PaperlayerView exact revision on arXiv

Source-linked AI summary

Using Machine Learning and Data Mining to Leverage Community Knowledge for the Engineering of Stable Metal-Organic Frameworks

Aditya Nandy, Chenru Duan, Heather J. Kulik

arXiv:2106.13327v1cond-mat.mtrl-scics.LGphysics.chem-ph

TL;DR

A limited understanding of how to improve MOF stability constrains its practical use. The paper mines stability reports and trains machine-learning models, revealing limits of existing heuristics and predicting stability orders of magnitude faster than conventional physics-based modeling.

  • Problem

    Limited understanding of how to improve MOF stability constrains its practical use.

    Source: Conclusions · p. 21

  • Method

    The paper extracts published MOF stability reports and trains machine-learning models to predict stability properties.

    Source: Conclusions · p. 21

  • Results

    The analysis reveals limits of heuristics derived from smaller MOF sets, while machine-learning models predict both stability quantities orders of magnitude faster than conventional physics-based modeling.

    Source: Conclusions · p. 21

  • Takeaways & Limitations

    Interpreted model features point to strategies that could increase thermal stability for catalytic applications.

    Source: Conclusions · p. 22

    The main constraint is the amount of extractable data, rather than the representations or machine-learning models themselves.

    Source: Conclusions · p. 22

Abstract

from arXiv · show

Although the tailored metal active sites and porous architectures of MOFs hold great promise for engineering challenges ranging from gas separations to catalysis, a lack of understanding of how to improve their stability limits their use in practice. To overcome this limitation, we extract thousands of published reports of the key aspects of MOF stability necessary for their practical application: the ability to withstand high temperatures without degrading and the capacity to be activated by removal of solvent molecules. From nearly 4,000 manuscripts, we use natural language processing and automated image analysis to obtain over 2,000 solvent-removal stability measures and 3,000 thermal degradation temperatures. We analyze the relationships between stability properties and the chemical and geometric structures in this set to identify limits of prior heuristics derived from smaller sets of MOFs. By training predictive machine learning (ML, i.e., Gaussian process and artificial neural network) models to encode the structure-property relationships with graph- and pore-structure-based representations, we are able to make predictions of stability orders of magnitude faster than conventional physics-based modeling or experiment. Interpretation of important features in ML models provides insights that we use to identify strategies to engineer increased stability into typically unstable 3d-containing MOFs that are frequently targeted for catalytic applications. We expect our approach to accelerate the time to discovery of stable, practical MOF materials for a wide range of applications.

Dataset Curation

The curation pipeline assembled MOF structures and literature reports, then extracted solvent-removal labels and thermal decomposition temperatures using text analysis and TGA traces. The resulting datasets contain 2,179 solvent-removal structure–label pairs and 3,132 thermal decomposition temperatures.

Dataset Curation

  • Curation began with 10,143 all-solvent-removed CoRE MOF structures; revised autocorrelation and geometric descriptors could be computed for 9,597.The initial structures came from the 2019 CoRE MOF database, version 1.1.262.Source: Dataset Curation · p. 4
  • The literature pipeline linked 8,809 CSD refcodes to 5,152 DOIs and downloaded 3,809 manuscripts, yielding 7,004 featurizable structures for text mining.Source: Dataset Curation · p. 4
  • Regular expressions flagged sentences pairing solvent or activation terms with structural-integrity language, and dependency parsing helped distinguish affirmative from negative stability statements.Pattern matching alone could not resolve complex phrases, negations, or varied wording.Source: Dataset Curation · pp. 5–6
  • Solvent-removal labels were assigned only when all related sentences in a manuscript agreed; conflicting sentiment prevented automatic structure-specific labeling.Each labeled sentence received a binary stable or unstable flag.Source: Dataset Curation · p. 6
  • Thermal decomposition temperatures were measured from TGA traces rather than reported sentence values, using the intersection of tangent lines at the start and end of decomposition.This consistent trace-based protocol addressed ambiguity in reported temperatures, including guest removal and onset versus complete collapse.Source: Dataset Curation · pp. 6–7
  • The solvent-removal stability dataset contains 2,179 structure–label pairs; for this set, 1,209 manuscripts corresponded to 2,290 structures.In the CoRE MOF database, 2,649 manuscripts contained keywords associated with solvent-removal.Source: Dataset Curation · p. 7
  • The thermal stability dataset contains 3,132 decomposition temperatures extracted from TGA traces in 1,886 manuscripts.The 1,886 manuscripts were drawn from 2,366 manuscripts containing TGA keywords and 4,366 structures.Source: Dataset Curation · p. 7

MOF Solvent Removal Stability Classification

Across 2,179 MOFs, no single heuristic strongly predicts solvent-removal stability, while machine-learning classifiers capture structure–stability relationships and reveal that linker and metal-node chemistry both contribute.

MOF Solvent Removal Stability Classification

  • The 2,179-MOF dataset slightly favored stable labels (59%), and no single heuristic such as pore volume or linker size correlated strongly with stability.The authors attribute the stable-label majority to expected publication bias toward successful results.Source: MOF Solvent Removal Stability… · p. 8
  • AUC 0.81 for the feature-selected GP and 0.79 for the ANN both exceeded random guess (0.5), with the ANN offering greater interpretability.Both models were evaluated on a set-aside test set; perfect classification would have an AUC of 1.0.Source: Figure 2 · p. 9, MOF Solvent Removal Stability… · p. 9
  • About 75% of selected features described linker chemistry, while about 20% captured metal-SBU chemistry; selected linker features emphasized substituent identity, size, and functional-group placement.Only one geometric descriptor, gravimetric surface area, was selected; metal-node features included elemental identity, relative size, and, to a lesser extent, electronegativity.Source: MOF Solvent Removal Stability… · p. 10
  • The model correctly classified rigid-linker UKIQOV as stable but XEJTIR, with small linkers and mononuclear Cu/Al SBUs, as unstable, showing linker rigidity alone can mislead.Training-set neighbors indicated that mononuclear transition-metal SBUs can be unstable despite rigid linkers, so linker design must account for SBU chemistry.Source: MOF Solvent Removal Stability… · pp. 11–12
  • The classifier identified porphyrinic MEGBAD with trinuclear Mn SBUs as unstable upon solvent removal, and correctly predicted porphyrinic AVILEY with an Hf-containing SBU to be stable with solvent removal; predictions were least accurate for 3d transition-metal MOFs.Source: MOF Solvent Removal Stability… · p. 12

MOF Thermal Stability Regression

Across 3,132 MOFs, machine-learning models predict thermal decomposition temperatures with test errors near experimental uncertainty. Feature analysis points to roles for linker and metal chemistry in thermal stability.

MOF Thermal Stability Regression

  • Across 3,132 chemically diverse MOFs, thermal decomposition temperatures ranged from 35°C to 650°C, far exceeding the typical TGA uncertainty of about 25°C.No single heuristic strongly correlated with decomposition temperature, and observed metal dependence was minor, likely because the dataset includes mixed-metal MOFs.Source: MOF Thermal Stability Regression · pp. 12–13
  • GPR and ANN models achieved test-set MAEs of 44–47°C, with errors balanced across metals and only somewhat above typical experimental uncertainty.The paper reports that the models are faster and more accurate than conventional simulation.Source: MOF Thermal Stability Regression · p. 13
  • Analysis of features governing thermal stability indicates both metal-node and linker chemistry are important, with linkers selected most frequently; about 65% of features encode nuclear charge or electronegativity.For the metal-centered subset, this share rises to about 73%. Unlike solvent removal, thermal stability is less dependent on the size of atoms around the metal.Source: MOF Thermal Stability Regression · p. 14
  • Thermal stability features suggest that linker rigidity and metal–ligand bond strength matter, while linker chemistry can exert greater influence on decomposition temperature than metal substitution.For example, a Zr MOF with piperazine phosphonic acid linkers had lower decomposition temperature than UiO-67 with biphenyl dicarboxylate linkers.Source: MOF Thermal Stability Regression · pp. 15–16
  • For the same Mn SBU, the GPR model predicts that the larger bis(oxy)dibenzoate-substituted carboxylate linker YINCIJ reduces Td by 100°C; the biphenyl dicarboxylate Mn MOF JOYKUF is predicted exceptionally stable at 427°C.Source: MOF Thermal Stability Regression · p. 16

Engineering Stability into Existing MOFs

The section applies confident ML predictions to synthesized MOFs lacking extracted stability data, then uses learned structure–stability patterns to propose modifications that improve one or both stability measures.

Engineering Stability into Existing MOFs

  • In about an hour, the models predict both stability measures for nearly 1,500 synthesized MOFs—several orders of magnitude faster than the fastest alternatives.Direct manual inspection could take minutes to hours per MOF, while new synthesis and measurement could take days to weeks per MOF.Source: Engineering Stability into Existing MOFs · p. 17
  • MOFs with alkali, alkaline-earth, or heavy metals are predicted to be stable on both measures, with lighter Zn MOFs as the main exception.The predictions also identify Zn examples with good solvent-removal stability, illustrating that metal identity alone does not determine both measures.Source: Engineering Stability into Existing MOFs · p. 17, Figure 7 · p. 18
  • Nearly all 3d transition-metal MOFs and late 4d Cd MOFs are predicted solvent-stable but thermally unstable, although exceptions occur, including alkali- and alkaline-earth-metal MOFs.This shows that a metal’s identity does not by itself determine both stability measures.Source: Engineering Stability into Existing MOFs · p. 18
  • Cl-, B-, or Si-containing MOFs are often thermally unstable, whereas fluorine-containing MOFs are more stable; oxygen-coordinating linkers yield more stable MOFs than nitrogen-coordinating linkers.The elemental and coordinating-atom patterns are general tendencies, not universal rules.Source: Engineering Stability into Existing MOFs · pp. 18–19
  • Despite expectations of reduced thermal stability45, several phosphonate metal-coordinating MOFs are predicted stable by both models.Source: Engineering Stability into Existing MOFs · p. 19
  • For naphthalene dicarboxylate linkers, the model distinguishes HEBTAL and correctly predicts its collapse upon solvent removal, confirmed by manual literature inspection.The article could not be automatically downloaded and mined, making this a naturally blinded test; the passage cites literature73.Source: Engineering Stability into Existing MOFs · p. 19
  • Removing methoxy groups from Zn MOF XAMHEA is predicted to raise solvent-removal stability from 0.11 to 0.89 while retaining thermal stability (433°C vs 430°C).Feature analysis suggested that linker heteroatoms distant from the metal reduce solvent-removal stability; the redesign maintains the MOF connectivity.Source: Engineering Stability into Existing MOFs · p. 19, Figure 8 · p. 20
  • For SODQOT, whose Td is 182°C under light-alkane-oxidation catalytic conditions, the model predicts a 130°C Td increase without compromising solvent-removal stability, making this MOF suitable for catalysis (28, 74).The improvement is likely due to the more electron-deficient fluorinated linker enabling stronger metal–ligand bonds.Source: Engineering Stability into Existing MOFs · pp. 20–21

Conclusions

The conclusions show that broad MOF-stability data challenge family-specific heuristics, enable rapid prediction and interpretation of stability, and point to ways to improve practical materials.

Conclusions

  • Existing MOF stability heuristics derived within a single family fail when applied more globally, as revealed by relationships across chemical and geometric structures.The analysis examined stability in relation to MOF chemical and geometric structures.Source: Conclusions · p. 21
  • Trained models using graph- and pore-structure-based representations predict thermal and solvent-removal stability orders of magnitude faster than conventional physics-based modeling.The models encode structure–property relationships for both stability measures.Source: Conclusions · p. 21
  • ML interpretation identifies linker connectivity, or bond saturation, as a determinant of solvent-removal stability and linker chemistry as a determinant of thermal stability.The models also supported testing stability heuristics, including predictions of stability for specific large-pore MOFs confirmed against literature reports.Source: Conclusions · p. 22
  • Expanded analysis of the CoRE MOF database found that oxygen-coordinating linkers are typically much more stable than nitrogen-coordinating linkers, which may be preferred for catalysis.Automated extraction of both thermal and solvent-stability information was unavailable for this broader set.Source: Conclusions · p. 22
  • For frequently heat-unstable 3d transition-metal MOFs targeted for catalysis, linker-chemistry modifications could increase thermal stability for catalytic applications.The models were used to identify strategies for engineering more stable MOF structures.Source: Conclusions · p. 22
  • First-generation model performance was constrained primarily by the amount of literature data that could be readily extracted.The models and extracted properties are provided in a user-friendly interface that welcomes community feedback and additional data.Source: Conclusions · p. 22
  • The authors expect this approach, including extensions to other properties, to accelerate discovery of stable, practical MOFs by computational and experimental researchers.The interface provides access to the models and extracted datasets.Source: Conclusions · p. 22

Software and Workflows

The workflow linked MOFs to article DOIs, assembled and parsed a literature corpus, and digitized thermogravimetric traces to extract decomposition-onset temperatures.

  • The CSD v5.41 Python application and API supplied article DOIs for MOFs identified by unique CSD refcodes.Source: Software and Workflows · p. 23
  • A modified ArticleDownloader automated MOF-article downloads from RSC, Wiley, AAAS, Springer, and Nature, while ACS content came through a direct-download agreement.Source: Software and Workflows · p. 23
  • ChemDataExtractor searched structured HTML and XML, while Stanza dependency parsing helped disambiguate complex sentences tokenized by ChemDataExtractor.Source: Software and Workflows · p. 23
  • Webplotdigitizer was used to digitize thermogravimetric traces from main-text figures and supporting information to identify thermal decomposition-onset temperatures.Source: Software and Workflows · p. 23

Machine Learning Models and Representations

The study combined graph-based and pore-geometry features to train machine-learning models that classify solvent-removal stability and predict thermal decomposition temperature. Kernel models and neural networks were tuned and evaluated using separate training, validation, and test data.

Machine Learning Models and Representations

  • 134 revised autocorrelation descriptors built from five atomic properties and paths up to three bonds were combined with 14 Zeo++ geometric features describing pore size, surface area, and void space.The 134 descriptors were retained from 160 possible RACs after invariant features were removed.Source: Machine Learning Models and… · pp. 23–24
  • Models independently classified solvent-removal stability and predicted thermal decomposition temperature using a random 80%/20% train/test split, with 16% of all data reserved for validation.The validation subset comprised 20% of the training data and was used for hyperparameter selection.Source: Machine Learning Models and… · p. 24
  • Kernel-based models were tuned by exhaustive grid search, while ANNs used Hyperopt; random-forest-ranked recursive feature addition selected features for KRR, GPR, SVC, and GPC models.Feature addition was evaluated on validation metrics and stopped when performance ceased improving.Source: Machine Learning Models and… · p. 24
  • RBF kernels were used for KRR, GPR, SVC, and GPC, with additional Matern-3/2 and Matern-5/2 kernels for GPR and GPC; final models were tested on the held-out test set.Validation used MAE for regression and AUC for classification.Source: Machine Learning Models and… · p. 24

TOC GRAPHICS · RACs: [625.0,

This supporting-information section describes the curation, featurization, and analysis of MOF structures and stability evidence, including structure-based classification, feature correlations, and stability prediction. It also reports uncertainty-cutoff filtering and the resulting prediction set of 1,492 CoRE MOFs after removing training, validation, and test overlaps.

RACs: [625.0,

  • RACs were generated for 9,597 of 10,143 all-solvent-removed CoRE MOFs; failures included residual solvent, atom overlap, irreducible cells, missing metals, tiny linkers, and geometric-property errors.The starting set was the 10,143-structure ASR subset of CoRE MOF.Source: Table S1 · p. 36
  • Dependency parsing checks keyword-matched solvent-removal stability sentences for negation, while keyword filters distinguish solvent evacuation from water or air stability.The workflow uses solvent plus stability/collapse keywords, then parses sentence dependencies so negations such as “prevents collapse” are labeled correctly.Source: Table S4 · p. 39, Table S5 · p. 41, Figure S2 · p. 42
  • The thermal dataset contains 3,132 curated decomposition temperatures from 1,886 manuscripts, measured from TGA traces at tangent-line intersections.A distinct solvent-removal step and stability plateau precede decomposition in solvent-removal-stable MOFs; unstable MOFs can decompose upon solvent removal.Source: Figure S4 · p. 43, Table S7 · p. 44
  • Metal identity, linker coordinating atoms and size, and pore volume rank among features associated with both stability labels and thermal decomposition temperatures, but no individual feature correlates strongly.The solvent-removal analysis uses point-biserial correlations; thermal decomposition uses Pearson correlations.Source: Table S9 · p. 48, Table S13 · p. 57
  • 64% test accuracy and 0.60 test AUC were achieved by the linker-and-geometry-only decision tree, which performed worse on unseen data than the GPC classifier.Despite including pore-size features, its tree splits were determined by linker size and topology.Source: Figure S10 · p. 51
  • At LSE 0.19, solvent-removal predictions covered 66% of test data with 81% accuracy and 0.82 AUC; scaled LSD 0.21 retained 51% with thermal MAE of 39°C.The corresponding test-set baselines shown are AUC 0.79 for solvent-removal classification and MAE 47°C for thermal stability.Source: Figure S22 · p. 64
  • After excluding MOFs overlapping the training, validation, or test sets and applying the UQ cutoffs, 1,492 of the 9,597 CoRE MOFs remained for prediction.The solvent-removal LSE cutoff was 0.19 and the thermal scaled LSD cutoff was 0.21.Source: Table S17 · p. 65, Figure S23 · p. 66
  • The timing analysis profiles structure-specific descriptor generation separately from normalization and prediction, and estimates predictions for 1,500 compounds parallelized across 8 CPU cores.The supplied caption gives no elapsed-time values.Source: Table S18 · p. 65
  • The 1,492 predicted MOFs were profiled by pore size, metal and linker composition, and linker identities, with examples of phosphonate candidates and a solvent-removal-unstable MOF.The HEBTAL example was reported unstable on solvent removal despite its small, rigid linker.Source: Figure S24 · p. 67, Figure S26 · p. 69, Figure S27 · p. 70
  • The predicted MOFs were classified into stability quadrants using solvent-removal probability 0.5 and thermal decomposition temperature 359°C as cutoffs.The 359°C threshold is the average thermal decomposition temperature in the thermal stability dataset.Source: Table S19 · p. 68, Figure S26 · p. 69
Loading 2106.13327v1…