Rapid catalyst performance prediction method based on XRD (X-Ray Diffraction) data

Abstract

The invention provides a catalyst performance rapid prediction method based on XRD data. The method comprises the steps of data acquisition and preprocessing, feature extraction, prediction model training and unknown sample prediction. According to the method, the ORR performance can be predicted only through XRD testing, the traditional evaluation period of several weeks is shortened to the hour level, and consumption of precious metal and time can be greatly reduced; through multi-dimensional and high-coverage XRD feature engineering and in combination with an extreme random tree algorithm with high over-fitting resistance and a small sample optimization strategy, high-precision performance prediction can be performed on a catalyst with unknown components; the method does not depend on known chemical components of the catalyst, predicts only based on crystal structure information of the catalyst, and is particularly suitable for synthetic samples with complex or unknown components; a standardized process is formed from data preprocessing, feature extraction to model prediction, integration and automation are easy, and a powerful tool is provided for high-flux catalyst screening.

Description

Technical Field

[0001] This invention belongs to the field of electrochemical catalyst performance prediction technology, specifically relating to a rapid method for predicting catalyst performance based on XRD data. Background Technology

[0002] The oxygen reduction reaction (ORR) is a core cathode process in sustainable energy devices such as fuel cells and metal-air batteries. Its slow kinetics result in high usage of precious metal catalysts, especially platinum-based materials. Global annual demand for platinum group metals (PGMs) for fuel cell vehicles and distributed power generation systems has exceeded 120 tons, with approximately 60% used in ORR catalyst layers. However, limited platinum reserves and volatile prices (averaging approximately $1050 / ounce in 2024) severely restrict the large-scale promotion of the hydrogen economy.

[0003] To reduce platinum usage, researchers commonly employ a platinum-transition metal / carbon (Pt-M / C, where M = Co, Ni, Fe, etc.) alloying strategy, enhancing intrinsic activity by controlling electronic structure, lattice strain, and surface defects. However, Pt-M / C powders obtained through liquid-phase synthesis often suffer from unknown composition and highly heterogeneous microstructure, leading to a traditional "trial-and-error" experimental process—powder preparation → thin film coating → electrochemical testing—taking 4-6 weeks per cycle and consuming large amounts of precious metals.

[0004] In recent years, the integration of machine learning and high-throughput experiments has provided new insights for catalyst development. However, existing studies largely rely on density functional theory (DFT) calculations or elemental descriptors, making it difficult to directly correlate the crystallographic characteristics and macroscopic catalytic performance of actual synthesized powders. Furthermore, the scarcity of strictly paired XRD-ORR datasets makes deep learning models highly susceptible to overfitting. Therefore, under "small sample" conditions, how to rapidly and accurately predict the ORR activity of unknown Pt-M / C components using only readily available experimental XRD patterns remains a critical technical bottleneck that urgently needs to be overcome in this field. Summary of the Invention

[0005] (a) Technical problems to be solved This invention proposes a rapid prediction method for catalyst performance based on XRD data to solve the technical problems of long cycle, high cost, low efficiency, and lack of rapid performance prediction means for unknown component systems in traditional ORR catalyst trial and error methods.

[0006] (II) Technical Solution To address the aforementioned technical problems, this invention proposes a rapid prediction method for catalyst performance based on XRD data. This rapid prediction method for catalyst performance includes the following steps: S1. Data Acquisition and Preprocessing Obtain the original XRD pattern of the catalyst to be tested, and preprocess the XRD pattern; S2. Feature Extraction Multiple features, including global statistical features, main peak analysis features, secondary peak identification features, ratio features and interaction features, are extracted from the preprocessed XRD pattern to form a high-dimensional feature vector. S3. Prediction Model Training Using strictly paired XRD features and ORR half-wave potential data as the training set, the extreme random tree regression algorithm was used for training, and the hyperparameters were optimized through nested cross-validation to obtain the final half-wave potential prediction model. S4. Prediction of Unknown Samples For the unknown component catalyst to be tested, steps S1 and S2 are performed to extract features and input them into the half-wave potential prediction model trained in step S3, and the predicted ORR half-wave potential value is output.

[0007] Furthermore, in step S1, the catalyst to be tested is Pt-M / C catalyst powder.

[0008] Furthermore, in step S1, the preprocessing of the XRD pattern includes smoothing and noise reduction using a Savitzky-Golay filter; and baseline correction using an adaptive iterative weighted penalized least squares algorithm.

[0009] Furthermore, in step S1, during baseline correction, the penalty factor λ is dynamically adjusted to a value range of 10. 2 ~10 6 This is to control the background subtraction peak intensity error to within 1%.

[0010] Further, in step S2, the global statistical features include the mean, standard deviation, skewness, kurtosis, and integral area of ​​the diffraction intensity; the main peak analytical features include the peak position, intensity, full width at half maximum (FWHM), and symmetry of the main peak; the secondary peak identification features include the peak position, intensity, and width of the first three secondary peaks; the ratio features include calculating the peak position ratio and intensity ratio between the secondary peak and the main peak; and the interaction features are second-order polynomial interaction features generated based on the global statistical features, the main peak analytical features, the secondary peak identification features, and the ratio features, which together constitute a high-dimensional feature set.

[0011] Furthermore, in step S2, the parameters of each peak are accurately extracted through Gaussian fitting, and the final high-dimensional feature set has a dimension of 217.

[0012] Furthermore, in step S3, the hyperparameter optimization space of the extreme random tree regression algorithm includes the number of trees, maximum depth, minimum number of split samples, minimum number of leaf node samples, and maximum feature ratio.

[0013] Furthermore, in step S3, the training process of the prediction model includes optimizing the model hyperparameters using a nested cross-validation strategy, wherein the outer cross-validation is used to evaluate the model's generalization performance, and the inner cross-validation is used to perform grid search to determine the optimal combination of hyperparameters.

[0014] Furthermore, in step S3, nested cross-validation consists of an outer 5-fold cross-validation layer for generalization evaluation and an inner 3-fold cross-validation layer for grid search, with the validation set's coefficient of determination R... 2 To optimize the objective.

[0015] (III) Beneficial Effects This invention proposes a rapid catalyst performance prediction method based on XRD data, including data acquisition and preprocessing, feature extraction, prediction model training, and prediction for unknown samples. This method only requires XRD testing to predict ORR performance, shortening the traditional evaluation cycle of several weeks to hours, significantly reducing the consumption of precious metals and time. Through multi-dimensional, high-coverage XRD feature engineering, combined with an extreme random tree algorithm with strong anti-overfitting capabilities and a small-sample optimization strategy, it can achieve high-precision performance prediction for catalysts with unknown components. This method does not rely on the known chemical composition of the catalyst, but only on its crystal structure information for prediction, making it particularly suitable for synthetic samples with complex or unknown compositions. From data preprocessing and feature extraction to model prediction, a standardized process is formed, which is easy to integrate and automate, providing a powerful tool for high-throughput catalyst screening. Attached Figure Description

[0016] Figure 1a Automatic preprocessing results for XRD data with baseline correction via Savitzky-Golay filter and improved airPLS algorithm; Figure 1b To obtain the position, full width at half maximum (FWHM), and intensity data of each peak in the XRD data; Figure 2a This is a scatter plot showing the correlation between experimental values ​​and model predictions for the ExtraTreesRegressor model. Figure 2b This is a residual distribution chart showing the difference between experimental values ​​and model predictions. Figure 3a The cumulative importance curves, used for in-depth mining of feature interactions and statistical properties, show the number of features required to reach 80% and 90% contribution, and include annotations of key statistical indicators. Figure 3b The graph showing the contribution ratios of six types of features reveals that intensity parameters and feature interactions are the core factors in predicting catalyst performance. Detailed Implementation

[0017] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0018] This embodiment proposes a method for rapid prediction of catalyst performance based on XRD data, which specifically includes the following steps: First, standardized XRD tests were performed on the Pt-M / C powder to be tested. A CuKα radiation source was used, with continuous scanning in 0.02° steps within a 2θ range of 10°–80°. The tube voltage was 40 kV, the tube current was 30 mA, and the integration time was 0.5 s / step to ensure sufficient signal-to-noise ratio and angular resolution in the diffraction pattern. After scanning, the raw data was saved in a two-column format (2θ angle-intensity) to provide a unified input for subsequent preprocessing.

[0019] Subsequently, the raw XRD data underwent three preprocessing steps. First, a Savitzky-Golay filter was used for smoothing and noise reduction, with a window length of 21 points and a polynomial order of 3, to remove high-frequency noise while preserving peak shape information. Second, an improved adaptive iterative weighted penalized least squares (airPLS) algorithm was used to subtract baseline drift, dynamically adjusting the penalty factor λ∈

[10] . 2 10 6 The first step involves controlling the background subtraction error to within 1%. The second step involves automatically identifying all diffraction peaks based on a dual-threshold peak detection strategy and performing Gaussian fitting on each peak to accurately extract parameters such as peak position, full width at half maximum (FWHM), integral intensity, and Lorentz ratio, forming a 217-dimensional original feature pool. Figure 1a Automatic preprocessing results for XRD data with baseline correction via Savitzky-Golay filter and improved airPLS algorithm; Figure 1b To obtain the position, half-peak width, and intensity data of each peak in the XRD data.

[0020] In the feature engineering stage, the 217-dimensional original features are further constructed into five types of complementary descriptors: the first type is global statistical features, including the mean, standard deviation, skewness, kurtosis and cumulative integral area of ​​the whole spectrum intensity, used to characterize the overall crystallinity; the second type is the main peak analytical features, which lock the peak position, intensity, FWHM and symmetry of the Pt(111) crystal plane in the 39–41° range, and directly correlate with the lattice strain; the third type is secondary peak identification features, which locate the first three secondary peaks through an automatic peak finding algorithm, record their 2θ position, relative intensity and FWHM, reflecting the secondary phase content; the fourth type is ratio features, which calculate the peak position ratio and intensity ratio of the secondary peak to the main peak, used to quantify the phase purity; the fifth type is second-order polynomial interaction features, which combine the above features in pairs to expand the feature space to 217 dimensions and capture nonlinear coupling effects.

[0021] For small sample scenarios, ExtraTreesRegressor is chosen as the core algorithm. This algorithm uses an extremely random splitting strategy to randomly select feature thresholds when splitting nodes, effectively reducing variance and avoiding overfitting under small sample conditions. The hyperparameter space is set as follows: The number of trees, n_estimators∈{100,200,300}, The maximum depth is max_depth ∈ {None, 10, 20, 30}. The minimum number of split samples is min_samples_split∈{2,5,10}. The minimum number of leaf node samples is min_samples_leaf∈{1,2,4}. The maximum feature ratio is max_features∈{'sqrt','log2',0.5,0.8}.

[0022] A nested cross-validation strategy is employed: the outer 5-fold stratified sampling ensures a consistent ratio of high and low activity samples, while the inner 3-fold grid search uses the validation set's coefficient of determination R0. 2 To optimize the objective, the optimal combination of hyperparameters was ultimately determined.

[0023] After obtaining the optimal hyperparameters, ExtraTreesRegressor is retrained using all training data and saved as a binary model file. For Pt-M / C catalysts with unknown components, only the above XRD testing and feature extraction process needs to be repeated. By inputting the 217-dimensional feature vector into the trained model, the predicted ORR half-wave potential can be output in seconds. Experimental results show that this method achieves R0 on strictly paired datasets. 2 =0.9882, MAE=0.0083V; Blind testing was performed on 10 independent unknown samples, and the mean absolute error was still less than 0.05V, achieving high-precision and rapid mapping from XRD patterns to ORR half-wave potential.

[0024] Figure 2a This is a scatter plot showing the correlation between experimental values ​​and model predictions for the ExtraTreesRegressor model. Figure 2b This is a residual distribution chart showing the difference between experimental values ​​and model predictions. Figure 3a The cumulative importance curves, used for in-depth mining of feature interactions and statistical properties, show the number of features required to reach 80% and 90% contribution, and include annotations of key statistical indicators. Figure 3b The graph showing the contribution ratios of six types of features reveals that intensity parameters and feature interactions are the core factors in predicting catalyst performance.

[0025] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for rapid prediction of catalyst performance based on XRD data, characterized in that, The method for rapid prediction of catalyst performance includes the following steps: S1. Data Acquisition and Preprocessing Obtain the original XRD pattern of the catalyst to be tested, and preprocess the XRD pattern; S2. Feature Extraction Multiple features, including global statistical features, main peak analysis features, secondary peak identification features, ratio features and interaction features, are extracted from the preprocessed XRD pattern to form a high-dimensional feature vector. S3. Prediction Model Training Using strictly paired XRD features and ORR half-wave potential data as the training set, the extreme random tree regression algorithm was used for training, and the hyperparameters were optimized through nested cross-validation to obtain the final half-wave potential prediction model. S4. Prediction of Unknown Samples For the unknown component catalyst to be tested, steps S1 and S2 are performed to extract features and input them into the half-wave potential prediction model trained in step S3, and the predicted ORR half-wave potential value is output.

2. The method for rapid prediction of catalyst performance based on XRD data as described in claim 1, characterized in that, In step S1, the catalyst to be tested is Pt-M / C catalyst powder.

3. The method for rapid prediction of catalyst performance based on XRD data as described in claim 1, characterized in that, In step S1, the preprocessing of the XRD pattern includes smoothing and noise reduction using a Savitzky-Golay filter, and baseline correction using an adaptive iterative weighted penalized least squares algorithm.

4. The method for rapid prediction of catalyst performance based on XRD data as described in claim 3, characterized in that, In step S1, during baseline correction, the penalty factor λ is dynamically adjusted to a value range of 10. 2 ~10 6 This is to control the background subtraction peak intensity error to within 1%.

5. The method for rapid prediction of catalyst performance based on XRD data as described in claim 1, characterized in that, In step S2, the global statistical features include the mean, standard deviation, skewness, kurtosis, and integral area of ​​the diffraction intensity; the main peak analytical features include the peak position, intensity, full width at half maximum (FWHM), and symmetry of the main peak; and the secondary peak identification features include the peak position, intensity, and width of the first three secondary peaks. The ratio feature includes calculating the peak position ratio and intensity ratio between the secondary peak and the primary peak; the interaction feature is a second-order polynomial interaction feature generated based on global statistical features, primary peak analysis features, secondary peak identification features, and ratio features, which together constitute a high-dimensional feature set.

6. The method for rapid prediction of catalyst performance based on XRD data as described in claim 1, characterized in that, In step S2, the parameters of each peak are accurately extracted through Gaussian fitting, and the final high-dimensional feature set has a dimension of 217.

7. The method for rapid prediction of catalyst performance based on XRD data as described in claim 1, characterized in that, In step S3, the hyperparameter optimization space of the extreme random tree regression algorithm includes the number of trees, maximum depth, minimum number of split samples, minimum number of leaf node samples, and maximum feature ratio.

8. The method for rapid prediction of catalyst performance based on XRD data as described in claim 1, characterized in that, In step S3, the training process of the prediction model includes optimizing the model hyperparameters using a nested cross-validation strategy, where the outer cross-validation is used to evaluate the model's generalization performance and the inner cross-validation is used to perform grid search to determine the optimal combination of hyperparameters.

9. The method for rapid prediction of catalyst performance based on XRD data as described in claim 8, characterized in that, In step S3, nested cross-validation consists of an outer 5-fold cross-validation layer for generalization evaluation and an inner 3-fold cross-validation layer for grid search, with the validation set's coefficient of determination R0 as the benchmark. 2 To optimize the objective.

Images