MOF Synthesis Prediction Enabled by Automatic Data Mining and Machine Learning**

Yi Luo, Saientan Bag, Orysia Zaremba, Adrian Cierpka, Jacopo Andreo, Stefan Wuttke, Pascal Friederich, Manuel TsotsalasAngewandte 2022 · DOI

TL;DR

Researchers have developed a machine learning (ML) approach to predict the synthesis conditions for metal-organic frameworks (MOFs) directly from their crystal structures. By automatically extracting synthesis data from literature and training ML models, they've created a tool that can suggest experimental parameters, significantly accelerating the discovery of new MOFs and outperforming human experts in prediction tasks.

What Problem Does This Paper Address?

The field of Metal-Organic Frameworks (MOFs) has exploded, with hundreds of thousands of known structures and millions more theoretically possible. However, synthesizing these materials is still largely a trial-and-error process, relying heavily on expert intuition and extensive literature searches. This makes discovering new MOFs and optimizing their synthesis incredibly time-consuming, resource-intensive, and a major bottleneck in the field. The paper addresses the challenge of "inverse synthesis design" for MOFs: given a desired MOF structure, how can we efficiently predict the experimental conditions needed to create it?

Methods

The core of this work is a three-pronged approach:

  1. Automated Data Mining: The researchers developed a pipeline to automatically extract synthesis parameters (metal source, linker, solvent, additive, synthesis time, and temperature) from scientific literature. This involved classifying paragraphs using string search and then employing a tool called ChemicalTagger to identify and extract experimental details. They also extracted structural information from Crystallographic Information Files (CIFs).
  2. ML Model Training: They trained various ML models using the extracted data. For predicting continuous parameters like temperature and time, regression models like random forests and neural networks were used. For discrete parameters like solvents and additives, they adopted a more nuanced approach: predicting solvent properties (like partition coefficients and boiling points) and then using a nearest-neighbor search to suggest suitable solvents, and classifying additives into acidic, basic, or no additive categories.
  3. Prediction and Evaluation: The trained ML models were used to predict synthesis conditions for new MOF structures. Crucially, their performance was compared against predictions made by human MOF synthesis experts.

Datasets

The primary dataset introduced and utilized in this paper is the SynMOF database. This database was created by automatically extracting synthesis parameters and structural information from literature for over 900 MOF structures. The authors also generated manually corrected versions (SynMOF-M and SynMOF-ME) for evaluation purposes, which are discussed in the supporting information.

Key Insights and Ideas

  • Bridging the Structure-Synthesis Gap: The most significant insight is that a MOF's crystal structure alone contains enough information to predict its synthesis conditions, moving beyond purely empirical or intuition-based approaches.
  • Automated Data Extraction is Key: The success hinges on the ability to automatically parse vast amounts of unstructured text (scientific literature) into structured, usable data. This NLP-driven approach is a major step forward for data-driven materials science.
  • ML Outperforms Human Experts (Surprisingly): The paper's finding that ML models, even in their initial stages, outperform human experts in predicting synthesis conditions is a powerful testament to the hidden patterns that ML can uncover. This highlights the limitations of human intuition when dealing with the vast complexity of MOF chemistry.
  • Predicting Properties Over Specifics: For parameters like solvents, predicting underlying properties (e.g., boiling point) rather than specific chemical names proved more robust and flexible, allowing for the suggestion of novel or less common solvents.

Key Results

  • The developed ML models achieved good prediction performance for synthesis parameters, particularly temperature and time, with reproducible positive correlation coefficients (r²) on unseen test datasets.
  • The ML solvent prediction accuracy for single-solvent MOFs exceeded random prediction methods, reaching over 90% for the top 1-3 predictions.
  • Additive classification showed promise, though differentiation between certain categories was less pronounced due to dataset imbalances and the complex role of additives.
  • A direct comparison with 11 human MOF synthesis experts revealed that ML models achieved better correlation coefficients (r²) for predicting temperature and time, even when experts were asked to report high certainty.

Limitations and Future Work

  • Dataset Size and Imbalance: The current SynMOF database, while substantial, is still limited, especially for less common synthesis parameters like additives. Imbalanced datasets (e.g., for additives) pose a challenge for model accuracy.
  • Complexity of Synthesis: The authors acknowledge that MOF synthesis is an "ill-defined task" with a range of acceptable conditions, meaning a perfect r² score of 1 is unlikely. The influence of factors like additive concentration and strength, which are not fully captured, also limits current models.
  • Future Directions:
    • Expanding the SynMOF database with more data and improving NLP extraction techniques.
  • Developing more sophisticated ML models, such as graph neural networks, which are better suited for complex structural representations and can learn from larger datasets.
  • Improving the representation of additives and exploring their concentration and function.
  • The authors have made their automated synthesis prediction available via a web-tool, encouraging community adoption and further development.

Summary updated

Abstract

Despite rapid progress in the field of metal-organic frameworks (MOFs), the potential of using machine learning (ML) methods to predict MOF synthesis parameters is still untapped. Here, we show how ML can be used for rationalization and acceleration of the MOF discovery process by directly predicting the synthesis conditions of a MOF based on its crystal structure. Our approach is based on: i) establishing the first MOF synthesis database via automatic extraction of synthesis parameters from the literature, ii) training and optimizing ML models by employing the MOF database, and iii) predicting the synthesis conditions for new MOF structures. The ML models, even at an initial stage, exhibit a good prediction performance, outperforming human expert predictions, obtained through a synthesis survey. The automated synthesis prediction is available via a web-tool on https://mof-synthesis.aimat.science.
BibTeX
@inproceedings{luo2022mof,
  author = {Yi Luo and Saientan Bag and Orysia Zaremba and Adrian Cierpka and Jacopo Andreo and Stefan Wuttke and Pascal Friederich and Manuel Tsotsalas},
  title = {{MOF Synthesis Prediction Enabled by Automatic Data Mining and Machine Learning**}},
  booktitle = {Angewandte},
  year = {2022},
  doi = {10.1002/anie.202200242},
}
Figure