EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Target-driven inverse design of battery cathodes via constraint-prioritised cross-entropy search

Ali Badakhshan, Naeem Adibi, Andrew Gallant, Mehdi Keshavarz Hedayati

Machine Learning: Science and Technology · 01.08.2026 · 7 (4) · 045066

DOI: 10.1088/2632-2153/ae97d7 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Machine Learning: Science and Technology, online erschienen am 24.08.2026. Verfasst von Ali Badakhshan und 3 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: WeLOOP, EPSRC Centre for Doctoral Training in Renewable Energy Northeast Universities.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Batteries are central to transport electrification, grid-scale storage, and low-carbon energy systems, yet cathode discovery remains constrained by the size and sparsity of composition space. Inverse design can map desired electrochemical targets to candidate materials, but cathodes must satisfy coupled constraints involving charge balance, redox accessibility, capacity, stability, volume change, and electronic properties. Here, we introduce a physics-aware, chemically type-conditioned, voltage-stratified cross-entropy method (CEM) for inverse cathode design, combining structured chemical blueprints with constraint-prioritised lexicographic elite selection rather than a single manually weighted scalar objective. The framework searches oxide–fluoride, phosphate, and sulphate cathode spaces. Candidate blueprints define mobile-ion identity, redox-active elements, spectator cations, and stoichiometric coefficients on a discrete grid. A deterministic feasibility solver converts blueprints into charge-balanced formulas, while a redox-capacity function limits charge storage by mobile-ion charge transfer and available redox capacity. Surrogate models trained on density-functional-theory data estimate voltage, redox-limited capacity, formation-energy-related stability, relative volume change, band gap, and predictive uncertainty. The structured blueprint space contains approximately 4.28 × 10 10 possible sampled compositions before charge-balance filtering, of which approximately 5.5 × 10 9 ( ∼ 13 % ) satisfy the formal charge-balance and redox-feasibility rules. Uniform random search confirmed that jointly feasible candidates are rare, especially in high-voltage buckets. Across the 2.5–4.5 V range, chemically type-conditioned CEM identified 4315 ± 42 unique all-satisfying candidates across three independent seeds and achieved sample-efficiency gains of 38.6–292 × relative to random search, with a mean gain of 174.5 × , while retaining a wall-clock advantage after CEM overhead. The final archive contained oxide–fluoride and phosphate subsets, with no sulphate candidates under the current thresholds, and showed enrichment in selected mobile-ion/framework combinations and vanadium-based redox chemistry. The study is limited by the discretised blueprint space, surrogate-model domain coverage, and indirect computational assessment of synthesis feasibility.

Förderung

  • WeLOOP
  • EPSRC Centre for Doctoral Training in Renewable Energy Northeast Universities

Zitieren

APA

Badakhshan, A., Adibi, N., Gallant, A., & Hedayati, M. K. (2026). Target-driven inverse design of battery cathodes via constraint-prioritised cross-entropy search. Machine Learning: Science and Technology, 7(4), 045066. https://doi.org/10.1088/2632-2153/ae97d7

Vancouver

Badakhshan A, Adibi N, Gallant A, Hedayati MK. Target-driven inverse design of battery cathodes via constraint-prioritised cross-entropy search. Machine Learning: Science and Technology. 2026;7(4):045066. doi:10.1088/2632-2153/ae97d7

BibTeX
@article{Badakhshan2026,
  author = {Badakhshan, Ali and Adibi, Naeem and Gallant, Andrew and Hedayati, Mehdi Keshavarz},
  title = {Target-driven inverse design of battery cathodes via constraint-prioritised cross-entropy search},
  journal = {Machine Learning: Science and Technology},
  year = {2026},
  volume = {7},
  number = {4},
  pages = {045066},
  doi = {10.1088/2632-2153/ae97d7}
}
Kostenlos starten →