A high-dimensional data-driven, artificial intelligence-based approach to next-generation catalyst design

Abstract

The present invention discloses a high-dimensional data-driven, artificial intelligence-based next-generation catalyst design method with the following innovative components: (1) Multidimensional data fusion module: Uses novel data processing technology to integrate various experimental and computational data. (2) Adaptive neural network architecture: Dynamically adjusts the network structure according to the performance requirements of the catalyst to optimize the selectivity and activity of the catalyst. (3) Knowledge-driven feature engineering module: Combines chemical knowledge graphs to automatically extract features that have the greatest impact on catalytic performance. (4) Iterative feedback optimization module: Uses online learning mechanisms to adjust parameters, digital twin technology to simulate behavior, and machine learning prediction models and sensor data fusion to monitor status. (5) Interactive visualization and decision support module: Provides advanced visualization tools and AI decision support to help R&D personnel understand data and make scientific decisions.

Description

Technical Field

[0001] The present invention relates to the fields of deep learning, computer vision, big data analysis, feature extraction and pattern recognition, and in particular to a high-dimensional data-driven, artificial intelligence-based next-generation catalyst design method. Background Art

[0002] Catalysts play a vital role in chemical reactions, and their performance directly affects the speed, selectivity, and purity of the reaction products. Traditional catalyst design methods mainly rely on repeated laboratory experiments and the professional experience of researchers, which is not only time-consuming and labor-intensive, but also difficult to achieve large-scale optimization and innovation in a short period of time. With the rapid development of deep learning and artificial intelligence technology, modern catalyst design is gradually moving towards automation and intelligence.

[0003] In existing research and applications, although some computer vision and machine learning technologies have begun to be used in catalyst design optimization, these methods are mostly limited to basic data processing and analysis, such as single variable optimization or simple feature extraction. For complex catalyst systems, these methods are difficult to efficiently process multidimensional data and conduct comprehensive performance evaluation and optimization. In addition, traditional methods often require a lot of manual intervention when processing large-scale experimental data, which greatly limits their promotion and popularization in practical applications. Summary of the invention

[0004] The present invention solves the problem that traditional methods mainly rely on the experience and repeated experiments of researchers, which is not only time-consuming and labor-intensive, but also easily affected by subjective judgment, and cannot provide quantitative and accurate analytical information about changes in catalyst performance. In addition, when dealing with complex and changeable catalyst characteristics, a large amount of manual intervention is often required, which limits its technical problem of widespread promotion in practical applications; the present invention provides a high-dimensional data-driven next-generation catalyst design method based on artificial intelligence, which utilizes multidimensional data fusion and deep learning analysis to achieve comprehensive optimization and dynamic adjustment of catalyst design and performance, and has significant advantages.

[0005] A high-dimensional data-driven, AI-based next-generation catalyst design platform, specifically including the following components:

[0006] Components (1) Multidimensional data fusion module: using data processing techniques to integrate experimental and computational data;

[0007] Component (2) Adaptive neural network architecture: Dynamically adjust the network structure according to the performance requirements of the catalyst to optimize the selectivity and activity of the catalyst;

[0008] Component (3) Knowledge-driven feature engineering module: Combined with chemical knowledge graph, it automatically extracts the features that have the greatest impact on catalytic performance;

[0009] Component (4) Iterative feedback optimization module: using online learning mechanism to adjust parameters, digital twin technology to simulate behavior, machine learning prediction model and sensor data fusion to monitor status;

[0010] Component (5) Interactive visualization and decision support module: Provides visualization tools and AI decision support to help R&D personnel understand data and make scientific decisions.

[0011] A high-dimensional data-driven, artificial intelligence-based next-generation catalyst design method comprises the following steps:

[0012] Step 1: Preprocess the multidimensional data, reduce the dimensionality of the high-dimensional data consisting of catalyst test data and computational simulation data, retain key features, and generate a simplified data set suitable for subsequent analysis; the processed data will be used as input for subsequent network architecture optimization and feature extraction.

[0013] Step 2: Using multi-level neural architecture search (NAS) and meta-learning techniques, based on the simplified data set obtained in step 1, the network architecture is automatically customized and optimized for the catalyst design task, and key performance features are extracted; the extracted features will provide a basis for in-depth analysis of the catalyst;

[0014] Step 3: Use chemical knowledge graphs and machine learning methods to conduct in-depth analysis of the key performance features extracted in step 2, and further explore the potential features and associations of the catalyst; this analysis provides data support for subsequent iterative optimization;

[0015] Step 4: Based on the analysis results of step 3, iterative feedback optimization is used to further adjust and optimize the catalyst design data; the optimized design will enter the next step of verification and discussion;

[0016] Step 5: An interactive user interface was developed to allow R&D personnel and engineers to view and discuss the catalyst design and analysis results optimized in Step 4, thereby achieving multi-party collaboration and decision support.

[0017] Preferably, in step 1, UMAP (Uniform Manifold Approximation and Projection) is used to perform dimensionality reduction processing on high-dimensional data, which can effectively map high-dimensional data to low-dimensional space while maintaining the global structure and local proximity relationship of the data;

[0018] in , is high-dimensional data, It is the data after dimensionality reduction.

[0019] Preferably, the step 2 specifically includes:

[0020] Step 21: Extract the key performance characteristics of the catalyst by integrating the convolutional neural network CNN, the recurrent neural network RNN ​​and the graph neural network GNN, wherein the graph neural network CNN is used to process the microstructure image of the catalyst and extract its surface morphology and structural characteristics; the recurrent neural network RNN ​​is used to analyze the time series data of the catalyst under different reaction conditions and predict its performance change trend; the graph neural network GNN is used to process the molecular structure data of the catalyst and capture the relationship between atoms and chemical bond characteristics; in, Indicates that the result of the convolution operation is at position The value of is the local weighted sum of the input image and the convolution kernel at that position. The convolution operation process is to traverse each pixel position of the input image, use the convolution kernel to perform weighted summation on the local area, and obtain the value of the corresponding position of the output feature map; is the convolution kernel, indicating that the convolution kernel is The value of the position; It is a double summation, which means that the summation operation is performed on each position of the convolution kernel; It indicates the position offset of the current convolution kernel in the input image. The pixel value after the reaction is calculated; the recurrent neural network (RNN) is used to analyze the time series data of the reaction conditions and predict the performance change trend;

[0021] in, yes The hidden state at the moment represents the hidden state of the time series; The weight matrix of the hidden state is used to update the hidden state, The input weight matrix is ​​used to update the input data; is the current time step The input data, is the bias term; is the activation function; the graph neural network GNN is used to process molecular structure data to capture the relationship between atoms and chemical bond characteristics;

[0022] ;in ,yes The feature vector of node i in the layer indicates the representation of the node in the layer; It is Node in layer The characteristic vector of It is The weight matrix of the layer is used to update the features of the nodes; Is a node The set of neighbor nodes of is the normalization coefficient; it is used to normalize the influence of neighboring nodes; is the activation function, used to introduce nonlinearity;

[0023] Step 22: Use multi-level neural architecture search (NAS) technology to automatically explore and optimize the network architecture to adapt to data processing requirements and improve model performance and efficiency.

[0024] Step 23: Combine meta-learning techniques for pre-training, so that the model can use the knowledge learned on previous tasks to quickly adapt to new tasks.

[0025] Preferably, step 3 specifically includes the following:

[0026] Step 31: Intelligently label and systematically classify catalyst data using chemical knowledge graphs; identify and classify key characteristics of catalysts using machine learning method support vector machine (SVM);

[0027] ;in, is the decision function of SVM, used for classification; is the Lagrange multiplier used to optimize the decision boundary of SVM; is the category label, indicating the classification of the i-th sample; It is a kernel function used to map input data into a high-dimensional space for easy classification; is the bias term, which adjusts the position of the classification boundary;

[0028] Step 32: Adopt adaptive graph neural network GNN and multimodal deep learning framework to perform deep feature analysis and intelligent matching of catalyst data.

[0029] Preferably, the step 4 specifically includes:

[0030] Step 41: Use deep Q network and policy gradient method to adjust and optimize catalyst design parameters; ;in, Yes Status and actions The value function of It’s an instant reward. is the discount factor, In the next state By choosing the best action The maximum value that can be obtained value, which represents the best estimate of future returns; Is to perform an action The next state to enter is is in the next state The action selected in

[0031] Step 42: Create a virtual copy of the catalyst through digital twin technology to simulate the behavior and reaction process of the catalyst in digital space; specific technologies include computational fluid dynamics (CFD) to simulate the fluid behavior and mass transfer process of the catalyst under different reaction conditions, finite element analysis (FEA) to evaluate the stability and life of the catalyst structure under different stress conditions, and molecular dynamics simulation (MD) to predict the reaction mechanism and performance of the catalyst at the atomic scale;

[0032] Step 43: Identify potential failures and performance bottlenecks through machine learning prediction models and Bayesian optimization, and make adjustments or replacements in advance; use long short-term memory networks (LSTMs) to monitor and predict catalyst performance trends over time;

[0033] in, yes The hidden state of the moment, The weight matrix of the hidden state is used to update the hidden state, The input weight matrix is ​​used to update the input data; is the current time step The input data, is the bias term; is the activation function;

[0034] Step 44: Utilize sensor data fusion, using Kalman filtering and particle filtering to integrate data from multiple sensors to provide comprehensive monitoring of the catalyst status, and anomaly detection algorithms to detect anomalies in operation in real time using statistical and machine learning methods.

[0035] Preferably, the step 5 specifically includes

[0036] Step 51: The interactive user interface supports multi-user operation, allowing team members to view and share catalyst design data and analysis results in real time; the interactive user interface provides various viewing and data presentation options, including catalyst structure diagrams, performance curves, and reaction path diagrams;

[0037] Step 52: The interactive user interface is capable of generating a detailed performance report and optimization recommendations based on the analysis results; the report includes catalyst performance evaluation results, optimization parameters, data charts, and predictive analysis.

[0038] Beneficial effects of the present invention:

[0039] 1. By integrating advanced machine learning models such as graph neural networks (GNNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs), it is possible to accurately simulate and predict the performance of catalysts under various reaction conditions; these networks accurately process and analyze multi-level data from the molecular to the macroscopic scale, thereby improving the accuracy and reliability of catalyst design.

[0040] 2. By adopting an adaptive neural network architecture and a multimodal deep learning framework, the network structure and learning process can be automatically optimized to adapt to different data types and design requirements. This not only improves learning efficiency, but also optimizes the overall performance of the catalyst.

[0041] 3. Using UMAP technology and automatic chemical knowledge graphs, we can effectively reduce the dimension and maintain key information when processing high-dimensional data, making the analysis of complex data more efficient and accurate. This is crucial for quickly identifying new catalyst candidates and optimizing existing designs.

[0042] 4. Combine online learning mechanisms and digital twin technology to monitor the performance of the catalyst in real time and make adjustments based on real-time feedback. This immediate feedback loop greatly improves the adaptability and responsiveness of the design. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A module diagram of the platform for designing next-generation catalysts based on artificial intelligence driven by high-dimensional data;

[0044] Figure 2 It is a functional diagram of the adaptive neural network architecture module;

[0045] Figure 3 Functional diagram for knowledge-driven feature engineering module;

[0046] Figure 4 Optimize module functional graphs for iterative feedback. DETAILED DESCRIPTION

[0047] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. It should be noted that the words "front", "rear", "left", "right", "upper" and "lower" used in the following description refer to directions in the accompanying drawings, and the words "inner" and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0048] This embodiment provides a high-dimensional data-driven next-generation catalyst platform based on artificial intelligence, which specifically includes the following components:

[0049] Components (1) Multidimensional data fusion module: using data processing techniques to integrate experimental and computational data;

[0050] Component (2) Adaptive neural network architecture: Dynamically adjust the network structure according to the performance requirements of the catalyst to optimize the selectivity and activity of the catalyst;

[0051] Component (3) Knowledge-driven feature engineering module: Combined with chemical knowledge graph, it automatically extracts the features that have the greatest impact on catalytic performance;

[0052] Component (4) Iterative feedback optimization module: using online learning mechanism to adjust parameters, digital twin technology to simulate behavior, machine learning prediction model and sensor data fusion to monitor status;

[0053] Component (5) Interactive visualization and decision support module: Provides visualization tools and AI decision support to help R&D personnel understand data and make scientific decisions.

[0054] In a specific embodiment, the present invention applies a next-generation catalyst design platform based on artificial intelligence, and conducts comprehensive catalyst performance evaluation and optimization by combining high-dimensional data processing, adaptive neural network architecture and digital twin technology, such as Figure 1 In this embodiment, the system specifically analyzes the microstructure and reaction kinetics of the catalyst, such as the active sites, reaction rate and stability of the catalyst, thereby providing comprehensive optimization of the catalyst performance.

[0055] In a specific embodiment, it is assumed that the user uploads high-dimensional catalyst experimental data and calculation data, including microstructure images, molecular structure data, reaction condition data, etc. Among them, the microstructure image:

[0056] 1000 catalyst microstructure images, each with a resolution of 1024x1024 pixels. Molecular structure data: 1000 catalyst molecular structures, expressed in SMILES format. Reaction condition data:

[0057] 1000 sets of reaction condition data, including temperature, pressure, reaction time, etc., each set of data has 10 different conditions.

[0058] First, the multidimensional data is preprocessed, and the high-dimensional data composed of catalyst test data and computational simulation data are reduced in dimension, the key features are retained, and a simplified data set suitable for subsequent analysis is generated; the processed data will be used as input for subsequent network architecture optimization and feature extraction; for microstructure images, all images are adjusted to 256x256 pixels and grayscale normalized. Molecular structure data: The molecular structure in SMILES format is converted into a molecular fingerprint vector. Reaction condition data: The temperature, pressure, reaction time and other data are normalized so that the data range is between 0 and 1. The processed high-dimensional data (microstructure image, molecular structure fingerprint vector and reaction condition data of each sample) are merged into an overall high-dimensional data matrix. The UMAP algorithm is used to reduce the dimension of the preprocessed high-dimensional data. From the original (1024x1024 + molecular fingerprint vector + 10) dimension to 2 dimensions.

[0059] ;in, It is a high-dimensional data of 1000 dimensions. It is the 2D data after dimensionality reduction.

[0060] UMAP achieves low-dimensional representation of high-dimensional data by maintaining the local neighborhood structure of the data. The process includes: Constructing a k-nearest neighbor graph: finding the k nearest neighbors for each data point. Optimizing low-dimensional embedding: By optimizing the objective function, the neighborhood structure in the low-dimensional space is made as close as possible to the neighborhood structure in the high-dimensional space. The data after dimensionality reduction retains the main features, which is convenient for subsequent feature extraction and analysis.

[0061] Using multi-level neural architecture search NAS and meta-learning technology, based on the simplified data set obtained in step 1, the network architecture is automatically customized and optimized for the catalyst design task, and key performance features are extracted; the extracted features will provide a basis for in-depth analysis of the catalyst; the pre-processed dimensionality reduction data is input into the adaptive neural network architecture, and the convolutional neural network CNN is used to process the microstructure image to extract the surface morphology and structural features;

[0062] in, Indicates that the result of the convolution operation is at position The value of is the local weighted sum of the input image and the convolution kernel at that position. The convolution operation process is to traverse each pixel position of the input image, use the convolution kernel to perform weighted summation on the local area, and obtain the value of the corresponding position of the output feature map; is the convolution kernel, indicating that the convolution kernel is The value of the position; It is a double summation, which means that the summation operation is performed on each position of the convolution kernel; It indicates the position offset of the current convolution kernel in the input image. The pixel value after the reaction is calculated; the recurrent neural network (RNN) is used to analyze the time series data of the reaction conditions and predict the performance change trend;

[0063] in, yes The hidden state at the moment represents the hidden state of the time series; The weight matrix of the hidden state is used to update the hidden state, The input weight matrix is ​​used to update the input data; is the current time step The input data, is the bias term; is the activation function; the graph neural network GNN is used to process molecular structure data to capture the relationship between atoms and chemical bond characteristics;

[0064] ;in ,yes The feature vector of node i in the layer indicates the representation of the node in the layer; It is Node in layer The characteristic vector of It is The weight matrix of the layer is used to update the features of the nodes; Is a node The set of neighbor nodes of is the normalization coefficient; it is used to normalize the influence of neighboring nodes; is the activation function, used to introduce nonlinearity;

[0065] Utilize multi-level neural architecture search (NAS) technology to automatically explore and optimize network architecture. The specific steps include: defining the search space: including different neural network layer types, number of layers, and connection methods; using reinforcement learning or evolutionary algorithms to find the best architecture in the search space; evaluating the performance of each candidate architecture and selecting the optimal architecture based on performance indicators.

[0066] Combine meta-learning technology for pre-training. By pre-training the model on a large number of related tasks, its generalization ability on new tasks can be improved. The specific steps include: training the basic model on multiple related tasks, such as training the basic model on other catalyst data sets; using the pre-trained basic model as the initial state, fine-tuning on new data to improve the generalization ability of the model. Figure 2 shown.

[0067] The optimized neural network model and catalyst data are input into the knowledge-driven feature engineering module, such as Figure 3 As shown in the figure, the catalyst data is intelligently annotated and systematically classified using the automatic chemical knowledge graph. The specific steps include: constructing a chemical knowledge graph that includes the chemical structure, reaction mechanism, and performance data of the catalyst; using natural language processing technology to extract chemical knowledge from literature and databases and update the knowledge graph. The key characteristics of the catalyst are automatically identified and classified using the support vector machine (SVM).

[0068] ;in, is the decision function of SVM, used for classification; is the Lagrange multiplier used to optimize the decision boundary of SVM; is the category label, indicating the classification of the i-th sample; It is a kernel function used to map input data into a high-dimensional space for easy classification; is the bias term, which adjusts the position of the classification boundary;

[0069] Adaptive graph neural network (GNN) and multimodal deep learning framework are used to deeply analyze catalyst data and mine key features. The specific steps include: using GNN to analyze the molecular structure of the catalyst to capture the relationship between atoms and chemical bond characteristics. Using a multimodal deep learning framework, the image, molecular structure and reaction condition data are jointly analyzed to extract comprehensive features.

[0070] The extracted features and real-time data are input into the iterative feedback optimization module. The catalyst design parameters are adjusted and optimized using the deep Q network DQN and the policy gradient method PGM. Figure 4 shown.

[0071] Use deep Q-network and policy gradient methods to adjust and optimize catalyst design parameters;

[0072] ;in, Yes Status and actions The value function of It’s an instant reward. is the discount factor, In the next state By choosing the best action The maximum value that can be obtained value, which represents the best estimate of future returns; Is to perform an action The next state to enter is is in the next state The action selected in

[0073] Through digital twin technology DTT, a virtual copy of the catalyst is created to simulate the behavior and reaction process of the catalyst in digital space. The specific steps include: using computational fluid dynamics CFD to simulate the fluid behavior and mass transfer process of the catalyst under different reaction conditions. Using finite element analysis FEA to evaluate the stability and life of the catalyst structure under different stress conditions. Using molecular dynamics simulation MD to predict the reaction mechanism and performance of the catalyst at the atomic scale. Using machine learning prediction models and Bayesian optimization to identify potential failures and performance bottlenecks, and make adjustments or replacements in advance.

[0074] in, yes The hidden state at the moment represents the hidden state of the time series; The weight matrix of the hidden state is used to update the hidden state, The input weight matrix is ​​used to update the input data; is the current time step The input data, is the bias term; is the activation function;

[0075] By using sensor data fusion, Kalman filtering and particle filtering are used to integrate data from multiple sensors to provide comprehensive monitoring of the catalyst status. Anomaly detection algorithms are used to detect abnormal conditions in operation in real time, using methods such as principal component analysis (PCA) and support vector machine (SVM).

[0076]

[0077] in, is the data matrix; It is a characteristic vector matrix that outputs optimized catalyst design parameters and real-time monitoring data.

[0078] Through interactive visualization and decision support modules, the interface provides various views and data presentation options, including catalyst structure diagrams, performance curves and reaction path diagrams, and generates detailed performance reports and optimization suggestions based on the analysis results.

Claims

1. A high-dimensional data-driven, artificial intelligence-based next-generation catalyst design method, characterized in that: The specific steps include: Step 1: Preprocess the multidimensional data, reduce the dimensionality of the high-dimensional data consisting of catalyst test data and computational simulation data, retain key features, and generate a simplified data set suitable for subsequent analysis; the processed data will be used as input for subsequent network architecture optimization and feature extraction; UMAP is used in step 1 to reduce the dimensionality of high-dimensional data; ;in, is high-dimensional data, It is the data after dimensionality reduction; Step 2: Using multi-level neural architecture search (NAS) and meta-learning techniques, based on the simplified data set obtained in step 1, the network architecture is automatically customized and optimized for the catalyst design task, and key performance features are extracted; the extracted features will provide a basis for in-depth analysis of the catalyst; step 2 specifically includes the following: Step 21: Extract the key performance characteristics of the catalyst by integrating the convolutional neural network CNN, the recurrent neural network RNN ​​and the graph neural network GNN, wherein the graph neural network CNN is used to process the microstructure image of the catalyst and extract its surface morphology and structural characteristics; the recurrent neural network RNN ​​is used to analyze the time series data of the catalyst under different reaction conditions and predict its performance change trend; the graph neural network GNN is used to process the molecular structure data of the catalyst and capture the relationship between atoms and chemical bond characteristics; in, Indicates that the result of the convolution operation is at position The value of is the local weighted sum of the input image and the convolution kernel at that position. The convolution operation process is to traverse each pixel position of the input image, use the convolution kernel to perform weighted summation on the local area, and obtain the value of the corresponding position of the output feature map; Is the convolution kernel, indicating that the convolution kernel is at position The value of It is a double summation, which means that the summation operation is performed on each position of the convolution kernel; It indicates the position offset of the current convolution kernel in the input image. The pixel value after the reaction is calculated; the recurrent neural network (RNN) is used to analyze the time series data of the reaction conditions and predict the performance change trend; in, yes The hidden state at the moment represents the hidden state of the time series; The weight matrix of the hidden state is used to update the hidden state, The input weight matrix is ​​used to update the input data; is the current time step The input data, is the bias term; is the activation function; the graph neural network GNN is used to process molecular structure data to capture the relationship between atoms and chemical bond characteristics; in, yes The feature vector of node i in the layer indicates the representation of the node in the layer; It is Node in layer The characteristic vector of It is The weight matrix of the layer is used to update the features of the nodes; Is a node The set of neighbor nodes of is the normalization coefficient; it is used to normalize the influence of neighboring nodes; is the activation function, used to introduce nonlinearity; Step 22: Use multi-level neural architecture search (NAS) technology to automatically explore and optimize the network architecture; Step 23: Combine meta-learning technology for pre-training; Step 3: Use chemical knowledge graphs and machine learning methods to deeply analyze the key performance features extracted in step 2, and further explore the potential features and associations of the catalyst; this analysis provides data support for subsequent iterative optimization; Step 3 specifically includes the following: Step 31: Intelligently label and systematically classify catalyst data using chemical knowledge graphs; identify and classify key characteristics of catalysts using machine learning method support vector machine (SVM); in, is the decision function of SVM, used for classification; is the Lagrange multiplier used to optimize the decision boundary of SVM; is the category label, indicating the classification of the i-th sample; It is a kernel function used to map input data into a high-dimensional space for easy classification; is the bias term, which adjusts the position of the classification boundary; Step 32: Adopting adaptive graph neural network GNN and multimodal deep learning framework to perform deep feature analysis and intelligent matching of catalyst data; Step 4: Based on the analysis results of step 3, the catalyst design data is further adjusted and optimized by iterative feedback optimization; the optimized design will enter the next step of verification and discussion; step 4 specifically includes: Step 41: Use deep Q network and policy gradient method to adjust and optimize catalyst design parameters; ;in, Yes Status and actions The value function of It’s an instant reward. is the discount factor, In the next state By choosing the best action The maximum value that can be obtained value, which represents the best estimate of future returns; Is to perform an action The next state to enter is is in the next state The action selected in Step 42: Create a virtual copy of the catalyst through digital twin technology to simulate the behavior and reaction process of the catalyst in digital space; specific technologies include computational fluid dynamics (CFD) to simulate the fluid behavior and mass transfer process of the catalyst under different reaction conditions, finite element analysis (FEA) to evaluate the stability and life of the catalyst structure under different stress conditions, and molecular dynamics simulation (MD) to predict the reaction mechanism and performance of the catalyst at the atomic scale; Step 43: Identify potential failures and performance bottlenecks through machine learning prediction models and Bayesian optimization, and make adjustments or replacements in advance; use long short-term memory networks (LSTMs) to monitor and predict catalyst performance trends over time; in, yes The hidden state of the moment, The weight matrix of the hidden state is used to update the hidden state, The input weight matrix is ​​used to update the input data; is the current time step The input data, is the bias term; is the activation function; Step 44: Utilize sensor data fusion to integrate data from multiple sensors using Kalman filtering and particle filtering to provide comprehensive monitoring of the catalyst status, and anomaly detection algorithms to detect anomalies in operation in real time using statistical methods and machine learning methods; Step 5: An interactive user interface was developed to allow R&D personnel and engineers to view and discuss the catalyst design and analysis results optimized in Step 4, thereby achieving multi-party collaboration and decision support.

2. The high-dimensional data-driven next-generation catalyst design method based on artificial intelligence according to claim 1, characterized in that: The step 5 specifically includes Step 51: The interactive user interface supports multi-user operation, allowing team members to view and share catalyst design data and analysis results in real time; the interactive user interface provides various viewing and data presentation options, including catalyst structure diagrams, performance curves, and reaction path diagrams; Step 52: The interactive user interface is capable of generating a detailed performance report and optimization recommendations based on the analysis results; the report includes catalyst performance evaluation results, optimization parameters, data charts, and predictive analysis.

3. A high-dimensional data driven next-generation catalyst design platform based on artificial intelligence that runs the high-dimensional data driven next-generation catalyst design method based on artificial intelligence as claimed in claim 1, characterized in that: Specifically, it includes the following components: Components (1) Multidimensional data fusion module: using data processing techniques to integrate experimental and computational data; Component (2) Adaptive neural network architecture: Dynamically adjust the network structure according to the performance requirements of the catalyst to optimize the selectivity and activity of the catalyst; Component (3) Knowledge-driven feature engineering module: Combined with chemical knowledge graph, it automatically extracts the features that have the greatest impact on catalytic performance; Component (4) Iterative feedback optimization module: using online learning mechanism to adjust parameters, digital twin technology to simulate behavior, machine learning prediction model and sensor data fusion to monitor status; Component (5) Interactive visualization and decision support module: Provides visualization tools and AI decision support to help R&D personnel understand data and make scientific decisions.

Images