Entirely AI-generated and automatically published after automated validation. Not independently human fact-checked. Automated checks do not establish scientific accuracy.

This edition highlights differentiable crystal design, uncertainty-aware interatomic potentials, and AI methods for modeling materials properties and calculations. All cited works are arXiv preprints, and the available evidence is limited to their abstracts; reported results have not been independently assessed here.

DEFEAT enables gradient-based optimization of crystal composition, occupancy, atomic positions and lattice parameters using a pretrained Equiformer model. In its preprint-reported searches, calculations identified 39 of 45 novel superconducting designs with predicted isotropic Allen–Dynes critical temperatures above 10 K, and ten photovoltaic candidates with calculated band gaps near the target and energies within 70 meV/atom of the DFT hull.

Why it matters: Optimizing composition and structure together could make inverse design more direct than searching only among fixed candidate structures. The calculations also show how functional-dependent band-gap errors and finite-temperature phase competition can affect which candidates look promising.

Limitations: This is a preprint, not peer-reviewed evidence. The reported outcomes are computational predictions, not experimental demonstrations of superconductivity or photovoltaic performance; the authors also report that finite-temperature stability can reorder candidates.

Inverse design of superconductors and semiconductors with differentiable foundation models — Preprint; Abstract only.

Uncertainty methods address model misspecification in interatomic potentials

Two preprints explore uncertainty estimates that account for errors caused by imperfect model assumptions. One adds user-specified physical priors to POPS-Regression ensembles and demonstrates the approach for phononic vibrations and elastic properties. The other uses a jackknife-conformal method for linear ACE potentials, reporting near-target force coverage on held-out silicon data and property intervals for 21 silicon properties, while also identifying under-coverage for screw dislocations and relaxed point defects.

Why it matters: Uncertainty estimates that expose model failures could help researchers judge when a potential is suitable for a simulation, rather than relying on confidence estimates that overlook shared biases or physical constraints.

Limitations: Both works are preprints, and the reported evaluations are specific to their methods and test cases. The jackknife study reports failures on some defect and dislocation settings; the physical-prior approach is demonstrated on selected quantities and does not establish universal calibration.

Physics Informed Priors for Machine Learning Interatomic Potentials with Uncertainty Quantification — Preprint; Abstract only.

Jackknife-conformal uncertainty for linear ACE potentials: calibrated forces and silicon properties compared with Bayesian and POPS baselines — Preprint; Abstract only.

Fine-tuned MACE model probes temperature-dependent ion motion in Li₆PS₅Cl

A preprint uses molecular dynamics with a fine-tuned MACE model to study lithium-ion dynamics in the superionic conductor Li₆PS₅Cl. The authors report localized, phonon-like lithium motion below 500 K, more complex competition between lithium and framework motion at 500–700 K, and different behavior near 800 K.

Why it matters: The work illustrates how machine-learned potentials can support temperature-dependent dynamical analysis in an ionic conductor, where collective motion and framework vibrations both matter.

Limitations: These are simulation results reported in a preprint, not direct measurements. The available abstract does not establish how the model’s accuracy transfers beyond the studied material or conditions.

Temperature-Dependent Ion Dynamics in Argyrodite Superionic Conductor Li$_{6}$PS$_{5}$Cl with Fine-Tuned MACE Foundation Model — Preprint; Abstract only.

DP-Learner predicts electronic-structure Hamiltonians in operator space

DP-Learner performs equivariant message passing directly on Hamiltonian-like blocks in a direct-product representation. Its authors report improved accuracy and inference efficiency over a state-of-the-art method on two datasets, and introduce MP-HAM-110K, a dataset with more than 110,000 structures across 68 elements.

Why it matters: Predicting Hamiltonians efficiently could support downstream electronic-structure calculations, while a large dataset may provide a useful benchmark for this class of models.

Limitations: The results and dataset description are from a preprint; the abstract does not provide enough detail to assess dataset availability, benchmark comparability or performance across applications beyond those evaluated.

Equivariant Deep Learning in Direct-Product Space for Electronic-Structure Hamiltonian Prediction — Preprint; Abstract only.

AI forecasts grain growth under changing heat-treatment schedules

A thermally conditioned neural-network framework forecasts grain-size distributions under non-isothermal thermal histories. Trained and evaluated on data from validated front-tracking simulations, it reports mean relative errors of 2.99% for one-hour cycles and 4.51% for four-hour cycles in cross-validation, with errors of 1.91%–14.66% on four unseen thermal histories.

Why it matters: Fast forecasts could help explore complex heat-treatment schedules without running a costly simulation for every trajectory, potentially aiding microstructure process design.

Limitations: This preprint evaluates predictions against simulation-generated data, not experimental grain-growth measurements. The abstract reports a limited set of unseen histories, so performance on other materials and processing conditions remains uncertain.

Thermally Conditioned Time-Series Forecasting of Grain Growth in Polycrystalline Materials — Preprint; Abstract only.

DefectSkills packages guidance for AI-assisted defect calculations

DefectSkills is a set of structured prompts intended to guide AI agents through first-principles point-defect calculations in semiconductors and insulators using VASP. The workflow covers setup, charged-defect calculations, corrections, formation energies, analysis and data backup.

Why it matters: Reusable, domain-specific guidance could help make agent-assisted computational workflows more consistent and reproducible, while keeping human expertise in the process.

Limitations: The work is a preprint describing prompts and workflow guidance; the abstract does not report a quantitative evaluation of agent reliability or demonstrate that the prompts prevent calculation errors across systems.

DefectSkills: Guiding AI agents through first-principles calculations for point defects in semiconductors and insulators — Preprint; Abstract only.

Corrosion model combines historical tests with salt-preparation context

A preprint compiles 291 reported molten-chloride corrosion experiments and trains an interpretable machine-learning framework. A graded salt-preparation index captures differences in treatment, atmosphere and impurity conditions; the reported blended tree ensemble achieves cross-validated R² of 0.85 and mean absolute error of 0.41 mg cm⁻² day⁻¹.

Why it matters: Representing experimental context explicitly may help models make more useful comparisons across historical corrosion studies and highlight where standardized reporting or new experiments are needed.

Limitations: The dataset is assembled from literature reports, so inconsistent measurements and unreported factors may affect results. Cross-validation on this dataset does not establish predictive performance for new alloys, salts or laboratories.

Data-Driven Modeling of Molten Chloride Corrosion Using Static Test Data — Preprint; Abstract only.