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
TL;DR
Limited understanding of MOF stability constrains practical use. The paper uses machine-learning models and reports stability-engineering results, including preserved thermal stability and improved solvent-removal stability.
Problem
Limited understanding of how to improve MOF stability constrains their practical use.
Method
Machine-learning models use graph- and pore-structure-based representations to predict MOF stability and guide stability engineering.
Results
0.11 vs 0.89 solvent-removal stability was achieved while predicted thermal stability remained 433°C vs 430°C.
Takeaways & Limitations
The approach supports rapid stability engineering and can preserve good thermal stability while improving solvent-removal stability.
Takeaways & Limitations
Predictions are least accurate for 3d transition-metal MOFs.
Abstract
from arXiv · showhide
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 study curates MOF stability datasets from published structures and manuscripts using automated text and image analysis. It resolves linguistic ambiguity and extracts labeled solvent-removal stability data and thermal decomposition temperatures.
- Structure and manuscript collection: 10,143 solvent-removed crystal structures were screened, with structures filtered for featurization and linked to downloadable manuscripts.The workflow began from the ASR structure set and used CoRE MOF/CSD mappings to connect structures with papers.
- Text processing: 3,809 manuscripts were processed by sectioning and sentence tokenization, while identifiable introductions or their first 40% were excluded from keyword analysis.The first-40% cutoff was selected by trial and error when no introduction heading was available.
- Solvent-removal stability extraction: Regular-expression matching combined solvent or activation terms with structural-integrity terms to identify solvent-removal stability statements.Dependency parsing then mapped related words and distinguished affirmative, negative, double-negative, and complex stability statements.
- Solvent-removal stability extraction: 2,179 structure–label pairs formed the solvent-removal stability dataset after manuscript-level labeling and removal of duplicate structures with conflicting labels.Labels were assigned when all solvent-evacuation-related sentences agreed; conflicting cases were not automatically assigned.
- Dataset yields: 1,209 manuscripts covering 2,290 structures were unambiguously labeled for solvent-removal stability, after 2,649 manuscripts contained relevant keywords.The final dataset was called the solvent-removal stability dataset (SSD).
MOF Solvent Removal Stability Classification
Solvent-removal stability varies across MOFs in ways that defeat single heuristics, so the study uses interpretable ML classification to connect linker, metal-node, and structural features with stability.
- Dataset and motivation: 2,179 MOFs were assigned solvent-removal stability labels from extracted literature data.
- Dataset and motivation: No single heuristic, including pore volume or linker size, correlates strongly with solvent-removal stability across the dataset.
- ML classification: The feature-selected GP classifier achieved AUC 0.81 on a set-aside test set, while the ANN classifier achieved AUC 0.79.
- Interpretable features: Approximately 75% of selected features describe linker chemistry, emphasizing substituent identity, size, and functional-group placement over linker rigidity.
- Interpretable features: Metal secondary-building-unit chemistry contributes approximately 20% of selected features, especially elemental identity and relative size.
- Design implications: Correct predictions show that solvent-removal stability depends jointly on linker design and SBU chemistry, including cases where rigid linkers do not ensure stability.
MOF Thermal Stability Regression
Thermal decomposition temperatures span a broad range, and ML regression predicts them accurately while indicating that linker and metal electronic chemistry are important design variables.
- Dataset and prediction: Both feature-selected GPR and ANN models achieved mean absolute errors of 44–47°C on a set-aside test set.
- Dataset and prediction: ML predictions were more rapid and accurate than conventional simulation, with errors only somewhat larger than typical experimental uncertainty.
- Interpretable features: Metal-node and linker chemistry are important predictors, with linker features selected most frequently and approximately 65% of features primarily electronic.
- Interpretable features: Approximately 73% of metal-centered features are electronic, emphasizing nuclear charge or electronegativity rather than atom-size effects.
- Design implications: The analysis suggests orthogonal design of stability properties by tuning electronic SBU properties for thermal stability and geometric properties for solvent-removal stability.
Engineering Stability into Existing MOFs
The authors use ML predictions to identify stability patterns and propose targeted structural modifications for improving existing MOFs. These predictions support rapid screening and suggest ways to tune solvent-removal and thermal stability independently.
- Predictive screening: ML models predict both stability measures for nearly 1,500 synthesized MOFs within their domain of applicability.The models were applied conservatively where they were reasonably confident.
- Predictive screening: The models evaluate the full prediction set in about an hour, several orders of magnitude faster than available alternatives.Conventional evaluation of each MOF could require minutes to hours, or days to weeks under conservative assumptions.
- Stability patterns: Metal chemistry separates stability trends but does not determine both properties: heavier and alkaline metals are often stable, whereas many 3d metals show good solvent stability but poor thermal stability.Lighter Zn MOFs are a primary exception among metals otherwise associated with both stability measures.
- Stability patterns: Linker chemistry also differentiates stability: fluorine-containing and oxygen-coordinating linkers are generally more stable than several alternatives, while Cl, B, Si, and nitrogen-coordinating linkers are often associated with instability.The authors also identify phosphonate and linker-isomer exceptions to these broad trends.
- Targeted engineering: Feature analysis assigns distinct roles to linker connectivity and linker chemistry in solvent-removal and thermal stability, enabling orthogonal design strategies.The models identify metal-proximal chemistry as important for solvent removal and more distant functional groups as important for thermal stability.
- Targeted engineering: Fluorinating the linker of a Cu MOF is predicted to improve thermal stability by 130°C without compromising solvent-removal stability.The proposed improvement is attributed to a more electron-deficient linker enabling stronger metal–ligand bonds.
- Scope and limitations: The approach is constrained primarily by the amount of stability data that can be extracted from the literature.The authors state that data limitations constrain first-generation model performance more than the representations or ML models themselves.
Software and Workflows
The workflow builds a searchable corpus of MOF literature, extracts structured chemical information, and digitizes thermal-analysis figures to obtain stability data.
- Corpus construction: Article downloads from multiple publishers are automated to construct a corpus of MOF articles.The workflow includes direct downloads and an article-downloading toolkit for several publishers.
- Corpus construction: ChemDataExtractor parses downloaded documents and performs tree-based searches for relevant chemical information.The workflow parses structured documents before downstream stability extraction.
- Text mining: ChemDataExtractor tokenization and Stanza dependency parsing disambiguate complex sentences and identify stability-related statements.Dependency parsing is used to handle sentence structure and negation.
- Thermal-data extraction: TGA traces from article text and supporting information are digitized with WebPlotDigitizer to identify thermal decomposition onset temperatures.The digitized results are supplied in the Supporting Information.
Machine Learning Models and Representations
The study represents MOFs with graph-based RAC descriptors and geometric pore features, then trains classification and regression models for the two stability properties.
- Representations: RAC descriptors encode atomic properties along bond paths, while Zeo++ features quantify pore size, surface area, and void space.The representations include 134 non-invariant RACs and 14 geometric features.
- Modeling: The models independently classify solvent-removal stability and predict thermal decomposition temperature.The workflow uses both classification and regression formulations.
- Modeling: Kernel models and artificial neural networks are trained with hyperparameter optimization and an 80%/20% train/test split.A portion of the training data is reserved for validation during hyperparameter selection.
- Modeling: Feature selection evaluates validation performance using metrics such as MAE for regression and AUC for classification.Recursive feature addition stops when validation metrics no longer improve.
- Evaluation: Final models are trained on the full training set and evaluated on a held-out test set.The workflow uses kernel-based models and neural networks implemented with established ML software packages.
RACs: [625.0,
The supporting workflow documents literature filtering, structure mapping, stability-keyword extraction, and dataset construction for MOF stability analysis.
- Structure processing: The CoRE MOF starting set contains 10,143 all-solvent-removed structures, from which RACs can be generated for 9,597 MOFs.The supporting table records structure attrition and featurization failure modes.
- Stability extraction: Solvent-removal stability extraction combines solvent keywords with stability or collapse keywords and checks dependency-based negation.The resulting sentence labels are aggregated to assign material-level stability labels.
- Data curation: The workflow excludes false positives concerning water or air stability and removes duplicate structures with conflicting solvent-removal labels.The supporting figures document duplicate removal and checks for other stability classes.
- Thermal-data extraction: Thermal-data curation maps TGA-related manuscript evidence to thermal decomposition temperatures.The supporting statistics distinguish manuscripts containing TGA keywords, traces, and mapped temperatures.
- Dataset composition: The solvent-removal dataset includes 2,179 MOFs, with most containing one metal and smaller fractions containing two or three metals.The supporting distribution counts 1,815 single-metal, 352 two-metal, and 11 three-metal MOFs.