Research Digest — 2026-09-23¶
Machine Learning Interatomic Potentials¶
1. Small-supercell and Small-dataset Training Strategy of Machine Learning Interatomic Potentials for Point Defects¶
Source: arXiv (2609.24293) · 📅 2026-09-21 · ↗ Open paper
Proposes an efficient MLIP training scheme for neutral and lowly charged point defects that sidesteps the large defect supercells normally required to suppress finite-size effects and long-range Coulomb artifacts, sharply reducing the DFT data burden. Small-supercell, small-dataset training still recovers accurate defect structures and energetics for charge-stabilized defects.
Relevance to DENG.Group
Directly useful for Cheng Peng (charged defects such as Li vacancies and antisites at grain boundaries, where giant charged supercells are the DFT bottleneck) and Yanhao Deng (defect-rich training sets for halide electrolyte potentials) - this scheme could cut the DFT cost of their defect datasets by an order of magnitude.
2. ECENet: An Edge Cluster Expansion Line-Graph Neural Network¶
Source: arXiv (2609.23134) · 📅 2026-09-19 · ↗ Open paper
Introduces the edge cluster expansion (ECE), an analogue of the atomic cluster expansion in which local environments are expanded around edges between atom pairs rather than single atoms, and builds the line-graph neural network ECENet on top of it. The design targets accuracy competitive with O(3)-equivariant graph MLIPs at substantially lower computational cost.
Relevance to DENG.Group
A cheaper non-equilibvariant alternative matters for long-MD studies of compositionally disordered halide electrolytes (Yan Li, Mengke Li): edge-based expansions echo cluster-expansion thinking for Br/Cl mixing and antisite disorder, and Yanhao Deng should benchmark cost-accuracy against MACE-style models before committing to production trajectories.
3. Colour me shocked: Exact Molecular Hessians from local MLIPs in O(N) time using sparse differentiation!¶
Source: arXiv (2609.24720) · 📅 2026-09-21 · ↗ Open paper
Derives the closed-form sparsity pattern of an MLIP's Hessian and exploits it with sparse differentiation to compute exact molecular Hessians in O(N) time, removing the traditional O(N^2) bottleneck of N Hessian-vector products. This makes vibrational analysis of large MLIP-described systems practical at near-linear cost.
Relevance to DENG.Group
Phonons at scale are a missing piece in several group workflows: Mengke Li (thermal properties and transport mechanisms of halide electrolytes) and Yanhao Deng (dynamical-stability screening of candidate potentials) can now contemplate finite-difference-free Hessian spectra for large supercells, interfaces, and grain-boundary structures rather than small primitive cells.
Orbital-Free & Electronic-Structure ML Methods¶
4. Orbital-Free Surrogate Functionals Yield Transferable Interatomic Potentials and Electron Densities¶
Source: arXiv (2609.21882) · 📅 2026-09-18 · ↗ Open paper
Extends machine-learned orbital-free density functional theory by introducing (weak) surrogate functionals that make the density-optimization convergence robust to hyperparameter and architectural choices, yielding transferable interatomic potentials and electron densities. This is a step toward scalable electronic-structure calculations computed directly from the electron density, without orbitals.
Relevance to DENG.Group
Worth tracking as the methods horizon beyond MLIPs: orbital-free ML-DFT could eventually handle dense metallic lithium, molten SEI-like phases, and charged interfaces where ordinary MLIPs discard electronic structure - context for Umang Agarwal's electrolyte/electrode interface work and Yanhao Deng's potential-development toolkit.