molecular and generic condensed-phase systems; the Cambridge group tests on metal-water interfaces and \(\ce{SiO2}\) vacancies. Neither benchmarks on ionic conductors, charged defects in halides, or electrochemical interfaces — exactly the systems where long-range electrostatics matter most for Jerry's group.
  • GPUMD/NEP ecosystem remains separate from the MACE/NequIP ecosystem. There's growing fragmentation in MLIP tools. Groups need to choose between MACE (more expressive, better equivariance) and NEP (faster, GPU-native). Cross-architecture benchmarking on the same system (e.g., \(\ce{Li3YCl6}\)) is badly needed but rarely done.

Contradictions and unexplored directions:

  • EquiEwald takes a reciprocal-space approach while the Cambridge group works in real space with self-consistent charges. Both claim to solve the long-range problem, but they're fundamentally different — EquiEwald is more physically principled (Ewald formalism) while the self-consistent approach is more flexible. Which works better for solid electrolytes? Nobody has tested this yet. The ideal test case: \(\ce{Li+}\) migration through a halide electrolyte with and without charged grain boundaries.

Research Ideas

Idea 1: Benchmark Long-Range MLIP Extensions for Halide Electrolyte Defect Chemistry

  • Based on: EquiEwald, Self-consistent electrostatic MLIPs
  • Core hypothesis: Both EquiEwald and self-consistent electrostatic MACE will significantly outperform standard local-only MACE for predicting charged defect formation energies and migration barriers in \(\ce{Li3YCl6}\), but the two approaches will differ in which types of defects they handle best (EquiEwald for long-range polarization, self-consistent MACE for charge-transfer defects).
  • Why non-obvious: Current MLIP benchmarks focus on bulk energies and forces. Defect properties involve subtle long-range electrostatic rearrangements that local models fundamentally cannot capture — but nobody has quantified how much this matters for halide electrolytes specifically.
  • Minimal validation plan: (1) Compute \(\ce{Li}\) vacancy, \(\ce{Y}\) vacancy, and antisite defect formation energies in \(\ce{Li3YCl6}\) with standard MACE, self-consistent electrostatic MACE (use their code), and EquiEwald. (2) Compare against DFT reference values. (3) If differences exceed 50 meV/defect, long-range electrostatics matters and the group should adopt one of these approaches.

Idea 2: Percolation-Driven Ion Transport in Composite Polymer Electrolytes

  • Based on: Anion-exchange membrane MLIP study
  • Core hypothesis: In composite polymer electrolytes (Naibing's project), \(\ce{Li+}\) transport undergoes a percolation transition controlled by the connectivity of ion-conducting pathways through the polymer-filler interface, analogous to the water-content-driven percolation in anion-exchange membranes.
  • Why non-obvious: Most composite polymer electrolyte studies focus on the filler as a static structural modifier. The membrane study shows that transport is governed by dynamic network connectivity — applying this framework to polymer-filler composites could reveal that filler loading has a threshold effect on conductivity that isn't captured by simple tortuosity models.
  • Minimal validation plan: (1) Use a fine-tuned MLIP to run MD of PEO-\(\ce{LiTFSI}\) with varying LLZO nanoparticle loadings. (2) Analyze \(\ce{Li+}\) trajectory networks and compute percolation probability as a function of filler content. (3) Check for a sharp conductivity transition at a critical loading fraction.

Idea 3: NEP vs MACE Benchmark on \(\ce{Li3YCl6}\) Ionic Conductivity

  • Based on: GPUMDkit
  • Core hypothesis: NEP (trained via GPUMDkit) will match or exceed MACE accuracy for \(\ce{Li3YCl6}\) ionic conductivity predictions at 10× lower computational cost for MD, but will be less accurate for off-equilibrium properties like defect formation energies.
  • Why non-obvious: NEP is designed for speed with graph-based message passing, while MACE is designed for expressiveness with higher-body-order equivariance. For transport properties that depend on the quality of the potential energy surface along migration paths, the faster model might actually perform better due to enabling longer simulation times and better statistics.
  • Minimal validation plan: (1) Train NEP and MACE on identical DFT datasets for \(\ce{Li3YCl6}\). (2) Run NVT-MD at 400-700 K with both models. (3) Compare predicted ionic conductivities, activation energies, and \(\ce{Li+}\) diffusion coefficients against experimental data. (4) Measure wall-clock time per MD step for both.