Entirely AI-generated, including the editorial review. Not independently human fact-checked.

Published September 20, 2026. This September 14 edition covers August 31–September 14; the selected preprints were posted September 13–14. This is a bounded source collection, not a comprehensive survey.

This edition follows programmable laboratory measurements, reproducible synthesis, and tougher tests for learned materials models. One toolkit exposes RHEED measurements and analysis to AI agents; another platform uses AI-generated control code for automated synthesis. Computational benchmarks examine whether prediction accuracy translates into stable dynamics or faster electronic-structure calculations. All summaries below rely on preprint abstracts. Peer-review status has not been independently verified, and reported results remain the authors’ claims.

Researchers built an open-source hardware/software toolkit that automates reflection high-energy electron diffraction (RHEED) alignment, rocking-curve acquisition, and quantitative structural analysis during pulsed laser deposition. A Model Context Protocol server lets AI agents interpret natural-language requests to select and run analyses, with an extensible adapter interface for community-developed RHEED simulation and AI-analysis methods.

Why it matters: RHEED is one of the richest real-time probes of surface evolution during epitaxial growth, but manual operation has long been a bottleneck to closing the loop on autonomous synthesis. Making RHEED control and interpretation machine-addressable is a concrete step toward agent-directed thin-film deposition, with potential applications in automated materials research.

Limitations: This is a preprint demonstrating toolkit capability and a single agent-driven use case on one pulsed laser deposition system; the abstract does not establish closed-loop optimization of film quality or generalization across deposition techniques and materials classes.

An Open-Source Hardware and Software Toolkit to Enable Agentic RHEED-Guided Thin-Film Synthesis — Preprint; Abstract only.

AI-generated control code supports reproducible MOF synthesis

A team combined commercial robot arms, electric pipettes, web cameras, and an electronic balance, controlled by code generated from an AI agent based on a large language model, to automate and precisely record a two-solution synthesis of the metal-organic framework ZIF-8. The platform revealed that particle size distribution depends strongly on pipette dispensing speed, a parameter rarely controlled or logged in manual synthesis, and reproduced this dependence across repeated runs.

Why it matters: Rather than chasing full autonomy, this work targets reliable, quantitative recording of process parameters that are typically invisible in manual lab work, addressing a key bottleneck for data-driven materials development: generating high-quality, reproducible synthesis data at scale using commercially available components.

Limitations: Demonstrated on a single well-characterized system (ZIF-8 precipitation); broader applicability to more complex, multi-step syntheses and to materials with less forgiving reaction windows is not established by the abstract.

Quantitative control and recording of materials-synthesis processes using an automated experimentation platform — Preprint; Abstract only.

Benchmark shows validation accuracy doesn’t predict MD stability for learned potentials

Comparing five graph neural network interatomic potentials (SchNet, PaiNN, NequIP, Allegro, MACE) on a chloride-passivated CdSe nanocluster, Allegro achieved the lowest force errors but this did not translate to stable molecular dynamics. NequIP ran stably for a full nanosecond without extra data, while MACE needed active-learning augmentation to stabilize, and PaiNN and Allegro remained unstable after adding 100 and 95 configurations, respectively. SchNet also failed within its tested augmentation budget.

Why it matters: This directly challenges the common practice of selecting machine-learning interatomic potentials purely by held-out force/energy error. For finite, surface-dominated nanostructures relevant to quantum dots and catalysts, long-timescale dynamical testing and uncertainty-guided active learning appear essential for trustworthy deployment.

Limitations: Findings are based on a single 149-atom CdSe nanocluster system and one thermodynamic condition (300 K, 1 ns); generalization to other nanocrystal chemistries, larger systems, or longer timescales is not established.

Benchmarking Machine-Learning Interatomic Potentials for Dynamical Stability in Inorganic Semiconductor Nanocrystals: A CdSe Case Study — Preprint; Abstract only.

Active-learning CALPHAD search identifies Ru-stabilized refractory BCC/B2 alloys

Combining CALPHAD thermodynamic calculations with random-forest-guided active learning, researchers screened a ten-element Nb-based, Ru-bearing composition space at 1 at.% resolution across 500 evaluated alloys. The approach reproduced known trends (RuHf and RuZr staying stable to the solidus) and identified 100 candidate alloys meeting criteria for high-temperature BCC+B2 dual-phase stability, including 19 Ru-lean compositions.

Why it matters: Refractory alloys with ductile BCC matrices strengthened by ordered B2 precipitates are a high-temperature analogue to Ni-superalloy gamma/gamma-prime microstructures. Coupling active learning with CALPHAD offers a way to prioritize a large compositional space, offering practical design rules for next-generation high-temperature structural alloys.

Limitations: The screening results are computational predictions of phase stability. The abstract does not provide new experimental or mechanical validation of the candidate set, which includes two independently reported alloys.

Machine Learning Guided CALPHAD Design of Ru-Stabilized BCC B2 Refractory Alloys — Preprint; Abstract only.

Generative-predictive framework reports ~90% success on computational singlet-fission criteria

A framework integrating a molecular structure generator, a properties predictor, and multi-criteria validation was used to search for singlet-fission (SF) chemical space, generating about 100 million candidate structures. Time-dependent DFT validation of a random 1% subset confirmed a 90.8% success rate for meeting target SF energetics, yielding a database of 283,559 candidates and an interpretable structural fragment associated with favorable SF behavior.

Why it matters: Singlet fission could help photovoltaics exceed the Shockley-Queisser efficiency limit, but SF-capable molecules are extremely rare in chemical space, making conventional virtual screening inefficient. Coupling generative design with predictive filtering is proposed as a way to concentrate screening on this narrow target property.

Limitations: Success rate is based on TDDFT validation of only a 1% random subsample of generated candidates, and does not demonstrate experimental singlet-fission performance. The abstract provides computational screening evidence, not experimental confirmation of synthesis or device behavior.

Navigating Sparse Singlet Fission Chemical Space: An Intelligent Generative-Predictive Paradigm — Preprint; Abstract only.

AIDEN neural network reconstructs real-space charge density with zero-shot transfer

AIDEN (Atomic-Interaction Density Equivariant Network) separates element-dependent one-center density from environment-induced redistribution, using a continuous low-rank Gaussian decoder to reconstruct charge density at arbitrary spatial coordinates. It reports state-of-the-art accuracy on periodic crystal benchmarks, competitive performance on molecules, zero-shot transfer to out-of-distribution structures, and substantially faster inference than baseline models or full self-consistent-field DFT calculations.

Why it matters: Fast, accurate surrogates for the Kohn-Sham charge density could meaningfully cut the cost of electronic-structure-driven materials screening, a persistent computational bottleneck for high-throughput computational materials design.

Limitations: Speed and accuracy gains are reported on benchmark datasets; the abstract does not establish downstream property accuracy or performance in a discovery pipeline.

Neural-Network Solutions to Real-Space Charge Density and Generalization — Preprint; Abstract only.

Large SCF benchmark shows learned initializers don’t always speed up DFT convergence

Using a new 43,851-crystal dataset (ρHD-43K) with aligned charge-density, Hamiltonian, and density-matrix labels, researchers tested whether learned predictions actually accelerate the self-consistent-field (SCF) loop of DFT, rather than just minimizing offline prediction error. Direct charge-density prediction accelerated about 91% of test crystals with a 1.18x SCF-loop speedup, but tested residual-density and Hamiltonian/matrix initializers did not consistently beat the standard atomic-density-superposition baseline.

Why it matters: This is a useful reality check for the growing field of ML-accelerated DFT: it shows that a learned model’s compatibility with the nonlinear SCF solver trajectory, not just its target-space accuracy, determines real-world usefulness, which should reshape how such models are evaluated going forward.

Limitations: Benchmark is restricted to non-magnetic crystals on a frozen test set; extrapolation beyond the reported test conditions is not justified by the abstract, and the modest observed speedups (median 3 fewer SCF iterations) may not justify added model complexity in all workflows.

Evaluating Predicted Densities, Hamiltonians, and Density Matrices as Periodic SCF Initializers — Preprint; Abstract only.

Physics-guided ML models prescreen point defects in SiC for quantum applications

Ridge, kernel ridge, and multilayer perceptron models were trained to predict defect formation energies and zero-phonon lines for point defects in 4H-SiC without requiring costly structural relaxation. For vacancies and substitutions, mean absolute errors reached 0.437 eV (formation energy) and 0.202 eV (zero-phonon line), described as useful even beyond prescreening in some cases; interstitial defects showed larger errors (1.101 eV and 0.230 eV) still deemed useful for prescreening.

Why it matters: High-throughput DFT defect screening for quantum and power-electronics applications is computationally expensive; models that skip the relaxation step could substantially widen the search space explored before committing to costly first-principles calculations.

Limitations: Error levels for interstitial defects are explicitly described as insufficient for detailed characterization, and the abstract reports validation on 4H-SiC; it does not establish transferability to other semiconductor hosts or defect chemistries.

Prescreening Point Defects in Semiconductors With Machine Learning — Preprint; Abstract only.