Catalysis is a process that speeds up chemical reactions by using a substance called a catalyst. Catalysts are widely used in various fields, such as energy, environment, and medicine. The performance of a catalyst depends on its surface structure, which can affect how molecules interact with it. Therefore, understanding and predicting the fine structures of catalytic surfaces is crucial for designing and optimizing new catalysts.
However, this is not an easy task, as catalytic surfaces are complex and diverse. Traditional methods of characterizing and modeling catalytic surfaces are often time-consuming, expensive, and inaccurate. To overcome these challenges, researchers have been exploring the use of machine learning (ML) techniques to analyze and predict the properties of catalytic surfaces based on data.
A novel ML framework that combines global and local features
Recently, a research team led by Prof. Yong Wang from Zhejiang University, China, developed a novel ML framework that can capture the fine structures of catalytic surfaces without losing any information. The framework, named GLCNN, combines two types of features: global and local. Global features are derived from two-dimensional grids that represent the atomic positions and types on the catalytic surface. Local features are derived from one-dimensional descriptors that represent the adsorption sites where molecules bind to the surface.
The GLCNN framework uses a convolutional neural network (CNN) to process the global features and a fully connected neural network (FCN) to process the local features. The CNN can extract the geometric and electronic information of the catalytic surface from the grids, while the FCN can capture the specific characteristics of the adsorption sites from the descriptors. The outputs of the CNN and FCN are then concatenated and fed into another FCN to generate the final prediction.
The GLCNN framework also employs a data augmentation (DA) technique to increase the size and diversity of the training dataset. DA is a method of creating new data samples by applying transformations to existing data samples. For example, rotating or flipping a grid can create a new grid with a different orientation but the same structure. DA can help improve the generalization and robustness of the ML model by exposing it to more variations of data.
A high-precision prediction of catalytic properties
The researchers tested the GLCNN framework on a dataset of carbon-based transition metal single-atom catalysts (TMSACs), which are promising candidates for various catalytic reactions. The dataset contains 1,440 samples of TMSACs with different metal atoms, carbon supports, and adsorption sites. The target property is the adsorption energy of hydroxyl (OH) groups on the TMSACs, which is an important indicator of their catalytic activity.
The results showed that the GLCNN framework achieved a high accuracy in predicting the OH adsorption energy, with a mean absolute error (MAE) of less than 0.1 eV. This is better than any other popular ML models trained on large datasets so far. The GLCNN framework also outperformed other ML models that use descriptors or graphs as inputs, demonstrating its ability to capture the fine structures of catalytic surfaces.
The researchers also analyzed the importance of different features in the GLCNN framework by using a sensitivity analysis method. They found that both global and local features contribute significantly to the prediction performance, but global features have more influence than local features. This suggests that the geometric and electronic information of the catalytic surface is more critical than the specific characteristics of the adsorption sites.
A new tool for exploring catalysis
The GLCNN framework is a new tool for exploring catalysis based on ML techniques. It can accurately and efficiently predict the properties of catalytic surfaces based on their fine structures without any complicated encoding methods. It can also provide insights into the key factors that affect catalytic performance from both geometric and chemical/electronic perspectives.
The researchers hope that their work can inspire more studies on developing interpretable ML frameworks that can simultaneously capture the features of electronic and geometric fine structures in heterogeneous catalysis. They also plan to apply their framework to other types of catalysts and reactions in future work.
