Accelerating the Selection of Covalent Organic Frameworks with Automated Machine Learning.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Covalent organic frameworks (COFs) offer high thermal stability and surface area, making them promising for gas storage and catalysis.
• Predicting methane (CH4) working capacity in COFs is crucial but computationally intensive due to the vast number of possible structures.

Purpose of the Study:

• To apply automated machine learning (AutoML) for accurate prediction of CH4 working capacity in COFs.
• To compare the performance of AutoML with traditional machine learning (ML) models and simulation methods.

Main Methods:

• Utilized AutoML, specifically the Tree-based Pipeline Optimization Tool (TPOT), to analyze CH4 working capacity across 403,959 COFs.
• Explored relationships between 23 COF features (structural, chemical, atomic) and CH4 working capacity.
• Compared TPOT performance against multiple linear regression, support vector machines, decision trees, and random forests.

Main Results:

• AutoML (TPOT) demonstrated superior performance compared to traditional ML models.
• TPOT significantly reduced the time and complexity associated with data preprocessing and model parameter tuning.
• AutoML offers a substantial time saving compared to traditional grand canonical Monte Carlo simulations.

Conclusions:

• AutoML democratizes advanced material screening by enabling non-expert researchers to achieve accurate predictions.
• AutoML facilitates rapid identification of high-performance COFs for applications like methane storage.
• This approach accelerates the discovery process for novel materials with tailored properties.