Research Digest — 2026-08-25¶
ML Potentials & Simulation Methods¶
1. Universal Machine-learning Molecular Dynamics at the Speed of Empirical Potentials¶
Source: Unknown · 📅 · ↗ Open paper
Introduces DPA4C, a family of equivariant universal ML interatomic potentials whose architecture and compressed CUDA operators are co-designed for deployment speed. The largest variant approaches MACE-Omat accuracy at ~100x higher measured throughput, while the smallest cuts energy/force/stress errors of the fastest existing universal MLIP by 34-61% at 1.9x its throughput. All variants run multimillion-atom MD on a single GPU, with a 2-billion-atom demonstration at 83-91% weak-scaling efficiency on 1,024 V100s.
Relevance to DENG.Group
Directly relevant to Yanhao's ML potential program — DPA4C is the next iteration of the DPA lineage the group likely builds on, and its speed/accuracy frontier defines the new baseline for universal-potential benchmarking. Also matters for anyone in the group wanting large-scale MD of electrolyte interfaces and grain boundaries at DFT-level accuracy.
2. Machine Learning Guided Discovery of Corundum High Entropy Oxides¶
Source: Unknown · 📅 · ↗ Open paper
Uses ML interatomic potentials to predict synthesizability of A2O3 high-entropy oxides from ~500 candidate trivalent-cation combinations, identifying 16 for experimental validation via solid-state and combustion synthesis. Discovers three new corundum HEOs and a novel cation-ordered phase, but finds that HEO occurrence is far rarer than combinatorial arguments suggest and strongly synthesis-method dependent (only 3 of 16 compositions gave equivalent outcomes across both routes).
Relevance to DENG.Group
Methodological template for the group's ML-driven discovery ambitions: MLIP-based synthesizability screening plus targeted experimental validation, including honest accounting of synthesis-route dependence. The high-entropy framework also connects to recent high-entropy halide electrolyte directions (entropy-stabilized disorder is where ML potentials shine).
3. Atomistic Structure Generation and Neural-Network Screening of Hard Carbons to Identify High-Capacity Sodium Storage¶
Source: Unknown · 📅 · ↗ Open paper
Combines universal machine-learned interatomic potentials with the RAFFLE structure-generation framework to build 13,096 realistic hard-carbon models (up to 4,378 atoms) matching measured densities, porosities, and sp2/sp3 fractions. Explicit Na intercalation reproduces sloping-to-plateau voltage profiles, and a lightweight NN surrogate trained on frozen universal-potential descriptors plus void features screens the library, identifying candidates exceeding 800 mAh/g that survive full intercalation validation.
Relevance to DENG.Group
The structure-generation + uMLIP + surrogate-screening workflow is directly transferable to the group's disordered electrolyte modeling (polymer composites, amorphous/glassy electrolytes, grain-boundary networks). Demonstrates how to bridge the gap between realistic disordered structures and affordable high-throughput property prediction — a pattern Naibing and Yanhao could adopt.
Solid-State Batteries: Dendrites & Electrolyte Design¶
4. Extracting a nitrile-centered, ether-assisted motif hierarchy for lithium-battery electrolyte design from billion-scale molecular space¶
Source: Unknown · 📅 · ↗ Open paper
Screens nearly one billion GDB13 structures for lithium-battery electrolyte molecules using electronic-solvation descriptors without scaffold constraints, finding that high-ranking candidates organize into a nitrile-dominant regime and a nitrile/ether coexistence regime. Encodes this motif hierarchy in a generative model to expand beyond GDB13, producing high-scoring fluorinated structures without an explicit fluorination reward. Explicit-solvent MD confirms weak, exchangeable coordination of representative candidates around Li+.
Relevance to DENG.Group
While focused on liquid electrolytes, the interpretable motif-hierarchy extraction from massive molecular screening is a nice methodology contrast to the group's solid-electrolyte discovery work — relevant to how we frame design rules (motifs/descriptors) rather than just ranked candidates. Useful context for Umang's interface chemistry and for any future electrolyte-molecule screening projects.