Discovery of descriptors for stable monolayer oxide coatings through machine learning
Summary¶
Screening monolayer oxide coatings on oxide supports with exhaustive density functional theory is expensive, yet simple linear models may suffice if informative descriptors can be identified. Jonayat, van Duin, and Janik train supervised LASSO regression models that map DFT-computed monolayer oxide stabilities for coating/support pairs onto physically interpretable features drawn from tabulated ionic, electronic, and surface properties. They separately treat stoichiometric and nonstoichiometric monolayer oxides to reflect different defect chemistries important in catalysis. The motivation ties to supported catalyst design where monolayer oxides can stabilize interfaces but where brute-force DFT across composition grids is prohibitive. The article’s machine-learning layer is deliberately sparse and interpretable rather than a black-box neural network, aiding experimental guidance.
Methods¶
Training labels come from DFT (PBE family with Hubbard corrections where noted in companion work) evaluations of monolayer adhesion or stability relative to bulk references across transition-metal oxide coatings on oxide supports; exact dataset scope appears in the article and Supporting Information. Input features include substrate surface energies, ionic/atomic radii, ionization potentials, and parent oxide stability metrics as enumerated in the feature table. LASSO regression selects sparse descriptor sets with cross-validation protocols defined in the SI. Feature scaling and colinearity checks matter because tabulated radii and energies can correlate. Reproducibility requires the same DFT settings, dataset splits, LASSO hyperparameters, and random seeds documented in DOI 10.1021/acsaem.8b01261.
Static QM / DFT details. Functional / Hubbard: PBE-family DFT with DFT+U on localized d states where noted for transition-metal oxides (exact U values and XC treatment in Methods/SI). Dispersion: N/A — explicit DFT-D3 / pairwise vdW correction not highlighted in this wiki summary; confirm whether the publication applies a vdW correction in pdf_path. Basis / potentials: plane-wave expansion with PAW pseudopotentials as in companion oxide-screening work from the same series. k sampling: Monkhorst–Pack k-point meshes sized per slab / bulk calculation in the article. Structures: relaxed oxide surfaces and monolayer configurations for coating/support pairs; N/A — NEB transition-state pathways as the primary object—the study targets adhesion/stability energies rather than reaction barriers. Properties: formation/adhesion-type energies feeding LASSO labels plus tabulated electronic/ionic descriptors.
Findings¶
Outcomes. For stoichiometric mixed oxides, LASSO selects substrate surface energy, orbital radii, and ionization energies as dominant descriptors of monolayer stability. For nonstoichiometric films, parent oxide stability and oxidation-state mismatch between coating and support dominate.
Comparisons. Sparse linear models benchmark against brute-force DFT screening costs while staying interpretable versus black-box ML.
Sensitivity / practice levers. Oxidation state metadata and explicit stoichiometry tags materially change which descriptors survive LASSO—mixing regimes collapses interpretability even if error improves marginally.
Limitations and PDF grounding. DFT+U errors on correlated oxides and narrow training chemistry imply extrapolation risk; coefficients are not portable if the functional/U recipe changes. Numeric dataset sizes, cross-validation splits, and bootstrap uncertainty should be taken from the ACS Appl. Energy Mater. PDF/SI (pdf_path).
Limitations¶
DFT errors on strongly correlated oxides and limited coverage of synthesis environments mean extrapolation risk outside the training chemistry.
Relevance to group¶
van Duin-group coauthorship on descriptor learning for oxide interfaces—complements ReaxFF applications by targeting DFT screening.
Citations and evidence anchors¶
- DOI:
10.1021/acsaem.8b01261.