The article “An end-to-end artificial intelligence prediction method for carbon capture performance of metal-organic framework materials based on deep learning” was selected as the cover article of the Journal of Chemical Information and Modeling courtesy of the Wanxili Research Group
How to make performance predictions for hundreds of thousands of materials in minutes to find carbon capture and storage (CCS) metal-organic framework materials (MOFs) with high-quality performance? Recently, Lu Cunxing, a 2020 master’s student under the guidance of Wan Xili, associate professor of the School of Computer Science and Technology of Nanjing University of Technology, solved this problem through a calculation method.
Recently, the related research was “an end-to-end artificial intelligence prediction method for carbon capture performance of metal-organic framework materials based on deep learning”, which was accepted by the Journal of Chemical Information and Modeling published by the American Chemical Society, and was selected as the cover article of the current issue.
How to find MOFs with excellent CCS performance in a large number of MOFs is a difficult point in current academic research. Metal-organic frame materials are considered ideal for carbon capture and storage due to their excellent performance in gas adsorption, but the highly adjustable MOFs lead to the generation of millions of different MOFs.
According to Wan Xili, Lu Cunxing elaborated an end-to-end artificial intelligence prediction method for carbon capture performance of metal-organic framework materials based on deep learning, which has two major advantages. First, it has the characteristics of deep learning, and can find the optimal prediction method through continuous self-training and adjustment; Second, its end-to-end characteristics allow scholars without computer foundation to use it directly.
In recent years, although there have been research in the academic community to solve this problem, there are some limitations. For example, molecular simulation methods consume a lot of computing resources and are very inefficient when dealing with large data sets. In addition, the method requires both a large number of descriptors to be constructed, a strong prior knowledge and constant trial and error, and a lot of time to invest in feature engineering to obtain descriptor values that pass molecular simulations.
Lu Cunxing’s calculation method avoids the time-consuming and labor-intensive drawbacks of existing calculations, and develops an end-to-end prediction method without constructing descriptors, using only crystallographic information files (CIF) as inputs, and learning high-dimensional features that affect performance adaptively through deep learning, so as to quickly and accurately predict the performance of MOFs.
According to Lu Cunxing, he innovatively uses the projection method to transform the three-dimensional structure in the field of materials into computer-readable two-dimensional information, and realizes end-to-end performance prediction after combining the research hotspots of deep learning in the computer field. Specifically, this self-learning computing method evaluates the gap between its prediction and the true value at the end of each cycle, and then adjusts its own parameters to reduce the gap through this gap, and minimizes the error after multiple cycles, so as to achieve a more accurate quasi-function. “Experiments have shown that our calculation method can predict hundreds of thousands of MOFs in minutes, and the top 12% of MOFs contain 99.3% of real high-performance materials.” In practical applications, this calculation method shortens the calculation time by 1/10, that is, saves the calculation time by nearly an order of magnitude. (Source: China Science Daily, Wen Caifei, Yang Fang)
Related paper information:https://doi.org/10.1021/acs.jcim.2c00092