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

This week’s eight selections connect machine learning to battery interfaces, crystal screening, microstructure generation and materials characterization. The featured study pairs atomistic predictions with coating experiments; two other studies scrutinize crystal novelty and interatomic-potential benchmarks. All selected sources are preprints available here only as abstracts, not peer-reviewed findings.

Researchers combined a fine-tuned PET-MAD interatomic potential with experiments to study molecular coatings on the sulfide electrolyte Li6PS5Cl. Simulations predict that long-chain lithiated carboxylates form hydrophobic layers, with protection emerging near 0.8 nm thickness. Experiments found stronger suppression of hydrogen sulfide evolution with longer chains, while impedance measurements showed lithiated coatings retained 40–55% of initial ionic conductivity after powder processing.

Why it matters: The combination connects molecular adsorption and coating morphology to measurable protection and conductivity. Intermediate chain lengths emerge as a practical design direction because thicker barriers can exclude water while also obstructing lithium-ion transport.

Limitations: This abstract-only preprint does not establish long-term cycling performance or full-cell durability. The approximately 0.8 nm protection threshold is a simulation result; the experimental trends support the proposed mechanism without independently verifying every predicted structural detail.

Combining Simulations and Spectroscopy to Design Protective Coatings on Sulfide Solid Electrolytes — Preprint; Abstract only.

A consolidated crystal database challenges claims of generative novelty

A preprint reports structure-based integration of 55 databases into a collection of 14.6 million crystals and 131.8 million property values. Only 4.5% of crystal-property pairs have recorded values. At an equal verification budget, a stiffness-conditioned generator produced 100 stiff hits versus 55 from screening, but every generated hit contained osmium and 82 repeated recorded compositions.

Why it matters: Incomplete property coverage can make known crystals look exhausted as a discovery resource. Structural deduplication and source-traced records could improve both forward screening and the reference used to assess whether generated candidates are genuinely new.

Limitations: Only the preprint abstract is supplied. Repeated compositions are not necessarily duplicate structures, and the generator comparison concerns one stiffness task. Predicted stiff candidates are not experimental discoveries; the cited 428 GPa measurement for ReN2 comes from a retrospective audit of existing synthesis evidence.

Probing the limits of forward screening and inverse design of crystal materials — Preprint; Abstract only.

Alloy-potential rankings depend on the quality of their DFT controls

Researchers benchmarked 14 universal machine-learned interatomic potentials against 1.45 million ordered intermetallic structures from AFLOW, which the authors report none of the models trained on. Comparisons among AFLOW, Alexandria and OQMD revealed substantial reference disagreement. After re-referencing, the best model agreed with AFLOW about as closely as Alexandria did, while the next four were 2–7 meV per atom above that disagreement floor.

Why it matters: Benchmark error can mix model shortcomings with inconsistent reference calculations. Explicit database controls and elemental-reference checks could prevent misleading rankings and help identify which apparent failures merit new calculations rather than model retraining.

Limitations: This abstract-only preprint primarily addresses ordered intermetallic energy and stability comparisons. Agreement within cross-database variation does not establish experimental accuracy or reliable behavior for defects, finite-temperature dynamics and other configurations outside the benchmark.

How good are universal MLIPs for alloys? A benchmark is only as trustworthy as its controls — Preprint; Abstract only.

ML-assisted screening maps ordering and distortion in sodium cathodes

Using a library of 41 sodium/vacancy orderings, researchers screened more than 120,000 layered-oxide structures with machine-learning interatomic potentials before performing 1,025 first-principles calculations. The resulting phase diagrams explain reported stacking sequences and voltages. Their analysis favors combining high-voltage redox chemistry with low Jahn–Teller strain and identifies Na_xMn0.5Ni0.5O2 as a practical composition within the study.

Why it matters: The workflow connects composition, charge transfer and cooperative distortion rather than relying on a single structural descriptor. It also reports release of an ordering-resolved dataset that could support more accurate potentials for related sodium-ion electrode searches.

Limitations: The supplied evidence is a preprint abstract describing computational results, not new battery testing. The search applies a fixed ordering library to selected transition metals, and key stacking competitions involve only a few meV per atom, making small energy differences consequential.

Decoding Composition-Structure-Performance Relations in Sodium Layered Oxide Positive Electrode Materials — Preprint; Abstract only.

Symmetry-guided fine-tuning reduces the cost of thermal-conductivity calculations

A low-rank adaptation strategy fine-tunes MatterSim 1M on a small selection of compounds sharing a crystal symmetry, then applies the potential to other compounds with that symmetry. Across half-Heuslers, space-group-225 materials and spinels, the authors report thermal-conductivity relative errors 2–3 times smaller than with the untuned model. For 100 compounds, the reported computational time was about 5–7 times lower than individual compound fine-tuning.

Why it matters: Sharing adaptation across a structural family could make thermal-transport screening more affordable without requiring a separate fine-tuning workflow for every material. That would be useful when comparing thermoelectrics or heat-management materials within related crystal classes.

Limitations: This abstract-only preprint compares conductivity calculations with references obtained using moment tensor potentials and polyMLPs, not direct experimental measurements. The reported accuracy and speed gains apply to the tested families and do not establish transfer across arbitrary symmetries or chemistries.

Symmetry-Based Parameter Efficient Low-Rank Fine-Tuning for High-Throughput Calculation of Lattice Thermal Conductivity — Preprint; Abstract only.

Diffusion models generate microstructure inputs for additive-manufacturing digital twins

Researchers report denoising diffusion models that synthesize orientation-imaging micrographs containing morphological and crystallographic features of metal additive-manufacturing microstructures. They also introduce classifier-free guidance for property-to-structure and process-to-structure inverse mappings, with an example aimed at qualification-related uncertainty quantification.

Why it matters: Realistic synthetic microstructures could expand the inputs available to crystal-plasticity simulations when measured datasets are scarce. Conditional generation could help explore structural variability associated with a target property or processing condition.

Limitations: The preprint abstract reports quantitative fidelity assessment but provides no numerical metrics. Realistic generated images do not by themselves establish accurate mechanical properties, physically achievable processing routes or qualification-ready uncertainty estimates.

Microstructural Generative Inverse Models for Additive Manufacturing Digital Twins — Preprint; Abstract only.

Composable neural inference combines X-ray and neutron reflectometry

A probabilistic framework trains posterior estimators separately for individual scattering measurements, then combines them at inference time using a compositional sampler. Reflectometry demonstrations cover X-ray, time-of-flight neutron and polarized neutron observations, including co-refinement across neutron contrasts and across X-ray and neutron measurements.

Why it matters: Combining complementary probes can reduce structural ambiguity in thin films and interfaces. Reusing individual estimators without dedicated joint retraining could make uncertainty-aware analysis more modular as measurement combinations change.

Limitations: This abstract-only preprint demonstrates reflectometry applications, not general resolution of scattering inverse problems. The supplied evidence does not quantify uncertainty calibration, runtime or robustness to inconsistent measurements; broader scattering applications remain prospective.

Compositional neural networks for multimodal X-ray and neutron scattering inversion — Preprint; Abstract only.

ML potentials disentangle thermal expansion from heat transport in anharmonic crystals

Actively learned moment tensor potentials enabled molecular-dynamics studies of Rb2SO4, Rb2Ca2(SO4)3 and Ba3Lu(BO3)3. Simulations reproduced measured thermal expansion and supported previously proposed structural mechanisms. Calculated thermal conductivities differed by more than an order of magnitude between compounds with comparable anharmonicity measures, while predicted temperature-driven band-gap renormalization reached 2.7 eV in one compound.

Why it matters: The results illustrate how ML-enabled dynamics can connect atomic rearrangements to thermal and electronic properties in systems poorly described by harmonic models. They also caution against using a single anharmonicity metric as a shortcut for predicting heat transport.

Limitations: The evidence is limited to a preprint abstract. Thermal-expansion agreement provides an experimental check, but the conductivity and band-gap changes are computational predictions in the supplied record. Their accuracy depends on the underlying density-functional data and learned potentials.

Thermal expansion mechanisms, heat transport, and band-gap renormalization in strongly anharmonic Rb$_2$SO$_4$, Rb$_2$Ca$_2$(SO$_4$)$_3$, and Ba$_3$Lu(BO$_3$)$_3$ from machine-learned interatomic potentials — Preprint; Abstract only.