A Global-to-Local Diagnostic Scheme Identifies Residual-Defined Local Adsorption Regimes in Metal–Organic Framework Screening
High-throughput screening models for metal–organic frameworks (MOFs) achieve high global accuracy, yet this can conceal systematic failures in regions of framework space governed by distinct structure–property relationships. Here, we develop a residual-guided, global-to-local diagnostic scheme that couples extreme gradient boosting (XGBoost) residual analysis with sure independence screening and sparsifying operator (SISSO) symbolic regression to locate where accurate models fail and extract the local relationships. For light-hydrocarbon adsorption in 31 399 hydrophobic hypothetical MOFs (hMOFs), an XGBoost model reaches a global R2 of 0.9839 but retains a reproducible high-error tail enriched in specific descriptor-space regions. SISSO as an interpretable reference reveals a shift from a density/surface-area-dominated global expression to a local expression involving void fraction, largest cavity diameter and local chemistry (composition-level descriptors) in about 1% of the hMOFs, examined by accessibility and adsorption calculations. Methane and propane show analogous shifts with adsorbate-specific local expressions, establishing residual-guided diagnosis for materials screening.
Authors
- Xianyang Zhang (ORCID: https://orcid.org/0000-0001-6682-262X)
- Zhiwei Qiao (ORCID: https://orcid.org/0000-0002-1264-3762)
- Xue‐Qing Gong (ORCID: https://orcid.org/0000-0002-9566-7427)
- Yujuan Yang (ORCID: https://orcid.org/0000-0003-4635-2709)
- Xue-Ying Yuan
Institutions
- Shanghai Jiao Tong University (CN)
- Guangzhou University (CN)
Publication Details
- Journal
- The Journal of Physical Chemistry Letters
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1021/acs.jpclett.6c03526
- Primary Topic
- Machine Learning in Materials Science
- Type
- article
- Field-Weighted Citation Impact
- 0.00