CHEMICAL SCIENCE

Abandoned Data – An effective puzzle that enables AI to predict the separation performance of porous materials with high precision


On August 22, 2022, the team of Professor Xing Huabin of Zhejiang University published a research report entitled “Machine-learning-assisted exploration of anion-pillared metal organic frameworks for gas separation” in the journal Matter.

The research group proposed a general machine learning training strategy to achieve high-precision prediction of adsorption properties of porous materials by combining waste experimental data and computational chemical parameters, and for the first time realized the high-precision prediction of the experimental adsorption properties of anionic column support ultra-microporous materials acetylene, ethylene and carbon dioxide, and screened and prepared two new anion column bracing porous materials with high adsorption capacity and high separation selectivity. Professor Xing Huabin of Zhejiang University is the corresponding author of the paper; Hu Jianbo is the first author of the paper.

In recent years, machine learning methods have been introduced into materials science research, which can significantly improve the efficiency of material research and development. Machine learning enables computers to mine the associations between material structure and properties from the training data set and predict the performance of the new material in the test set. The key to the effectiveness of machine learning is to achieve high-precision predictions of material properties, and the accuracy of predictions depends heavily on the quality of the training data. In the field of adsorption separation, training data, including material structural feature descriptors and material property data, can be obtained by high-throughput computational simulations or experimental characterization tests. High-throughput simulation data is the main source of machine learning training data in the field of adsorption and separation, which has the characteristics of large amount of data, high consistency and good integrity, but it is limited by the force field parameters, and the accuracy of the material adsorption performance data obtained by the calculation simulation method is low, especially for materials with strong action sites and flexible features, which limits the prediction accuracy of the machine learning model. Compared with the calculation of simulation data, higher precision material performance data can be obtained based on experimental characterization test methods, but the experimental data reported in the existing literature are generated at different times and different researchers, and most of the reported data are material performance data with excellent performance, and the data consistency is insufficient, the integrity is poor, and it is difficult to meet the needs of machine learning training. Achieving high-precision predictions of the actual properties of materials based on machine learning is still challenging.

In this work, The team of Xing Huabin of Zhejiang University proposed a general machine learning training strategy that combines discarded experimental data and calculated chemical parameters to achieve high-precision prediction of material properties, which solves the problem of low prediction accuracy of machine learning models due to the difficulty of obtaining consistent, complete and accurate data in the current adsorption separation field. Based on this general machine learning strategy, this study for the first time realizes the high-precision prediction of the experimental adsorption performance of acetylene, ethylene and carbon dioxide of anion column support ultraporous materials, and screens and prepares two new ionic hybrid porous materials with high adsorption capacity and high separation selectivity. At the same time, the relationship between material structural characteristics and adsorption properties is quantitatively described, which provides more accurate and intuitive guidance for the design of new adsorbents.

Figure 1: Schematic diagram of a machine learning workflow

Firstly, the experimental adsorption isotherms discarded in the past research process of the research group were collected, and the published experimental adsorption isotherms data were further collected as a supplement. Further by computational chemical methods to calculate material structural characteristic descriptors. For each metal-organic skeleton material, calculate its three structural feature descriptors, including 1) crystal structure feature descriptors, which are used to describe the structural characteristics of the material’s pores; 2) density of chemical elements, used to describe the chemical characteristics of the surface of the material; 3) Maximum structural characteristic parameters, used to describe the flexible characteristics of the material. A machine learning prediction model is constructed using a random forest algorithm to identify the correlation between material structure feature descriptors and material adsorption data.

Figure 2: Prediction effect of machine learning model on the adsorption properties of acetylene, ethylene and carbon dioxide of anion column support ultra-microporous materials

The prediction results of the training set and the test set show that the introduction of the maximum structural feature parameter can significantly improve the prediction accuracy of the machine learning model. Both the predictions of the training and test sets matched the experimental data well, indicating that the machine learning model was able to capture the correlation between feature descriptors and adsorption performance. The comparison between different training datasets can be found that the prediction accuracy based on experimental data is significantly higher than that based on computational simulation data, and the introduction of discarded data can significantly improve the prediction accuracy of machine learning.

Figure 3: Machine learning-assisted exploration of anion column-braced ultraporous materials with optimal adsorption separation performance

Based on the high-precision machine learning model constructed in this paper, the adsorption properties of materials on acetylene, ethylene and carbon dioxide were predicted, and two adsorption separation materials ZU-96 and ZU-63 with high selectivity and high adsorption capacity were screened and synthesized according to the material adsorption performance data and separation performance data. Among them, ZU-96 material has benchmark co2/acetylene separation performance, and its carbon dioxide adsorption capacity reaches 83.2 cm3/cm3 and carbon dioxide/acetylene IAST selectivity reaches 81.5 at 298 K and 0.1 bar.

Figure 4: The relative importance of the structural characteristics of anion column bracing porous materials to the gas adsorption properties of materials

Based on the random forest algorithm, the influence weights of the structural characteristics of anion column-supported porous materials on the adsorption properties of acetylene, carbon dioxide and ethylene are quantitatively calculated, which provides a more accurate and intuitive guide for the subsequent design of adsorbents with ideal adsorption separation properties.

This research work shows that waste data is an effective puzzle to achieve AI accurate prediction of the actual performance of materials. Based on the findings of the study, the authors call on materials scientists to provide more material adsorption data in future published literature, especially those that have been abandoned during the material design process, which will greatly accelerate the development of materials. (Source: Science Network)

Related paper information:https://doi.org/10.1016/j.matt.2022.07.029



Source link

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button