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

This week’s preprints explore cheaper ways to screen and simulate materials, from million-atom moiré models to generative structures and adaptive discovery campaigns. The reported results are computational; they should not be read as peer-reviewed findings or experimental validation.

CrystalJev uses calibrated decisions to triage hypothetical materials

This preprint presents CrystalJev, which uses a frozen interatomic potential to classify unrelaxed structures with calibrated probabilities and recommend when more expensive calculations could change a decision. In a registered prospective test using 700 new density-functional calculations, its forecasts overstated the stable fraction of unseen candidates by at most 2.1 percentage points. The authors report that a single model pass can approach relaxation-based decisions at roughly one-thirtieth the cost.

Why it matters: If the reported calibration holds on relevant candidate sets, this approach could reserve costly relaxations for cases where they are most informative, making large-scale screening more efficient.

Limitations: This is an arXiv preprint, and the reported prospective test is computational, not experimental. Calibration and decision quality may depend on the models and candidate populations studied; the abstract does not establish performance across all materials domains.

CrystalJev: thinking fast and slow with atomistic foundation models for materials discovery — Preprint; Abstract only.

This preprint combines a quantum spectral thermodynamics framework with uncertainty-guided active learning to explore 7.7 million high-entropy ceramic configurations at density-functional-theory fidelity. The authors report that disorder-related phonon effects can contribute to stabilization beyond ideal configurational entropy, and identify a roughly 12 at.% solute threshold associated with a change from strengthening to softening in their studied chemical space.

Why it matters: The work suggests a computational route to surveying composition spaces too large for exhaustive first-principles calculations, while connecting disorder-related vibrational effects to predicted stability and mechanical behavior.

Limitations: These are computational results reported in a preprint, not experimental confirmation of the predicted materials or threshold. Conclusions depend on the modeled chemical space, thermodynamic framework, and active-learning workflow.

Quantum spectral thermodynamics and active learning enable million-scale exploration of high-entropy ceramics — Preprint; Abstract only.

Diffusion model generates amorphous oxide structures and ensembles

This preprint describes an SE(3)-equivariant diffusion model trained on 1,781 DFT configurations to generate amorphous oxide structures. The authors report that it reproduces selected structural statistics and can generate models exceeding 300,000 atoms. First-principles relaxation found a generated amorphous Zr-Ta-O structure 36 meV per atom below the previously known minimum; the paper also reports generation beyond training conditions.

Why it matters: Treating an amorphous material as an ensemble of structures may help researchers sample configurations that conventional simulations miss and propose candidates for subsequent first-principles study.

Limitations: The work is a preprint. The lower-energy structure is a computational result requiring further validation; the abstract does not establish experimental synthesis or broad generalization to arbitrary compositions and conditions. The authors caution that denoising loss does not rank generative quality.

Equivariant generative diffusion learns and generalizes the structural ensemble of amorphous oxides — Preprint; Abstract only.

Distilled potential scales moiré simulations to one million atoms

This preprint introduces MoiréMLIP, fine-tuned on a DFT-labeled dataset, and a distilled smaller model intended to retain its reported accuracy while enabling single-GPU inference on systems of up to one million atoms. The authors report force errors of 6–8 meV/Å and transfer to smaller twist angles and unseen structures. In a MoTe₂ trilayer application, simulations reveal a predicted polarization network associated with moiré-of-moiré domain walls.

Why it matters: Scaling atomistic models to much larger moiré structures could let researchers investigate reconstruction and polarization patterns beyond the reach of direct first-principles modeling.

Limitations: This is a preprint and the reported findings are simulation results, not experimental observations. Accuracy and transferability are established only for the tested data and structures; million-atom inference does not itself establish predictive accuracy for every such system.

Distilling universal machine-learning potentials for moiré lattices across one million atoms — Preprint; Abstract only.

SciKGExtract turns materials literature into structured process data

This preprint presents a schema-guided workflow that combines language-model extraction, chemical normalization, and agent-based evaluation to build knowledge graphs from atomic-layer-deposition literature. On 176 papers about ZnO and IGZO, the authors report that agentic refinement raised the best ZnO extraction F1 from 0.591 to 0.805, while the best IGZO result was 0.344. Tests with a deeply nested experimental schema exposed difficulties in process segmentation and assigning numerical values.

Why it matters: More structured, reusable process information could help materials researchers compare experimental procedures and provide machine-actionable data for downstream analysis.

Limitations: The results are from a preprint and a focused set of ALD papers; performance varied substantially between ZnO and IGZO. The abstract reports extraction accuracy, not evidence that the resulting knowledge graph is complete or reliable across broader materials literature.

Agentic schema-guided extraction of materials process knowledge from scientific literature — Preprint; Abstract only.

Adaptive resource allocation for multi-objective alloy discovery

This preprint compares fixed and LLM-based adaptive batch-allocation strategies for multi-objective Bayesian optimization in two alloy-design settings. Under simulated resource disruptions, the authors report that adaptive strategies adjusted allocation to campaign statistics and constraints, outperforming a fixed mixed policy; mixed strategies also accumulated more mutual information than an exploitation-focused baseline at a smaller cost to hypervolume and optimization speed.

Why it matters: Materials campaigns face changing time and budget constraints. Adapting the balance between exploration and exploitation could help make limited synthesis and characterization resources more useful.

Limitations: The reported evaluation is a computational benchmark, not a demonstrated autonomous laboratory campaign or experimental materials discovery. Results are limited to the two alloy systems and disruption scenarios described in the preprint.

Agentic Resource Allocation for Batch Multi-Objective Bayesian Optimization in Autonomous Materials Discovery — Preprint; Abstract only.

Reinforcement learning raises a crystal-generation benchmark score, with caveats

This preprint applies group-relative policy optimization to the discrete composition channel of a crystal generator. The authors report that their reward increased the benchmark’s metastable, unique, and novel structure rate from 13.4% to 45.5%. They also identify reward exploitation: distinct packings of single-element structures count toward the metric despite not yielding new compounds.

Why it matters: The work illustrates how reward-based fine-tuning can steer crystal generators toward design objectives, while showing why benchmark definitions and safeguards matter to interpreting gains.

Limitations: This is a preprint, and the reported gain is on a computational benchmark rather than experimental synthesis or validation. The authors identify metric inflation and reward hacking; stability also depends on the reference hull used.

Reinforcement Learning on the Discrete Composition Channel of a Crystal Generator: Validated Gains and Reward Hacking — Preprint; Abstract only.

Compressed magnetic potentials reduce model parameters

This preprint uses matrix and tensor factorizations to reduce the parameter count of magnetic Moment Tensor Potentials by about 1.5–3 times. The authors report no loss of predictive performance on the validation set and numerical consistency with the original models in tests of magnetic, structural, and vibrational properties and molecular dynamics for Fe-Al and CrN.

Why it matters: Smaller magnetic interatomic potentials could reduce storage and computational overhead in simulations while retaining the tested model’s predictive behavior.

Limitations: The claims are from a preprint and validation is limited to the reported properties and Fe-Al and CrN systems. Agreement with DFT and experimental data is reported for these tests, but does not establish general accuracy for other materials or conditions.

Compressed magnetic Moment Tensor Potentials via low-rank matrix and tensor factorizations — Preprint; Abstract only.