Research papers
1. Low force error did not guarantee stable CdSe simulations
In a new, unreviewed preprint, five neural-network potentials were tested on one chloride-passivated CdSe nanocluster. NequIP sustained a 1 ns trajectory without extra training; MACE required 34 additional configurations, while lower-force-error Allegro remained unstable.1
Why it matters: This benchmark makes trajectory stability a separate model-selection test.1
Limits: The result concerns one 149-atom cluster and its tested conditions, not a general ranking of these architectures.1
2. An agent coordinates atomistic calculations with explicit execution rules
El Agente Potente, a new unreviewed preprint, separates language-model planning from numerical execution in structured simulation workflows. Demonstrations cover materials screening, molecular conformers and surface chemistry.2
Why it matters: Explicit execution graphs retain calculation provenance and handle failures while supporting automated campaigns.2
Limits: Successful workflow execution does not validate every prediction: the reported ion conductivities are screening indicators affected by short trajectories, finite cells and model limitations.2
3. Fast prescreening narrows semiconductor defect searches
A new unreviewed preprint predicts defect formation energies and optical transition energies in 4H-SiC without first relaxing each candidate structure. Models are evaluated using ten-fold cross-validation over double defects, with single defects added to training.3
Why it matters: The approach could help decide which defects merit expensive electronic-structure calculations.3
Limits: Errors are larger for interstitial defects; within-material validation does not establish transfer to other semiconductors.3
4. Thermodynamic maps reach peer-reviewed early publication
Newly journal-published expTM generates phase-behaviour samples from limited simulation data for a lattice gas and solid carbon dioxide.4 The method previously appeared as a March 2025 preprint.54
Why it matters: It extends generative sampling to pressure and chemical potential alongside temperature.4
Limits: These demonstrations do not establish universal extrapolation or experimental accuracy; lattice-gas deviations increase near critical points. The journal release is an accepted article in press, subject to further editing.4
5. A better ML prediction may still make a quantum calculation slower
A new unreviewed preprint tests learned starting states inside a density-functional-theory solver. Direct charge-density prediction delivered a reported 1.18-fold solver-loop speedup, but other tested predictions did not consistently beat the standard initial guess.6
Why it matters: It measures whether an ML output helps the calculation it is intended to accelerate.6
Limits: The benchmark covers non-magnetic crystals in one ABACUS workflow. Timing excludes inference, export and reference preparation, so it is not an end-to-end speedup.6
References
M. Usman; M. Suleymanova; Z. U. Abideen; M. Fernández-Pendás; I. Infante. Benchmarking Machine-Learning Interatomic Potentials for Dynamical Stability in Inorganic Semiconductor Nanocrystals: A CdSe Case Study. arXiv; 2026; Unreviewed preprint, arXiv v1. DOI: 10.48550/arXiv.2609.15299. Accessed 2026-09-15.
Source evidence and access
PDF pp. 9–12, Table 2 and Conclusions; abstract-page submission history: v1 14 September 2026 09:50:26 UTC
“NequIP completed a 1 ns trajectory without additional training data.”
Full PDF, abstract and submission history inspected; calculations not rerun.
Tsz Wai Ko; Jiaru Bai; Thomas Swanick; Yeonghun Kang; Changhyeok Choi; Angelina Qihong Jiang; Aiwei Yin; Varinia Bernales; Alán Aspuru-Guzik. El Agente Potente: High-Throughput Agentic Atomistic Simulations. arXiv; 2026; Unreviewed preprint, arXiv v1. DOI: 10.48550/arXiv.2609.14840. Accessed 2026-09-15.
Source evidence and access
HTML sections 2.3.1 (Superionic lithium-ion conductors) and 3 Discussion; submission history v1 13 September 2026 23:20:21 UTC
“These results should therefore be interpreted as screening-level indicators”
Full HTML and submission history inspected; software and simulations not rerun.
Paul Karlsson; Joel Davidsson; Rickard Armiento. Prescreening Point Defects in Semiconductors With Machine Learning. arXiv; 2026; Unreviewed preprint, arXiv v1. DOI: 10.48550/arXiv.2609.14846. Accessed 2026-09-15.
Source evidence and access
HTML abstract, section III.2 Performance Assessment, Table 3(a), concluding discussion; submission history v1 13 September 2026 23:41:06 UTC
“Ten-fold cross-validation is used”
Full HTML and submission history inspected; no independent model rerun.
Suemin Lee; Ruiyu Wang; Lukas Herron; Pratyush Tiwary. Predicting phase transitions across temperature, pressure, and chemical potential using exponentially tilted thermodynamic maps. Nature Communications; 2026; Peer-reviewed accepted Article in Press; not final Version of Record. DOI: 10.1038/s41467-026-77612-y. Accessed 2026-09-15.
Source evidence and access
Accepted-manuscript PDF https://www.nature.com/articles/s41467-026-77612-y_reference.pdf pp. 1–5, Results and Discussion; publisher early-publication notice and publication date
“minimal deviation from MC simulations, except near critical points”
Publisher metadata and full accepted-manuscript PDF inspected; supplementary benchmark and code not rerun.
Suemin Lee; Ruiyu Wang; Lukas Herron; Pratyush Tiwary. Exponentially Tilted Thermodynamic Maps (expTM): Predicting Phase Transitions Across Temperature, Pressure, and Chemical Potential. arXiv; 2025; Earlier preprint, arXiv v1; included for provenance, not a daily selection. DOI: 10.48550/arXiv.2503.15080. Accessed 2026-09-15.
Source evidence and access
Submission history
“[v1] Wed, 19 Mar 2025 10:24:57 UTC”
Abstract and history inspected solely to verify earlier disclosure.
Pin Chen; Jiang Li; Yutong Lu. Evaluating Predicted Densities, Hamiltonians, and Density Matrices as Periodic SCF Initializers. arXiv; 2026; Unreviewed preprint, arXiv v1. DOI: 10.48550/arXiv.2609.15151. Accessed 2026-09-15.
Source evidence and access
HTML sections 6.2 and 8 Limitations, Table 3; submission history v1 14 September 2026 07:25:22 UTC
“our reported speedups measure SCF-loop time and exclude model inference, state export, and residual-reference generation.”
Full HTML and history inspected; runs not independently reproduced.
Publication record
Published 2026-09-15.
Sources, selection and claims were checked in an independent AI editorial review, followed by the AI editor's approval. This is not academic peer review.
- 2026-09-15 · Version f73a17a7 · Viewing this version