Skip to content
DraconiticResearch
Search Sources API
197,768 studies31 sourcesUpdated 9 min ago
PreprintCC BYChemRxiv

Flexible force field-based screening and machine learning predictions for CO2 sorption and diffusion in MOFs

  • Clare B. Martin1,2
  • Samir Budhathoki3,4,5
  • Janice A. Steckel4,6,7
7 affiliations
  1. 1Oak Ridge Institute for Science and Education, 100 ORAU Way, Oak Ridge, TN, 37830
  2. 2Department of Chemical Engineering, University of Michigan, 2800 Plymouth Road, Ann Arbor, MI, 48109
  3. 3Leidos Research Support Team, P.O. Box 10940, Pittsburgh, PA, 15236-0940
  4. 4US DOE, National Energy Technology Laboratory, 626 Cochran Mill Rd, Pittsburgh, PA, 15236
  5. 5Braskem, 550 Technology Dr, Pittsburgh, PA, 15219
  6. 6Department of Chemistry
  7. 7University of Pittsburgh
Open at sourcePDFDOICite

Abstract

Machine learning offers an efficient route for screening metal–organic frameworks (MOFs), but most existing models rely on rigid-framework simulations that neglect framework flexibility. In this work, DFT-derived flexible force fields were used to calculate CO2 sorption and self-diffusion for 5,418 MOFs, providing the training data for machine-learning models capable of rapidly predicting these properties. We investigated different ML algorithms, featurization methods, and the role of physical descriptors in predicting CO2 sorption and self-diffusion in MOFs. For CO2 sorption, a categorical boosting classifier coupled with Stoichiometric-45 featurization method used with scalar descriptors provided predictive accuracy of 76.1%. For CO2 self-diffusion, we tuned a categorical boosting regressor on revised autocorrelation functions coupled with scalar descriptors and achieved moderate predictive accuracy (R2 = 0.69). These results demonstrate that machine learning trained on properties generated from flexible force field simulations can substantially accelerate computational screening while incorporating key effects of framework flexibility, providing an efficient strategy for discovering MOFs for low-pressure CO2 capture. Summary Computational screening of 5,418 metal–organic frameworks using flexible DFT-based force fields captures framework motion effects on CO2 sorption and self-diffusion. Machine learning models were evaluated across descriptors and algorithms. A CatBoost classifier achieved 76.1% sorption accuracy, while an autocorrelation-based regressor predicted self-diffusion (R2 = 0.69), demonstrating the value of flexibility-aware modeling.

Related

  1. Reactive Machine-Learned Potentials Improve Predictions of Water Diffusion Behaviour in Defective Industrial Zeolite HY
    Indranil Saha, Alexander James Porter, Andreas Erlebach +4
    ChemRxivPreprintCC BYFull textInorganic & organometallic chemistryChemRxiv
  2. Thermodynamically Constrained Deep Learning for Hydrogen Sorption Prediction: Overcoming Classical Isotherm Generalization Failure through Physics-Informed Cross-Lithology Transferability
    Mohammad Nooraiepour, Mohammad Masoudi, Zezhang Song +1
    ACS OmegaPeer-reviewedCC BY 4.0Full textMaterials & polymersCrossref
  3. QALPA: Property-guided diffusion modeling for efficient exploration of chemical spaces of flexible molecules
    Michael Hanna, Julian Cremer, Zekiye Erarslan +1
    arXivPreprintCC BY-NC-ND 4.0Full textPhysical chemistryarXiv
  4. Embedded Morgan Fingerprints for more efficient molecular property predictions with machine learning
    Emilio Nuñez-Andrade, Isaac Vidal-Daza, Rafael Gomez-Bombarelli +2
    ChemRxivPreprintLicence not statedGeneral chemistryChemRxiv
  5. Flexible On-the-Fly Active Learning for Molecular Infrared Spectra.
    Giulio Benedini, Matti Hellström, Lucas Visscher
    Journal of chemical theory and computationPeer-reviewedCC BYFull textMaterials & polymersEurope PMC

Cite

BibTeXRIS
BibTeX
@misc{martin2026flexible,
  title = {{Flexible force field-based screening and machine learning predictions for CO2 sorption and diffusion in MOFs}},
  author = {Clare B. Martin and Samir Budhathoki and Janice A. Steckel},
  howpublished = {ChemRxiv},
  year = {2026},
  month = oct,
  doi = {10.26434/chemrxiv.15010813/v1},
  url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010813/v1},
  note = {CC BY},
}
RIS
TY  - UNPB
TI  - Flexible force field-based screening and machine learning predictions for CO2 sorption and diffusion in MOFs
AU  - Clare B. Martin
AU  - Samir Budhathoki
AU  - Janice A. Steckel
PY  - 2026
DA  - 2026/10/21
PB  - ChemRxiv
DO  - 10.26434/chemrxiv.15010813/v1
UR  - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010813/v1
AB  - Machine learning offers an efficient route for screening metal–organic frameworks (MOFs), but most existing models rely on rigid-framework simulations that neglect framework flexibility. In this work, DFT-derived flexible force fields were used to calculate CO2 sorption and self-diffusion for 5,418 MOFs, providing the training data for machine-learning models capable of rapidly predicting these properties. We investigated different ML algorithms, featurization methods, and the role of physical descriptors in predicting CO2 sorption and self-diffusion in MOFs. For CO2 sorption, a categorical boosting classifier coupled with Stoichiometric-45 featurization method used with scalar descriptors provided predictive accuracy of 76.1%. For CO2 self-diffusion, we tuned a categorical boosting regressor on revised autocorrelation functions coupled with scalar descriptors and achieved moderate predictive accuracy (R2 = 0.69). These results demonstrate that machine learning trained on properties generated from flexible force field simulations can substantially accelerate computational screening while incorporating key effects of framework flexibility, providing an efficient strategy for discovering MOFs for low-pressure CO2 capture. Summary Computational screening of 5,418 metal–organic frameworks using flexible DFT-based force fields captures framework motion effects on CO2 sorption and self-diffusion. Machine learning models were evaluated across descriptors and algorithms. A CatBoost classifier achieved 76.1% sorption accuracy, while an autocorrelation-based regressor predicted self-diffusion (R2 = 0.69), demonstrating the value of flexibility-aware modeling.
ER  -