in silico
Phase 1 : In Silico Validation
- Objectif
- To computationally test the core assumption of linear additivity and descriptor sufficiency using existing DFT datasets and surrogate models, before committing to expensive AET and synthesis.
- Coût estimé
- €0-2k (primarily compute time, potentially covered by existing allocations)
- Durée estimée
- 4-8 weeks
- Critères de réussite
- Cross-validated predictive power (Q²) of the linear additive model for ΔE_ads · Q² ≥ 0.50 · (5-fold cross-validation on the curated dataset; calculate Q² = 1 - (PRESS/SS_total) where PRESS is the prediction error sum of squares.)
- Evidence of significant non-additivity · F-test p-value > 0.05 for improvement with interaction terms · (Compare the linear additive model (restricted) to the model with ensemble-interaction terms (full) using an F-test on the residual sum of squares.)
- On continue si
- Q² ≥ 0.50 AND p-value > 0.05 (i.e., additive model performs reasonably well and adding complexity doesn’t significantly improve it).
- On arrête si
- Q² < 0.30 OR p-value < 0.01 (i.e., additive model fails or strong non-additivity dominates, falsifying core assumption).
- On réoriente si
- 0.30 ≤ Q² < 0.50 OR 0.01 ≤ p-value ≤ 0.05. Proceed to Phase 2 but expand descriptor set (e.g., include orbitalwise coordination number) or explicitly model nearest-neighbor interactions in the 3D-QSAR grid.
- Risques
- Insufficient or low-quality surrogate data in public DFT databases for relevant HE-NP surface ensembles.Probabilité : mediumParade : Generate a minimal set of 5-10 new DFT calculations for the most promising/compositionally diverse ensembles identified from literature to supplement the database.
- Computational cost of descriptor calculation for hundreds of ensembles is prohibitive.Probabilité : lowParade : Use machine-learned force fields (M3GNet) or graph neural network (GNN) surrogates (e.g., OrbNet) trained on a subset of DFT data to rapidly predict descriptors.