Classification and analysis method of sweet spots

Abstract

A method for classifying and analyzing sweet spots is provided. The method for classifying and analyzing sweet spots includes the step of classifying sweet spots of an analysis target using a clustering algorithm and the step of analyzing sweet spots using explainable artificial intelligence.

Description

Technology Field

[0001] The present invention relates to a sweet spot classification and analysis method. Background Technology

[0002] Although there are prior studies analyzing the influence of factors within the sweet spot, these studies are characterized by the fact that their explanations are valid only for points near the point intended to be explained. Consequently, while the analysis of points with similar correlations between factors is valid, inferences derived from the analysis of the entire dataset lack validity, leading to a decrease in the reliability of the analysis. The problem to be solved

[0003] The present invention provides a sweet spot classification and analysis method having excellent performance.

[0004] Other objects of the present invention will become clear from the following detailed description and the accompanying drawings. means of solving the problem

[0005] A method for classifying and analyzing sweet spots according to embodiments of the present invention includes the step of classifying sweet spots of an analysis target using a clustering algorithm and the step of analyzing sweet spots using explainable artificial intelligence. Effects of the invention

[0006] The sweet spot classification and analysis method according to the embodiments of the present invention can have excellent performance. The reliability of the analysis can be improved by combining a clustering algorithm with explainable artificial intelligence, and it is possible to analyze how the causal relationship between factors and target attributes differs for each sweet spot. Brief explanation of the drawing

[0007] Figure 1 shows a graph displaying optimal catalyst data in a two-dimensional space using UMAP. Figure 2 shows a graph in which each cluster is generated by applying HDBSCAN to the graph of Figure 1. Figure 3 shows a pairplot representing the distributions of representative factors for each cluster. Figure 4 shows a Random Forest regression model generated in an embodiment of the present invention. Figure 5 shows the classification of SHAP descriptions according to an embodiment of the present invention. Specific details for implementing the invention

[0008] The present invention will be described in detail below through examples. The purpose, features, and advantages of the present invention will be easily understood through the following examples. The present invention is not limited to the examples described herein and may be embodied in other forms. The examples introduced herein are provided to ensure that the disclosed content is thorough and complete, and to ensure that the concept of the present invention is sufficiently conveyed to those skilled in the art to which the present invention pertains. Accordingly, the present invention should not be limited by the following examples.

[0010] A method for classifying and analyzing sweet spots according to embodiments of the present invention includes the step of classifying sweet spots of an analysis target using a clustering algorithm and the step of analyzing sweet spots using explainable artificial intelligence.

[0011] The above clustering algorithm may include HDBSCAN (Hierarchical Density-based Spatial Clustering of Applications with Noise), and the above explainable artificial intelligence may include SHAP (Shapley Additive exPlanations).

[0012] The step of classifying the sweet spot may include a step of representing the data of the analysis target in a low-dimensional space based on the distance between data using UMAP (Uniform Manifold Approximation and Projection for dimensionality reduction), and a step of performing clustering in the low-dimensional space using HDBSCAN.

[0013] The step of classifying the sweet spot above may include the step of filtering the entire data of the analysis target to extract optimal data having a range of target attribute values.

[0014] The step of analyzing the sweet spot may include the step of generating a model to be analyzed by the SHAP, and the model may include a random forest regression model.

[0015] The step of analyzing the sweet spot may include the step of drawing a decision plot for the sweet spot. The part of the decision plot where the model prediction value changes the most may represent the key factor that enables the sweet spot to have optimal target attributes.

[0016] The step of analyzing the sweet spot may further include a step of comparing the decision plot with a pair plot obtained in the step of classifying the sweet spot, and factors affecting the sweet spot may be identified by comparing the decision plot and the pair plot.

[0018] [Dimensionality Reduction Techniques and UMAP (Uniform Manifold Approximation and Projection for Dimensionality Reduction)]

[0019] Dimensionality Reduction (DR) is a technique that reduces a high-dimensional data space to a low-dimensional space, which can improve the predictive performance of machine learning models and support more intuitive clustering algorithms by displaying data in 2D or 3D space.

[0020] UMAP (Uniform Manifold Approximation and Projection for Dimensionality Reduction) is a dimensionality reduction technique that represents data in a lower dimension based on topological data analysis. The UMAP algorithm first assigns a unique distance function to each point in the data to form simple complexes that are evenly distributed; based on this, it generates a graph representing the connectivity between points and then creates a new representation in a lower dimension. During this process, the new data space is constructed to preserve a topological structure similar to the original data by optimizing the cross-entropy between points in the original data space and the new lower-dimensional space.

[0022] [Clustering Algorithms and HDBSCAN (Hierarchical Density-based Spatial Clustering of Applications with Noise)]

[0023] Clustering algorithms are a type of unsupervised learning technique that forms clusters by automatically grouping similar data within a feature space.

[0024] HDBSCAN is a type of hierarchical clustering that specifies mutual reachability distances within a transformed data space, constructs cluster hierarchies based on these distances, and forms clusters.

[0026] [Explainable Artificial Intelligence (XAI) and SHAP (Shapley Additive exPlanations)]

[0027] Traditional machine learning corresponds to black-box models, which can only verify inputs and corresponding outputs but cannot obtain information regarding which factors led to those output values. This means that even if a model is obtained that accurately reflects real-world phenomena, it fails to provide insight into the causality of the results. However, explainable AI analyzes the causal relationship between the input and the results derived from already trained black-box models.

[0028] SHAP is an XAI technique based on cooperative game theory that can quantitatively analyze the influence of factors on results from the output of a trained black-box model.

[0030] [Filtering Analysis Targets]

[0031] Catalysts with optimal characteristics are filtered. The entire dataset (D) consists of a set of target property values ​​(y) and a set of various factors (X) that influence y. To classify various sweet spots and identify the key factors of each sweet spot, only data with an ideal range of target property values ​​is filtered from the entire dataset to create a dataset (D') consisting only of data for catalysts with optimal performance. The dataset (D') also consists of an optimal set of target properties (y') and a set of various factors (X').

[0033] [Identification of Clusters Using Dimensionality Reduction Techniques and Clustering Algorithms]

[0034] Unsupervised UMAP is applied, and clusters are identified using HDBSCAN (Hierarchical Density-based Spatial Clustering of Applications with Noise). Machine learning can be broadly classified into supervised learning, which includes y values ​​in the learning process; unsupervised learning, which proceeds with learning using only factors without including them; and reinforcement learning, which proceeds with learning using reward values ​​based on specific actions. UMAP is capable of both supervised and unsupervised learning; however, in this invention, since optimal catalyst data has already been filtered out through a filtering process, it is important to focus more on the factors themselves rather than the target property, so an unsupervised form of UMAP is performed. After representing X' data in a 2D space using UMAP, each cluster is identified through the HDBSCAN algorithm, which performs clustering based on the distance between data points.

[0036] [Create Model]

[0037] A model to be analyzed by SHAP is created. The SHAP used in the present invention is a TreeExplainer, which is used to analyze ensemble models such as Random Forest or XGBoost. Therefore, in the present invention, a Random Forest regression model to be analyzed by SHAP is created. This model is trained to predict y from X, and training is performed using the entire dataset (D) before filtering.

[0039] [Analysis using Explainable Artificial Intelligence (SHAP Analysis)]

[0040] The SHAP model is trained on the entire dataset (D), which includes both catalysts with optimal performance and those without. Subsequently, using the trained model, SHAP analysis is performed only on the data (D') with optimal performance, and a decision plot is drawn. The decision plot is a graph that aggregates the influence of each factor on a target value, allowing for a comprehensive view of the factors' effects across all target data. If the data generates various types of sweet spots even with similar target values, lines of various shapes may be observed in the decision plot. This can be interpreted as being due to diverse interactions between factors for each sweet spot. In the cluster verification process mentioned above By using UMAP and HDBSCAN to identify sweet spots composed of catalysts with similar relationships among various non-linear associations and redrawing the decision plots for each spot, a graph consisting of relatively consistent lines can be derived. Additionally, by identifying the factor that causes the result value to change most abruptly in each decision plot, the key factors that enable each sweet spot to have the optimal target attribute can be identified.

[0042] [Comparison of SHAP Analysis Results and Feature Distributions for Each Sweet Spot]

[0043] By comparing the pairplot obtained during the cluster verification process using dimensionality reduction techniques and clustering algorithms with the decision plot obtained during the aforementioned SHAP analysis process, it is possible to compare the influence on target attributes based on the distribution of factor values ​​for each sweet spot. Furthermore, since SHAP explanations may vary slightly even within a sweet spot, a waterfall plot—which analyzes the SHAP explanation for each point within the sweet spot—can be used to analyze which factor values ​​created the differences within the clusters.

[0045] [Example]

[0046] Target system

[0047] This study deals with alloy catalysts related to the chemical adsorption of carbon dioxide (CO2). For reactions that increase the optimal yield and selectivity of *COCOH, the CO adsorption energy to the catalyst must be between -0.65 eV and -0.5 eV. The CO adsorption energy (target property: y) is influenced by a total of 13 factors (X), including the d-band of the transition metal and the characteristics of the alloy. In this example, various sweet spots of catalyst data with optimal y values ​​are classified, and SHAP is used to investigate how much the 13 factors influence the CO adsorption energy. The entire dataset consists of data regarding the characteristics of a total of 298 different alloy catalysts.

[0048]

[0050] Catalytic filtering with optimal characteristics

[0051] A new dataset (D', 18) was created from the total dataset (D, 298) consisting only of data with CO adsorption energies (y) of -0.65 eV or higher and -0.5 eV or lower.

[0053] Application of unsupervised UMAP and cluster verification using HDBSCAN

[0054] As a result of performing UMAP and HDBSCAN on 18 filtered catalyst data, the following three clusters were identified, and the distribution of factors for each cluster could be confirmed. Figure 1 shows a graph displaying optimal catalyst data in a two-dimensional space using UMAP, Figure 2 shows a graph generating each cluster by applying HDBSCAN to the graph in Figure 1, and Figure 3 shows a pairplot representing the distribution of representative factors for each cluster. Based on the analysis of the factors in each cluster, it can be inferred that Group 0 consists of Pd or Au transition metal-based alloy catalysts, while Groups 1 and 2 consist of Cu transition metal-based alloy catalysts.

[0056] Creating a model to be analyzed by SHAP

[0057] A Random Forest regression model was created to predict y values ​​from the factor values ​​(X) of the entire data, and as shown in Figure 4, a model with high prediction performance was obtained with a Root Mean Square Error (RMSE) value of 0.078.

[0059] SHAP Analysis

[0060] Figure 5 illustrates the classification of SHAP descriptions according to an embodiment of the present invention. Referring to Figure 5, if SHAP analysis is performed directly on X' without a clustering process, lines with various shapes are observed, as in the raw decision plot. From this, it can be seen that even within X', which possesses the optimal target property, there are points where the interactions between factors differ. When the SHAP analysis of each sweet spot obtained through the application of UMAP and clustering using HDBSCAN is plotted on a decision plot, it can be classified into lines of relatively consistent shape. By identifying the part where the model prediction value changes the most in each decision plot (shaded area in the classified decision plot), the factor that contributed most significantly to the high CO adsorption energy for each sweet spot can be identified. When correlated with the application of UMAP and the interpretation of clustering using HDBSCAN, it can be inferred that for Pd or Au-based catalysts, d-band skewness has the greatest influence on the CO adsorption energy of the catalyst, while for Cu-based alloy catalysts, d-band filling has the greatest influence on the CO adsorption energy of the catalyst. Even within each sweet spot, the predicted CO adsorption energy varies depending on minute changes in factor values, and the SHAP analysis of each point within the sweet spot can be verified through a waterfall plot.

[0062] Comparison of SHAP analysis results and feature distributions for each sweet spot

[0063] When comparing the pair plot obtained during the clustering process using UMAP and HDBSCAN with the decision plot obtained during the SHAP analysis process, it can be predicted that for Pd or Au-based group 0, a moderate skewness that is neither too low nor too high has a major influence on the high adsorption energy, while for Cu-based groups 1 and 2, a high filling value has a major influence on the optimal CO adsorption energy.

[0065] Specific embodiments of the present invention have been examined so far. Those skilled in the art will understand that the present invention may be embodied in modified forms without departing from the essential characteristics of the invention. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. The scope of the present invention is defined by the claims, not by the foregoing description, and all variations within the scope of equivalents should be interpreted as being included in the present invention.

Claims

Claim 1 A method for classifying and analyzing sweet spots performed by a computing device, comprising: a step of classifying sweet spots of an analysis target using a clustering algorithm; and a step of analyzing sweet spots using explainable artificial intelligence, wherein the clustering algorithm includes HDBSCAN (Hierarchical Density-based Spatial Clustering of Applications with Noise) and the explainable artificial intelligence includes SHAP (Shapley Additive exPlanations), wherein the step of classifying sweet spots includes a step of filtering all data of the analysis target to extract optimal data having a range of target attribute values, and the step of analyzing sweet spots includes a step of drawing a decision plot for the sweet spots, wherein the part of the decision plot where the model prediction value changes the most represents a key factor that causes the sweet spot to have optimal target attributes. Claim 2 delete Claim 3 A method for classifying and analyzing sweet spots according to claim 1, wherein the step of classifying the sweet spot comprises: a step of representing the data of the analysis target in a low-dimensional space based on the distance between data using UMAP (Uniform Manifold Approximation and Projection for dimensionality reduction); and a step of performing clustering in the low-dimensional space using HDBSCAN. Claim 4 delete Claim 5 A method for classifying and analyzing sweet spots according to claim 1, wherein the step of analyzing the sweet spot includes the step of generating a model to be analyzed by the SHAP, and the model includes a random forest regression model. Claim 6 delete Claim 7 delete Claim 8 A method for classifying and analyzing a sweet spot according to claim 1, wherein the step of analyzing the sweet spot further includes the step of comparing the decision plot with a pair plot obtained in the step of classifying the sweet spot, and wherein a factor affecting the sweet spot is identified by the comparison of the decision plot and the pair plot.

Images