Multi-view machine learning assisted alcohol reforming catalyst screening method

Abstract

The invention relates to the technical field of machine learning, in particular to a multi-view machine learning assisted alcohol reforming catalyst screening method which comprises the following steps: S1, acquiring data to construct a first data set which comprises composition of an alcohol reforming catalyst, an alcohol reforming condition and a catalyst performance evaluation result; s2, weighting the composition data of the alcohol reforming catalyst in the first data set by adopting a multi-stage weighting element descriptor to construct a second data set; s3, inputting the second data set and the first data set into a parameter-free multi-view clustering model for training to obtain a catalyst performance prediction model; and S4, inputting the catalyst composition and the alcohol reforming condition which are acquired in real time into the catalyst performance prediction model to obtain a catalyst performance evaluation prediction index. The screening method can predict the performance of the catalyst, so that the screening process of the catalyst is accelerated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, in particular to a multi-view machine learning assisted alcohol reforming catalyst screening method. BACKGROUND

[0002] Due to the gradual depletion of non-renewable energy and environmental deterioration, there is increasing attention worldwide to the development of new energy and energy conservation. Hydrogen energy is recognized as the cleanest energy and has great development prospects. Biomass can be easily converted into various liquid fuels, including methanol, ethanol and biodiesel, which can be used as raw materials for hydrogen production. Methanol, as an important biomass-derived oxygen-containing hydrocarbon, has the advantages of renewable, high hydrogen-carbon ratio, convenient storage and transportation. Most importantly, compared with most other fuels (ethanol, glycerol, etc.), methanol can react to generate hydrogen at a lower temperature because it does not need to break C-C bonds. In 2002, Cortright R.D. et al. proposed a method for hydrogen production from methanol reforming, which opened up a way for efficient hydrogen production.

[0003] Methanol reforming for hydrogen production is a catalytic reaction. So far, researchers have been exploring the catalysts for this reaction for decades, which is a time-consuming task. In order to speed up the research and development of its catalysts, we hope to use computers to help us in a simple and effective way to screen methanol reforming catalysts. Traditional single-view clustering methods are applied to catalyst information technology \cite{HAJ2021, BS2023}, but this method can only describe objects from a single aspect of the data set, and cannot accurately grasp multiple information of catalysts, resulting in that the algorithm cannot obtain effective and interpretable clustering results in the field of chemical materials. SUMMARY

[0004] The purpose of the present application is to overcome the problem that the development of methanol reforming catalysts in the prior art is a complex and time-consuming process that requires a large amount of experiments and material costs, and the traditional catalyst screening method is inefficient and cannot meet the rapid development of energy demand. To provide a multi-view machine learning assisted alcohol reforming catalyst screening method, which can predict the performance of catalysts and thus accelerate the screening process of catalysts.

[0005] In order to achieve the above purpose, the present application provides a multi-view machine learning assisted alcohol reforming catalyst screening method, which comprises: S1, obtaining data to construct a first data set, including the composition of alcohol reforming catalysts, alcohol reforming conditions and catalyst performance evaluation results; S2, using multi-level weighted element descriptor (MWED) to weight the composition data of alcohol reforming catalysts in the first data set to construct a second data set; S3, inputting the second data set and the first data set into a parameter-free multi-view clustering model (PMC-IGM) to obtain a catalyst performance prediction model through training; S4, inputting the real-time collected catalyst composition and alcohol reforming condition into the catalyst performance prediction model to obtain a catalyst performance evaluation prediction index.

[0006] Through the above technical solution, the multi-view machine learning assisted alcohol (such as methanol) reforming catalyst screening method can predict the catalyst performance, thereby accelerating the catalyst screening process.

[0007] The application proposes a multi-level weighted element descriptor (MWED) method, which can more comprehensively capture the multi-dimensional characteristics of the catalyst by weighting the catalyst composition ratio vector through element descriptors of different levels.

[0008] The parameter-free multi-view clustering model is adopted, which can learn a consistent embedding from multi-view similarity matrices and spectral embeddings without pre-set parameters, thereby improving the universality and adaptability of the model. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 It is a schematic diagram of three modules of the PMC-IGM model; Figure 2 It is a clustering result graph drawn based on the volume space velocity (GHSV); Figure 3 It is a feature importance result graph; Figure 4 is an XRD spectrum of the selected catalyst after calcination and after reaction, Figure 4-1 Figure 4 is an XRD spectrum of the selected catalyst after calcination and after reaction, Figure 4-2 Figure 4 is an XRD spectrum of the selected catalyst after calcination and after reaction Figure 5 It is a graph of the influence of temperature on various performances of two catalysts; Figure 6 It is a comparison graph of the actual cluster and the predicted cluster of the catalyst. DETAILED DESCRIPTION

[0010] The endpoints of the ranges and any values disclosed in this document are not limited to the precise values recited. The ranges and values should be interpreted as being approximate to the exact ranges and values. For ranges, the endpoints are included between the various ranges, between the endpoints and individual points, and between individual points, and new ranges can be created that fall between the endpoints, the endpoints and individual points, and individual points, which should be considered as being specifically disclosed herein.

[0011] In the present application, the parameter-free multi-view clustering model (PMC-IGM) improves the accuracy and robustness of clustering by fusing information from multiple views. The core of the multi-view clustering model is a machine learning model that can effectively fuse information from different sources, even if the information may be contradictory or incomplete, and extract consistent clustering structure from it.

[0012] The present application provides a multi-view machine learning assisted alcohol reforming catalyst screening method, which comprises the following steps: S1, obtaining data to construct a first data set, including the composition of the alcohol reforming catalyst, the alcohol reforming conditions and the catalyst performance evaluation results; S2, using multi-level weighted element descriptor (MWED) to weight the composition data of the alcohol reforming catalyst in the first data set, and constructing a second data set; S3, inputting the second data set and the first data set into the parameter-free multi-view clustering model (PMC-IGM) for training to obtain a catalyst performance prediction model; S4, inputting the real-time collected catalyst composition and alcohol reforming conditions into the catalyst performance prediction model to obtain catalyst performance evaluation prediction indexes.

[0013] According to a preferred embodiment of the present application, in step S1, the obtained data includes: catalyst element composition, alcohol reforming reaction temperature, alcohol reforming reaction pressure, alcohol reforming steam / carbon ratio (S / C), alcohol reforming oxygen / carbon ratio (O / C), gas hourly space velocity (GHSV), catalyst mass (CM), hydrogen production rate, carbon monoxide selectivity.

[0014] According to a preferred embodiment of the present application, the alcohol is a C1-C6 monohydric alcohol.

[0015] According to a preferred embodiment of the present application, the alcohol is methanol.

[0016] According to a preferred embodiment of the present application, the multi-level weighted element descriptor includes electron cloud structure, microscopic level features and macroscopic level features, wherein the electron cloud structure includes five element descriptors: AN, AM, NOE, NON, NOP and P; its specific meaning is that it is the simplest descriptor that can determine the properties of elements and describe the elements themselves, indicating the periodicity of elements. Different elements in the region have different properties, for example, active metals are usually transition metals, which is determined by the different properties reflected by the different orbits of transition metals.

[0017] According to a preferred embodiment of the present application, the micro-level features include eight element descriptors: AR, CR, IR, AV, EGAR, EGP, FIP and EWF; it reflects the electronic properties of the elements. The gain and loss of electrons usually affect the probability of each element reaction in the whole catalytic reaction process, that is, the influence on the product selectivity has a mechanical connection.

[0018] According to a preferred embodiment of the present application, the macro-level features include five element descriptors: D, AM, HOF, MP and BP. These element descriptors show that various properties of the substances corresponding to the elements at the visible light level affect the final morphology of the prepared catalyst, and the difference in morphology also has different effects on the performance of the catalyst.

[0019] According to a preferred embodiment of the present application, in step S2, the feature importance analysis is performed by using tree models and SHAP values and the like, the importance of the descriptors is compared to see which descriptor has the greatest importance and has the greatest influence on the performance of the catalyst, revealing the influence of reaction conditions, element composition and physical properties on the performance of the catalyst and providing deeper insights.

[0020] According to a preferred embodiment of the present application, the catalyst performance prediction model obtains the clustering results by the hydrogen production rate and the carbon monoxide selectivity.

[0021] According to a preferred embodiment of the present application, the catalyst performance prediction model comprises one or more of the following: a clustering result graph drawn by the GHSV, a catalyst element composition clustering result graph, an alcohol reforming reaction temperature clustering result graph, an alcohol reforming reaction pressure clustering result graph, and an alcohol reforming steam-to-carbon ratio clustering result graph.

[0022] According to a preferred embodiment of the present application, the clustering results are evaluated by one or more of the following: accuracy (ACC), normalized mutual information (NMI), F-score, R-score, adjusted Rand index (ARI), mean variance, mean square error (MSE), mean absolute error (MAE), coefficient of determination (R 2 ) and the like.

[0023] According to a preferred embodiment of the present application, the clustering results are evaluated by the R-score.

[0024] According to an embodiment of the present application, as Figure 1As shown, the PMC-IGM model can be divided into three modules: 1) learning consistent embedding from multi-view similarity matrix through adaptive graph total variation; 2) learning consistent embedding from multi-view spectral embedding through graph construction on symmetric positive definite (SPD) manifold; 3) multi-objective learning with adaptive weights: through parameter-free iterative optimization formula. From the above three modules, consistent embedding can be effectively and concisely learned from multi-view similarity matrix and spectral embedding, and clustering results can be obtained by performing K-means on the learned embedding.

[0025] In the present application, an exemplary clustering result graph drawn based on GHSV is provided as Figure 2 As shown, it can be clearly divided, so we can know that the model gives GHSV a relatively high weight parameter. We use the classic random forest model to perform supervised learning on the database and obtain the ranking of feature importance affecting the performance of the catalyst. The feature importance result graph is shown as Figure 3 As shown, the importance of GHSV is much higher than that of other input features, which further proves the reliability of the clustering model.

[0026] According to the MWED descriptor, the pre-set copper-nickel bimetallic catalyst and the non-active metal catalyst were converted, and a total of 8 catalysts with different reaction temperatures were constructed. Each catalyst has three views as the input of the PMC-IGM model. The prediction results are shown in Table 1. It can be seen that at temperatures of 300, 350 and 400°C, the copper-nickel bimetallic catalyst is clustered into the first category, which means that the hydrogen production rate is general and the carbon monoxide selectivity is high; at 450°C, it is divided into the third category, which has a high hydrogen production rate and a low carbon monoxide selectivity. The non-active metal catalyst is clustered in the second category regardless of the temperature, and has almost no catalytic activity.

[0027] Table 1

[0028] Figure 4-1 、 Figure 4-2 The XRD patterns of the selected catalysts after calcination and after reaction are shown respectively. We found that there were no characteristic peaks of Al2O3 in either stage of the XRD pattern, which may be due to its good dispersion, being covered by other metal oxides or forming in amorphous phase. However, Cu, Ni and Ca elements always exist, proving that we have successfully synthesized the two pre-set catalysts. The reason why the non-active metal catalyst shows the characteristic peak of C after reaction is that it has no catalytic activity, resulting in the formation of a large amount of mixed carbon deposits in the catalyst.

[0029] Figure 5The effect of temperature on various properties of the two catalysts (methanol conversion, H2 yield, H2 and CO selectivity, and hydrogen production rate) is shown. It can be seen that the performance of the copper-nickel bimetallic catalyst also continuously improves as the temperature increases, and when the temperature reaches 450°C, the carbon monoxide selectivity significantly decreases. This result is consistent with the model prediction, which has a high hydrogen rate and low carbon monoxide selectivity at 450°C. In contrast, the performance of the inactive metal catalyst is not affected by temperature and remains catalytically inactive. Although this catalyst is not catalytically active, the methanol conversion and H2 yield are both greater than 0 due to the cracking of methanol at high temperatures, and the hydrogen selectivity is close to 100%. This is also the main reason for the appearance of C characteristic peaks in the XRD pattern of the reactive inactive metal catalyst.

[0030] Figure 6 The actual cluster data of the catalysts are shown. Compared with the results in Table 1, we can see that the actual results are basically consistent with the predicted results, proving the reliability of the algorithm and achieving the expected goal of catalyst screening. This allows us to refer to the results of the model in subsequent catalyst synthesis to determine whether the catalyst is worth preparing, which can save a lot of research time.

[0031] The above describes the preferred embodiments of the present application in detail, but the present application is not limited thereto. Within the technical concept of the present application, various simple modifications can be made to the technical solutions of the present application, including the combination of various technical features in any other suitable manner, and these simple modifications and combinations should also be considered as disclosed by the present application and fall within the protection scope of the present application.

Claims

1. A multi-view machine learning assisted alcohol reforming catalyst screening method, characterized in that, The method comprises: S1, obtaining data to construct a first data set, including the composition of the alcohol reforming catalyst, alcohol reforming conditions and catalyst performance evaluation results; S2, using a multi-level weighted element descriptor to weight the composition data of the alcohol reforming catalyst in the first data set to construct a second data set; S3, inputting the second data set and the first data set into a parameter-free multi-view clustering model for training to obtain a catalyst performance prediction model; S4, inputting the real-time collected catalyst composition and alcohol reforming conditions into the catalyst performance prediction model to obtain catalyst performance evaluation prediction indexes.

2. The screening method of claim 1, wherein, In step S1, the obtained data includes: catalyst element composition, alcohol reforming reaction temperature, alcohol reforming reaction pressure, alcohol reforming steam-carbon ratio, alcohol reforming oxygen-carbon ratio, gas hourly space velocity, catalyst mass, hydrogen production rate, and carbon monoxide selectivity.

3. The screening method according to claim 1 or 2, wherein, The alcohol is a C1-C6 monohydric alcohol.

4. The screening method according to claim 1 or 2, wherein, The alcohol is methanol.

5. The screening method of claim 1 or 2, wherein, The multi-level weighted element descriptor includes electron cloud structure, microscopic level features and macro level features, wherein the electron cloud structure includes five element descriptors: AN, AM, NOE, NON, NOP and P; The microscopic level features include eight element descriptors: AR, CR, IR, AV, EGAR, EGP, FIP and EWF; The macro level features include five element descriptors: D, AM, HOF, MP and BP.

6. The screening method of claim 1 or 2, wherein, The catalyst performance prediction model obtains clustering results through hydrogen production rate and carbon monoxide selectivity.

7. The screening method of claim 1 or 2, wherein, The catalyst performance prediction model includes one or more of: a clustering result graph plotted against volume space velocity, a catalyst element composition clustering result graph, an alcohol reforming reaction temperature clustering result graph, an alcohol reforming reaction pressure clustering result graph, and an alcohol reforming steam-carbon ratio clustering result graph.

8. The screening method of claim 1 or 2, wherein, The clustering results are evaluated by one or more of: accuracy, normalized mutual information, F-score, R-score, adjusted Rand index, mean squared error, mean absolute error, and determination coefficient.

9. The screening method according to claim 1 or 2, wherein, The clustering results are evaluated by R-score.

Images