Distributed catalyst data collaborative modeling and privacy protection screening system based on ai learning

Abstract

This invention provides an AI-based distributed catalyst data collaborative modeling and privacy-preserving screening system, belonging to the field of data processing technology. It includes: an acquisition module for acquiring catalyst-related data from multiple distributed data sources, cleaning and standardizing the data to obtain a standardized dataset; a generation module for analyzing data distribution characteristics based on the standardized dataset, determining a core reference set, and generating two sets of reference dimensions to determine the feature analysis range based on the core reference set; selecting a validation sample set within and outside the feature analysis range, calculating the distribution density of the validation samples in the feature space, identifying high-density regions as key evolution nodes, constructing data association paths, and generating data calibration parameters. This invention achieves distributed catalyst data collaborative modeling while ensuring privacy protection, completing data screening, and improving the accuracy and efficiency of catalyst performance prediction.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a distributed catalyst data collaborative modeling and privacy-preserving screening system based on AI learning. Background Technology

[0002] In the field of catalyst research and development, data-driven modeling methods have gradually become an important means of optimizing catalyst performance. Typically, the relevant data comes from multiple distributed nodes. However, existing technologies may have some limitations when processing such distributed data. For example, due to considerations of data privacy and security, most nodes cannot directly share the original data, resulting in data silos, which makes it difficult for the global model to fully learn and capture the complex nonlinear relationships in the data.

[0003] Taking catalyst performance prediction as an example, existing methods may model through centralized data collection or simple aggregation, but this approach may not effectively protect the sensitive information of each node. At the same time, due to differences in data distribution, the model sometimes performs poorly in terms of generalization ability and accuracy. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a distributed catalyst data collaborative modeling and privacy-preserving screening system based on AI learning, which realizes distributed catalyst data collaborative modeling while ensuring privacy protection, completes data screening, and improves the accuracy and screening efficiency of catalyst performance prediction.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] The first aspect is an AI-based distributed catalyst data collaborative modeling and privacy-preserving screening system, including:

[0007] The acquisition module is used to acquire catalyst-related data from multiple distributed data sources, and to clean and standardize the catalyst-related data to obtain a standardized dataset.

[0008] The generation module is used to analyze the data distribution characteristics based on the standardized dataset, determine the core reference set, and generate two sets of reference dimensions based on the core reference set to determine the feature analysis range; select a validation sample set inside and outside the feature analysis range, and identify high-density regions as key evolution nodes by calculating the distribution density of validation samples in the feature space, construct data association paths, and generate data calibration parameters.

[0009] The training module is used to calibrate parameters using data, deploy deep neural networks on each distributed node for local training, and generate local model parameter adjustment information.

[0010] The protection module is used to encrypt the parameter adjustment information based on the local model parameter adjustment information using differential privacy to obtain the privacy-protected parameter information.

[0011] The module is used to comprehensively consider the contribution of each node when fusing multi-node parameter adjustment information based on privacy-preserving parameter information, and to build a global collaborative model by obtaining complex nonlinear relationships between data through deep neural networks.

[0012] The screening module is used to predict and screen catalyst performance based on the global collaborative model, and obtain adjusted catalyst screening results.

[0013] In a second aspect, a computing device includes:

[0014] One or more processors;

[0015] A storage device for storing one or more programs that, when executed by one or more processors, enable the one or more processors to implement the system.

[0016] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, implements the system.

[0017] The above-described solution of the present invention has at least the following beneficial effects:

[0018] By acquiring and integrating catalyst-related data from multiple distributed data sources, we overcome the limitations of data volume and information silos caused by single data sources, achieving comprehensive convergence of multi-source data. By analyzing data distribution characteristics to determine the core reference set, defining the scope of feature analysis and identifying key evolutionary nodes, and constructing data association paths and calibration parameters, we can effectively mine the inherent patterns and core features of the data, reduce interference from irrelevant information, and improve the relevance and effectiveness of data dimensions.

[0019] By adopting a distributed local training mode, each node independently conducts deep neural network training based on calibration parameters, without the need to centrally transmit the original dataset to the central node. This reduces bandwidth consumption and latency in data transmission, fully utilizes the computing resources of each node, and avoids overloading the central node's computing power. Differential privacy is used to encrypt local model parameter adjustment information, enabling secure sharing of parameter information across multiple nodes without leaking the privacy of the original data, effectively preventing data leakage. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of a distributed catalyst data collaborative modeling and privacy protection screening system based on AI learning, provided in an embodiment of the present invention.

[0021] Figure 2 This is a flowchart illustrating how a global collaborative model is used to predict and screen catalyst performance to obtain adjusted catalyst screening results, as provided in an embodiment of the present invention. Detailed Implementation

[0022] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0023] like Figure 1 As shown, embodiments of the present invention propose a distributed catalyst data collaborative modeling and privacy-preserving screening system based on AI learning, comprising:

[0024] The acquisition module is used to acquire catalyst-related data from multiple distributed data sources, and to clean and standardize the catalyst-related data to obtain a standardized dataset.

[0025] The generation module is used to analyze the data distribution characteristics based on the standardized dataset, determine the core reference set, and generate two sets of reference dimensions based on the core reference set to determine the feature analysis range; select a validation sample set inside and outside the feature analysis range, and identify high-density regions as key evolution nodes by calculating the distribution density of validation samples in the feature space, construct data association paths, and generate data calibration parameters.

[0026] The training module is used to calibrate parameters using data, deploy deep neural networks on each distributed node for local training, and generate local model parameter adjustment information.

[0027] The protection module is used to encrypt the parameter adjustment information based on the local model parameter adjustment information using differential privacy to obtain the privacy-protected parameter information.

[0028] The module is used to comprehensively consider the contribution of each node when fusing multi-node parameter adjustment information based on privacy-preserving parameter information, and to build a global collaborative model by obtaining complex nonlinear relationships between data through deep neural networks.

[0029] The screening module is used to predict and screen catalyst performance based on the global collaborative model, and obtain adjusted catalyst screening results.

[0030] In this embodiment of the invention, by acquiring and integrating catalyst-related data from multiple distributed data sources, the limitations of data volume and information silos of a single data source are overcome, and comprehensive aggregation of multi-source data is achieved. By analyzing data distribution characteristics to determine the core reference set, dividing the scope of feature analysis and identifying key evolution nodes, constructing data association paths and calibration parameters, the inherent laws and core features of the data are effectively mined, irrelevant information interference is reduced, and the relevance and effectiveness of data dimensions are improved.

[0031] By adopting a distributed local training mode, each node independently conducts deep neural network training based on calibration parameters, without the need to centrally transmit the original dataset to the central node. This reduces bandwidth consumption and latency in data transmission, fully utilizes the computing resources of each node, and avoids overloading the central node's computing power. Differential privacy is used to encrypt local model parameter adjustment information, enabling secure sharing of parameter information across multiple nodes without leaking the privacy of the original data, effectively preventing data leakage.

[0032] In a preferred embodiment of the present invention, catalyst-related data is obtained from multiple distributed data sources, and the catalyst-related data is cleaned and standardized to obtain a standardized dataset, which may include:

[0033] In this embodiment of the invention, a secure data interaction channel is first established with each distributed data source node to clarify the scope and permissions of data sharing, ensuring that sensitive information is not leaked during data transmission. The data storage format and core field meanings of each node are communicated and confirmed, and the content dimensions of data collection are unified, including the basic properties of the catalyst, preparation process parameters, and performance test data. Data from each node is collected one by one, and a unique source identifier is added to each data entry. The integrity of the data is verified in real time during the collection process. If a node has problems such as missing fields or incorrect formats, the corresponding node is contacted in a timely manner to complete or correct the data, ensuring that the collected raw data can cover the key information dimensions of catalyst research and development.

[0034] First, a comprehensive screening of all fields in the collected data is conducted to identify missing data. For numerical missing data, if the missing rate is low, appropriate methods are used to supplement the missing data, taking into account common sense in the catalyst field and the distribution characteristics of similar data. If the missing rate is too high or the missing field is crucial to subsequent analysis, the invalid data is directly removed. For categorical missing data, priority is given to verifying and supplementing the missing data with the data source node. If verification is not possible, the missing data is marked as an unknown category to avoid arbitrary filling that could lead to data distortion. Combining professional knowledge of catalyst research and development with the overall data distribution, AI is used to intelligently identify abnormal data. If data that clearly violates logic is found, the data source node is first consulted to confirm whether it is a recording error. After verification, the data is corrected to be correct. If the data cannot be corrected or is confirmed to be true but is an extreme exception, it is marked separately and retained, with the reason for the anomaly noted in the description.

[0035] Based on the key unique identifier of the catalyst, duplicate data is screened. If multiple data entries have completely identical key identifiers and identical core fields, they are judged as completely duplicate data, and only one valid record is retained. If the key identifiers are identical but some non-core fields have slight differences, the record with more comprehensive information and a more authoritative source is retained, taking into account the completeness and credibility of the data, while marking the duplicate processing. The consistency of field logic and format of data from different data sources is checked. On the one hand, the logical relationship within a single data entry is verified, and logical contradictions are promptly communicated and corrected with the data source node. On the other hand, the field definitions and formats of different nodes are unified. For example, if the catalytic activity of node A and the reaction conversion rate of node B essentially refer to the same indicator, the field name is unified as catalytic activity. If node A uses K to record temperature and node B uses ℃, then all are uniformly converted to ℃. The description of classification fields is unified to ensure that data from different sources can be integrated.

[0036] A comprehensive review of the units of measurement for all numerical data was conducted, standardizing the measurement methods according to common standards in the catalyst field. For example, pressure was standardized to MPa, time to h, mass to g, and concentration to mol / L. This ensured consistency in measurement for the same type of data from different data sources, preventing analytical biases caused by unit confusion. For differences in the numerical scales of different indicators, scale adjustments were made, adhering to the principle of not altering the relative relationships of the data. The overall distribution range of all data was considered, adjusting the values ​​of various numerical indicators to a similar and reasonable range to avoid bias caused by excessively large numerical values ​​for any single indicator. The analysis is led by the standardization process, which preserves the inherent size and distribution characteristics of the data to ensure that the standardized data still accurately reflects the properties and performance differences of the catalyst. All cleaned and standardized data are integrated into a complete dataset according to a unified field order and format. The dataset must include all key information such as the core properties of the catalyst, process parameters, and performance indicators, and each data point must be accompanied by a source identifier and processing record. Finally, the integrated dataset is validated to confirm that there are no residual missing values, outliers, or duplicate data, and that the field definitions are unified, the units of measurement are consistent, and the numerical scales are coordinated, ultimately forming a standardized dataset that can be directly used for analysis.

[0037] In a preferred embodiment of the present invention, the data distribution characteristics are analyzed based on a standardized dataset to determine a core reference set. Two sets of reference dimensions are generated based on the core reference set to determine the feature analysis range. A validation sample set is selected within and outside the feature analysis range. The distribution density of the validation samples in the feature space is calculated to identify high-density regions as key evolutionary nodes. Data association paths are constructed and data calibration parameters are generated. This process may include:

[0038] In this embodiment of the invention, the local density and relative distance of data points in a standardized dataset are calculated and analyzed to identify data points with high local density in the data space. Specifically, this includes: firstly, determining that each data point in the standardized dataset contains key feature information of the catalyst, which covers multiple dimensions such as elemental composition ratio, preparation process parameters, performance test indicators, and experimental conditions, with specific numerical records for each dimension; when calculating the local density of each data point, a reasonable fixed distance threshold is first set based on the overall feature distribution range of the catalyst data, ensuring that the threshold can cover neighboring data points with similar features around each data point; for each target data point, all other data points in the standardized dataset are traversed, and the feature distance between the target data point and each other data point is calculated. The calculation method is to subtract the values ​​of the corresponding feature dimensions of the two data points one by one, take the absolute value of each difference, and then add them together to obtain the total absolute difference, which is the feature distance between the two; then, the number of other data points whose feature distance to the target data point is less than the set distance threshold is counted, and this number is recorded as the number of neighboring data points, and then the distance between these neighboring data points and the target data point is calculated. The differences in the values ​​of each feature dimension of the data points are summed to obtain the total difference. The local density value of the target data point is obtained by dividing the total difference by the number of neighboring data points. The larger the local density value, the more concentrated the similar data around the data point. When calculating the relative distance of each data point, all other data points with local density values ​​greater than the local density value of the current data point are first screened out. The feature distance between the current data point and these screened data points is calculated one by one. All calculated feature distances are compared, and the smallest distance is selected as the relative distance of the current data point. The relative distance reflects the proximity of the current data point to the core of the high-density area. After completing the calculation of the local density and relative distance of all data points, all data points are sorted from largest to smallest local density value. The top 30% of data points after sorting are selected as candidate high local density data points. This proportion can be adjusted according to the total amount and distribution of catalyst data. The relative distance of the candidate data points is then checked. If the relative distance of a candidate data point is much smaller than the relative distance of other candidate data points, it indicates that the data point may be an abnormally dense isolated point and is excluded. Finally, data points with high local density are determined.

[0039] Based on data points with high local density, high-density regions in the data distribution are identified, and the center points of each high-density region are calculated to form a core reference set. Specifically, this involves: first, grouping all identified high-local-density data points. The core of this grouping is based on the feature similarity between data points. The feature distance between any two high-local-density data points is calculated by subtracting the values ​​of each corresponding feature dimension of the two data points one by one, squared each difference, summing the results, and then taking the square root of the sum. This result is the feature distance between the two data points. Referring to the feature distribution span of the core reference set, a region division threshold is set. This threshold must ensure that data points within the same high-density region have strong feature correlations. When the feature distance between two high-local-density data points is less than this region division threshold, these two data points are grouped together. This operation is repeated, comparing the distances of each high-local-density data point with other data points sequentially, until all high-local-density data points are grouped together. High local density data points are assigned to corresponding groups, each group being a high-density region in the data distribution. Catalyst data within each region share similar core characteristics. For each high-density region, the sum of values ​​for each feature dimension of all data points within that region is calculated. Using each feature dimension, such as catalyst elemental composition, calcination temperature, reaction pressure, and catalytic activity, as units, the values ​​of all data points within that region are summed for that dimension to obtain the sum of values ​​for each feature dimension. The sum of values ​​for each feature dimension is divided by the total number of data points within that region to obtain the average value for that feature dimension. The average values ​​of all feature dimensions are combined to form the center point of the high-density region. The value of each feature dimension at the center point represents the average level of the corresponding region. All center points of all high-density regions are collected to form a core reference set. Each center point can accurately represent the core characteristics of the corresponding high-density region, comprehensively reflecting the main distribution patterns of the catalyst data.

[0040] Using the core reference set as the analysis object, principal component analysis is used to extract the main directions of change in the data, resulting in two mutually orthogonal principal component vectors. Specifically, each centroid in the core reference set is treated as an independent sample, and each feature dimension of each centroid is treated as an attribute of the sample. These are organized into a sample matrix. If the core reference set has K centroids, and each centroid has L feature dimensions, then the sample matrix is ​​a K x L matrix, with each row corresponding to a centroid and each column corresponding to a feature dimension. The variance of each feature dimension in the sample matrix is ​​calculated by first calculating the average value of all centroid values ​​for that feature dimension, and then using the variance of each centroid... Subtracting the mean from the value in this dimension yields the deviation value for each center point. Squaring all deviation values ​​and summing them, then dividing the sum by the total number of center points, yields the variance for this feature dimension. A larger variance indicates a higher contribution of this feature dimension to the differences in catalyst data. To calculate the covariance between different feature dimensions, take two feature dimensions, calculate the mean of all center point values ​​for each dimension, subtract the mean of that dimension from the value of each center point in the first dimension to obtain the deviation value for the first dimension, and so on for the second dimension. Multiply the deviation values ​​for the two dimensions one by one, sum them, and then divide by the center point value. The total number of features is used to obtain the covariance between these two feature dimensions. A positive covariance indicates a positive correlation between the two dimensions, a negative covariance indicates a negative correlation, and zero indicates no linear correlation. Based on the variances of all feature dimensions and the covariance between feature dimensions, a covariance matrix is ​​constructed. By analyzing the covariance matrix, the direction of the first principal component is identified. This direction best preserves data differences among all combinations of feature dimensions, and its corresponding variance is the largest among all possible directions, reflecting the most significant trends in catalyst data, such as the differences in core performance indicators of different catalysts. Then, directions orthogonal to the first principal component direction are found; orthogonality means that the covariance between the two directions is... Zero indicates that the catalyst data change trends reflected in the two directions are independent of each other, and the variance of this direction is the largest among all the remaining directions. This direction is identified as the second principal component direction, which can reflect the secondary change trends of the catalyst data, such as the differences in preparation process parameters. The first and second principal component directions are represented by vectors, and each vector contains elements corresponding to the number of feature dimensions. The element values ​​represent the contribution weight of that feature dimension in the principal component direction, resulting in two mutually orthogonal principal component vectors. These two vectors can condense the main feature information of the core reference set, achieving data dimensionality reduction while retaining key differences.

[0041] Two mutually orthogonal principal component vectors are defined as the first reference dimension and the second reference dimension, respectively, to define the scope of feature analysis. Specifically, the principal component vector with the largest variance is defined as the first reference dimension. This dimension primarily characterizes the core distribution features of the catalyst data, such as the correlation between catalyst elemental composition and catalytic activity, which is the main source of data distribution differences. The principal component vector with the second largest variance and orthogonal to the first reference dimension is defined as the second reference dimension. This dimension primarily characterizes secondary distribution features of the catalyst data, such as the combination and variation of preparation process parameters like calcination temperature and reduction time, supplementing the reflection of secondary data differences. The values ​​of all centroids in the core reference set on the first reference dimension are calculated. This is done by multiplying the feature dimension value of each centroid by the corresponding element of the first principal component vector one by one, and then summing all the products to obtain the value of that centroid on the first reference dimension. Similarly... Calculate the values ​​of all center points in the second reference dimension; find the maximum and minimum values ​​of all center points in the core reference set in the first reference dimension. To avoid the feature analysis range being too narrow and missing edge core features, add a preset expansion ratio to the maximum value of the first reference dimension and subtract the preset expansion ratio from the minimum value. This is used as the final boundary of the first reference dimension. The expansion ratio is set according to the distribution dispersion of the core reference set, usually 10% of the difference between the maximum and minimum values. Similarly, find the maximum and minimum values ​​of all center points in the core reference set in the second reference dimension and expand the boundary using the same expansion ratio to obtain the final boundary of the second reference dimension. The upper and lower boundaries of the expanded first and second reference dimensions together form a two-dimensional feature analysis range. This range can fully cover the main distribution area of ​​the core reference set, while also extending appropriately to include data points with related features in the surrounding area.

[0042] Within and outside the feature analysis range, sample points are randomly selected according to a preset ratio to form an initial validation sample set. Specifically, this includes: first, setting a clear preset ratio based on the proportion of data volume in the core and peripheral regions of the catalyst data; if the amount of data within the feature analysis range in the standardized dataset is large, the number of sample points selected within the feature analysis range can be set to 70% of the total number of sample points in the initial validation sample set, and the number of sample points selected outside the feature analysis range can be set to 30%; if the amount of data in the peripheral region is too small, it can be adjusted to 60% inside and 40% outside to ensure that the external samples are sufficiently representative; then, traversing all data points in the standardized dataset, it is determined whether the values ​​of each data point in the first and second reference dimensions are simultaneously within the corresponding dimensional boundaries of the feature analysis range; if so, the data point is determined to be within the feature analysis range; otherwise, it is determined to be outside the range. The total number of data points within and outside the specified range is recorded. Within the feature analysis range, random sampling without replacement is used to calculate the required number of sample points according to a preset ratio, ensuring that the sample points are evenly distributed across the range and cover different feature combinations. For example, data points corresponding to catalysts with different elemental compositions and process parameters can be sampled. Similarly, random sampling without replacement is used to calculate the required number of sample points outside the specified range, prioritizing data points close to the boundary of the feature analysis range. Although these data points are not in the core area, they are correlated with the core features. A small number of data points further away are also sampled to reflect extreme feature situations. All sample points sampled from both inside and outside the range are integrated to form the initial validation sample set.

[0043] Kernel density estimation is performed on the feature space of the initial validation sample set, which is composed of reference dimensions, to obtain the probability density estimate for each sample point. Specifically, this involves: first, selecting an appropriate kernel function based on the distribution characteristics of the catalyst data. The core features of the catalyst data mostly exhibit a continuous and smooth correlation trend, without drastic abrupt changes; therefore, a Gaussian kernel function is chosen. The core advantage of this kernel function is its ability to simulate the smooth correlation characteristics between catalyst data. Next, the bandwidth parameter of the kernel function is determined. The bandwidth parameter directly determines the effective range of the kernel function, thus affecting the accuracy of the density estimation. If the bandwidth parameter is too large, it will over-smooth local differences in the catalyst data, such as mixing highly active catalysts with moderately active catalysts. The dense regions of catalysts merge into a whole, masking the core distribution characteristics of catalysts with different performance levels. If the bandwidth parameter is too small, it will be overly sensitive to noise points in the data, such as misjudging abnormal data caused by single experimental errors as local dense regions, forming false core feature nodes. Therefore, the determination of the bandwidth parameter needs to be combined with the actual distribution characteristics of the catalyst data and carried out according to the following specific steps: First, define the candidate bandwidth range. First, calculate the difference between the upper and lower boundaries of the first reference dimension in the feature analysis range, and then calculate the difference between the upper and lower boundaries of the second reference dimension. Take the larger of the two differences and set 5% to 15% of the larger value as the range of candidate bandwidth values.

[0044] The second step involves using 10-fold cross-validation, combined with AI cross-validation logic, to select the final bandwidth. The initial validation sample set is randomly divided into 10 equal parts, each containing approximately the same number of samples and covering data from different regions outside the feature analysis scope. Nine of these parts are selected as the training set, and the remaining part as the test set. For each candidate bandwidth, the following operations are performed: using all sample points in the training set, a Gaussian kernel density calculation rule is constructed based on the candidate bandwidth, and the density estimate of each sample point in the test set under this rule is calculated. Simultaneously, the actual local density of each sample point in the test set within the training set is calculated, i.e., the number of sample points in the training set whose feature distance to the test sample point is less than the candidate bandwidth, divided by the total number of samples in the training set. Calculate the difference between the density estimate and the actual local density of each test sample point, and square the absolute value of the difference to obtain the error value of a single sample point. Sum the error values ​​of all sample points in the test set to obtain the cross-validation error of the candidate bandwidth under the current combination of training and test sets. Repeat the above process, taking each sample as the test set in turn, and calculate the 10 cross-validation errors corresponding to the candidate bandwidth. Sum the 10 errors and take the average value to obtain the average cross-validation error of the candidate bandwidth. Finally, compare the average cross-validation errors of all candidate bandwidths and select the candidate bandwidth with the smallest average error as the final kernel function bandwidth parameter to ensure that the density estimation result can accurately balance local feature capture and noise suppression.

[0045] For each target sample point in the initial validation sample set, the probability density estimate is calculated according to the following detailed process: First, locate the feature space center of the target sample point and extract the specific values ​​of the target sample point in the first reference dimension and the second reference dimension. These two values ​​together constitute the unique position of the target sample point in the feature space, which is the center position of the kernel density estimate, corresponding to a specific feature combination of the catalyst data. Second, calculate the feature space distance between the target sample point and all other sample points. Iterate through each other sample point in the initial validation sample set, and for each other sample point, perform the following distance calculation: first, extract the value of the first reference dimension of the other sample point. The first dimension difference is obtained by subtracting the first reference dimension value of the other sample points from the first reference dimension value of the target sample point. This first dimension difference is then multiplied by itself to obtain the square of the first dimension difference. Next, the second reference dimension value of the other sample points is taken, and the second reference dimension value of the target sample point is subtracted from it to obtain the second dimension difference. This second dimension difference is then multiplied by itself to obtain the square of the second dimension difference. The square of the first dimension difference and the square of the second dimension difference are added together to obtain the sum of squares. The square root of this sum of squares is then taken to obtain the straight-line distance in the feature space between the target sample point and the other sample points. The first step reflects the similarity of the catalyst feature combinations corresponding to two sample points; the smaller the distance, the more similar the feature combinations. The third step calculates the influence of each other sample point on the target sample point. Based on the characteristics of the Gaussian kernel function, the influence is calculated step by step: first, the feature space distance obtained in the second step is multiplied by itself to obtain the square of the feature distance; then, the square of the feature distance is multiplied by negative one to obtain a negative result; next, the bandwidth parameter is processed: the final determined bandwidth parameter is multiplied by itself to obtain the square of the bandwidth parameter; the square of the bandwidth parameter is multiplied by two to obtain twice the square of the bandwidth parameter; the negative result obtained earlier is divided by twice the square of the bandwidth parameter to obtain the input value for exponential operation; this input value... The natural index calculation yields the final value representing the influence of other sample points on the density estimation of the target sample point. In this process, the closer the other sample points are in the feature space, the greater their calculated influence and the more significant their contribution to the density of the target sample point. Conversely, the farther away the sample points are, the smaller their influence, even approaching zero, perfectly aligning with the distribution pattern of clustered similar feature combinations in catalyst data. The fourth step involves calculating the probability density estimate of the target sample point by adding the influence of the target sample point to the influence of all other sample points in the initial validation sample set, thus obtaining the total influence. This total influence comprehensively reflects the clustering degree of all catalyst data surrounding the feature space location of the target sample point.The total influence is then divided by the total number of samples in the initial validation sample set to obtain the standardized probability density estimate. This value directly corresponds to the density of catalyst data at the target sample point. A larger value indicates more catalyst data with similar characteristic combinations in the surrounding area, resulting in more universal and representative catalytic features that better reflect the core distribution patterns of the catalyst data. Conversely, a smaller value indicates sparser catalyst data at that location, suggesting that the corresponding catalytic features may be from special operating conditions or accidental test results.

[0046] The probability density estimate is compared with a preset density threshold to filter out all sample points that exceed the threshold, and the regions where these sample points are located are marked as key evolutionary nodes. Specifically, this involves: First, calculating the density threshold. The first step is to calculate the average probability density estimate of all sample points in the initial validation sample set by dividing the sum of the probability density estimates of all sample points by the total number of sample points. The second step is to calculate the standard deviation of the probability density estimate. This involves calculating the difference between the probability density estimate of each sample point and the average, squaring each difference, summing the squares, dividing the sum of squares by the total number of sample points to obtain the variance, and then taking the square root of the variance to obtain the standard deviation. The third step is to determine the density threshold. The threshold coefficient is adjusted based on the concentration of the catalyst data. If the data distribution is relatively concentrated and the standard deviation is small, the threshold coefficient is set to 0.8; if the distribution is relatively dispersed and the standard deviation is large, the threshold coefficient is set to 1.2; under normal circumstances, 1.0 is used. The density threshold = average + standard deviation × threshold coefficient. This threshold can both filter out truly dense regions and avoid missing important local dense points.

[0047] Next, high-density sample points are selected. The probability density estimate of each sample point in the initial validation sample set is compared with the preset density threshold one by one. All sample points with probability density estimates greater than the density threshold are retained. The feature space regions where these sample points are located are the core dense regions of catalyst data, and the corresponding catalytic features are highly representative. Sample points with probability density estimates less than or equal to the density threshold are removed. These sample points are located in sparse data regions, mostly random test results or atypical operating conditions, and are not very meaningful in reflecting the core distribution characteristics of the data. Then, the regional range of key evolution nodes is defined. First, the feature space coordinates of the retained sample points are determined, that is, the specific values ​​of each retained sample point in the first and second reference dimensions. Second, the node radius is set, and the average feature space distance between all retained sample points is calculated. 30% of this average value is taken as the base radius. If there are many other retained sample points around a certain retained sample point, that is, the local density is extremely high, the base radius is increased. The first step is to reduce the radius by 20%. If the number of surrounding retained sample points is small, i.e., the local density is relatively low, the base radius is increased by 20% to ensure that each node region can accurately cover the high-density associated data points around the sample point. The second step is to delineate the region boundary. Taking the feature space coordinates of each retained sample point as the center, the boundary is delineated by subtracting the node radius from the center value and adding the node radius to the center value in the first reference dimension. The same node radius is used in the second reference dimension to form a two-dimensional rectangular region with the sample point as the core. This region is the initial range of a single key evolution node. The third step is to merge overlapping regions. If the initial ranges of two key evolution nodes overlap, i.e., the boundaries intersect or contain each other, the average probability density estimate of all retained sample points in the two regions is calculated. Taking the sample point with the higher average value as the center, the merged node radius is recalculated. The maximum value of the two initial radii is taken to delineate the boundary of the merged region to avoid analysis redundancy caused by node overlap and to ensure that each key evolution node is an independent core dense region.

[0048] Finally, key evolution nodes are marked, and a unique identifier is assigned to each final key evolution node. Detailed information for each node is recorded, including the region boundary range, the feature dimension values ​​of all retained sample points within the region, the range of probability density estimates, and the corresponding catalytic feature type. These nodes collectively reflect the core dense region of catalyst data in the feature space and are important positions in the data distribution evolution. They can clearly show the distribution and aggregation of catalysts with different core features, such as the distribution of catalyst feature nodes corresponding to different reaction types.

[0049] Based on key evolution nodes, the distance between any two nodes is calculated as the connection cost. Connections between nodes are established sequentially in ascending order of connection cost, ensuring no loops are formed, until all nodes are connected and the overall connection cost is minimized, forming a data association path. Specifically, this includes: first, determining the center coordinates of key evolution nodes; for each key evolution node, calculating its center coordinates in the first and second reference dimensions. The calculation method is to add the values ​​of all retained sample points in the first reference dimension within the node's region one by one to obtain the sum of the values ​​in the first dimension, and then divide this sum by the number of retained sample points in the region to obtain the center coordinates in the first reference dimension; using the same method, the sum of the values ​​of all retained sample points in the second reference dimension within the node's region is calculated, and then divided by the number of sample points to obtain the center coordinates in the second reference dimension. The combination of the two coordinates is the center coordinates of the key evolution node, which represents the core position of the node.

[0050] Next, the connection cost between any two nodes is calculated. The connection cost is the straight-line distance between the center coordinates of the two key evolution nodes in the feature space. It is calculated by subtracting the center coordinates of the first reference dimension of the first node from the center coordinates of the second node, and then squaring the difference. Similarly, the difference between the center coordinates of the second reference dimension of the first node and the center coordinates of the second node is subtracted from the center coordinates of the second reference dimension of the first node, and then squaring the difference. The two squared results are added together, and the arithmetic square root of the sum is taken to obtain the connection cost between the two nodes. The physical meaning of the connection cost is the similarity of the catalytic features represented by the two nodes. The smaller the connection cost, the more similar the core features of the catalysts corresponding to the two nodes are. For example, they are both highly active catalysts for the same reaction, with only differences in the proportion of additives. The larger the connection cost, the more significant the difference in catalytic features. For example, one is a hydrogenation reaction catalyst node, and the other is an oxidation reaction catalyst node.

[0051] Then, a minimum spanning tree is constructed to form data association paths. The first step is to initialize the data structure: establish a connection set to store established node connections, where each connection contains the identifiers of the two nodes and their corresponding connection cost; establish a set of connected nodes to record nodes already included in the connection network (initially empty); and establish a set of unconnected nodes to record nodes to be included in the connection network (initially containing the identifiers of all key evolutionary nodes). The second step is to initiate connection construction: arbitrarily select a node from the unconnected node set as the starting node, add it to the connected node set, and simultaneously remove the node from the unconnected node set. The third step is to iteratively build connections: traverse all combinations of connected and unconnected nodes, calculate the connection cost for each pair, and form a candidate connection list; select the connection with the minimum connection cost from the candidate connection list, where one end of the connection is a connected node and the other end is an unconnected node; check if the connection will form a loop, traversing the connected nodes... For all connections in the set, determine whether the two nodes of the newly added connection are already connected through existing connection paths. That is, starting from one node, can another node be reached through existing connections? If it can be reached, it means that a loop will be formed, so abandon the connection and return to the candidate connection list to filter the next connection with the lowest connection cost. If it cannot be reached, a loop will not be formed, and proceed to the next step. Add the connection to the connection set, add the corresponding unconnected node to the set of connected nodes, and remove the node from the set of unconnected nodes. Fourth step, termination condition, repeat the iteration operation of the third step until the set of unconnected nodes is empty. At this time, all key evolution nodes are included in the set of connected nodes. The network composed of all connections in the connection set is the minimum spanning tree, which is the data association path. This path has the lowest overall connection cost and can connect all key evolution nodes, clearly showing the association between different core dense areas, such as the evolution path from the basic catalyst feature node to the optimized high-performance catalyst feature node.

[0052] Finally, optimize the data association path and calculate the average cost of all connections in the connection set. If the connection cost of a certain connection is more than twice the average, it indicates that the characteristics of the two corresponding nodes are too different. Check whether there are other key evolutionary nodes between these two nodes that have not been identified. If so, they are identified and added to the path. If not, the connection is retained, but marked as a cross-feature connection in the path to reflect the potential association between different types of catalysts, ensuring that the path can reflect the general evolutionary context without omitting special associations.

[0053] The density change rate between adjacent key evolutionary nodes is calculated along the data association path to obtain the path density gradient. Based on this gradient, data calibration parameters are generated to adjust and calibrate the data distribution. Specifically, this involves: first, identifying adjacent node pairs along the path. Following the connection order of the data association path, all adjacent key evolutionary node pairs are identified. For example, if the path connection order is node 1 to node 3 to node 5 to node 2 to node 4, then the adjacent node pairs are node 1 and node 3, node 3 and node 5, node 5 and node 2, and node 2 and node 4, ensuring no adjacent connection is missed. The order of each adjacent node pair is also clarified, i.e., which node is the preceding node and which is the following node. Next, the density change rate of adjacent node pairs is calculated. For each adjacent node pair, the following steps are performed: First, the node density value is obtained. The density value of each key evolutionary node is the average of the probability density estimates of all retained sample points within that node region. The calculation method is a regional... The first step is to divide the sum of the probability density estimates of all sample points by the number of sample points. The second step is to calculate the density difference by subtracting the density value of the preceding node from the density value of the subsequent node in the adjacent node pair. If the difference is positive, it indicates that the density increases from the beginning to the end, and the catalytic features are more concentrated. If the difference is negative, it indicates that the density decreases, and the catalytic features are gradually dispersed. If the difference is zero, it indicates that the density does not change, and the catalytic features are highly stable. The third step is to obtain the connection cost by extracting the connection cost recorded in the data association path of the adjacent node pair, which is the feature space distance between the center coordinates of the two nodes. The fourth step is to calculate the density change rate by dividing the density difference by the connection cost. The larger the absolute value of this value, the more drastic the density change of the catalyst data within a unit feature space distance, and the more significant the change in the corresponding catalytic feature difference. The smaller the absolute value, the smoother the density change and the more moderate the change in the catalytic feature difference. This change rate is the local value of the path density gradient.

[0054] Then, a complete path density gradient is constructed. According to the order of adjacent node pairs in the data association path, the density change rate of all adjacent node pairs is arranged sequentially to form a complete path density gradient sequence. This sequence can clearly reflect the continuous change trend of catalyst data density along the entire data association path. For example, the density change rate from node 1 to node 3 is 0.3, the density change rate from node 3 to node 5 is 0.5, the density change rate from node 5 to node 2 is -0.4, and the density change rate from node 2 to node 4 is -0.2. Through this sequence, it is possible to intuitively see where the catalytic characteristics move from dispersion to concentration and where they move from concentration to dispersion.

[0055] Next, data calibration parameters are generated hierarchically. The first step is to determine the gradient grading thresholds. The average absolute value of all rates of change in the path density gradient sequence is calculated. 1.5 times the average is used as the high gradient threshold, 0.8 times the average as the medium gradient threshold, and values ​​below the medium gradient threshold are considered low gradients. Thus, the density change rate is divided into high, medium, and low levels. High gradient absolute value > high gradient threshold, medium gradient threshold ≤ absolute value ≤ high gradient threshold, low gradient absolute value < medium gradient threshold. The second step is to set the calibration parameter range. The corresponding calibration parameter value range is set according to the gradient level: 0.8-1.2 for the high gradient region, 0.5-0.8 for the medium gradient region, and 0.2-0.5 for the low gradient region. The larger the calibration parameter value, the greater the adjustment to the data distribution. The third step is to refine the parameters based on the gradient direction. For positive gradient regions, increasing density indicates that the catalytic features of subsequent nodes are more concentrated. The calibration parameter is set to the upper limit of the corresponding level range, such as 1.2 for high-gradient positive gradients, to strengthen the adjustment of preceding node data towards concentrated features. For negative gradient regions, decreasing density indicates that the catalytic features of subsequent nodes are more dispersed. The calibration parameter is set to the lower limit of the corresponding level range, such as 0.8 for high-gradient negative gradients, to weaken the differences in dispersed features and make the data distribution more balanced. The fourth step is to adjust the associated feature dimensions. Based on the weight coefficients of the principal component vectors, the target of the calibration parameters is determined. Feature dimensions with high weight coefficients in the first reference dimension, such as the loading of active components, are assigned larger calibration weights. Feature dimensions with high weight coefficients in the second reference dimension, such as preparation process parameters, are assigned relatively smaller calibration weights to ensure that the calibration parameters can specifically adjust feature dimensions that have a significant impact on the distribution of catalytic data.

[0056] Finally, a complete set of data calibration parameters is formed, compiling the corresponding calibration parameters for each adjacent node, including parameter values, the feature dimensions they affect, the adjustment direction, and the applicable node region range. Each calibration parameter has a clearly defined application scenario. For example, the calibration parameters corresponding to high gradient positive gradients are used to adjust the data in the core feature enhancement region, while the calibration parameters corresponding to low gradient negative gradients are used to adjust the data in the edge feature transition region. This parameter set can accurately adapt to the distribution characteristics of catalyst data of different distributed nodes. For example, the distribution difference between small batches of fine data in university laboratory nodes and large-scale process data in enterprise production nodes can be addressed through targeted adjustments to make the data distribution of different nodes more consistent in the core feature dimensions.

[0057] By aligning local density and relative distances one by one along feature dimensions, and dynamically adjusting the screening ratio based on the total amount of data, we ensure that high local density data points can focus on the core feature combination of the catalyst. We set the grouping threshold based on the median of the pairwise feature distances, so that the division of high-density regions conforms to the natural aggregation law of catalytic data. The center point formed by the average values ​​of each dimension can accurately represent the core features of the region. The core reference set provides a highly condensed data source for dimensionality reduction analysis.

[0058] In a preferred embodiment of the present invention, using data calibration parameters, a deep neural network is deployed on each distributed node for local training to generate local model parameter adjustment information, which may include:

[0059] In this embodiment of the invention, data calibration parameters are used as input layer preprocessing coefficients and applied to the deep neural network input layer of each distributed node to calibrate the input features, resulting in calibrated input features. Specifically, the input features received by the deep neural network input layer of each distributed node are all from locally collected catalyst-related data. This data specifically covers basic characteristics related to catalyst performance, such as the content of active metal components in the catalyst, the type and proportion of additives, the specific surface area and pore size distribution of the support, reaction temperature and pressure, reaction time, and raw material ratio. The data calibration parameters are parameters specifically used to adjust and calibrate the data distribution, generated previously by analyzing the density gradient of the data association path. Each input feature dimension corresponds to a set of dedicated data calibration parameters. It includes a feature scaling factor and a feature offset factor. During the input layer preprocessing, for each input feature, the original value of the feature is first multiplied by the corresponding feature scaling factor to obtain an intermediate result after feature scaling. Then, the intermediate result is added to the corresponding feature offset factor. Through this combination of multiplication and addition, the scale adjustment and distribution correction of the input features are achieved. Since the catalyst data of different distributed nodes may come from different laboratories or industrial scenarios, there are differences in the accuracy of testing equipment, experimental operation procedures, environmental conditions, etc., which leads to heterogeneity in the numerical range and distribution trend of the same type of features. This calibration process can unify the input features of each node to the same numerical scale and distribution range, eliminating the interference caused by data heterogeneity.

[0060] Based on the calibrated input features, forward propagation computation of a deep neural network is performed on each distributed node using the local dataset to obtain the network output. Specifically, this involves: firstly, constructing a uniformly structured deep neural network for each distributed node, consisting of an input layer, 3 to 5 hidden layers, and an output layer. The network structure is designed to fully adapt to the feature dimensions and complex nonlinear relationships of the catalyst data. The number of neurons in the input layer is completely consistent with the calibrated input feature dimensions, with each neuron specifically corresponding to one calibrated input feature, ensuring that all features are fully received by the network; the hidden layers all employ the ReLU activation function, which effectively filters features... Effective information is extracted while ineffective signals are suppressed. The number of neurons in each hidden layer is reasonably set according to the complexity of the catalyst data. The number of neurons in the first hidden layer is 1.5 to 2 times that of the input layer, and the number of neurons in each subsequent hidden layer decreases sequentially by 0.7 to 0.8 times that of the previous layer. This hierarchical design gradually extracts high-order nonlinear information from the input features. The number of neurons in the output layer corresponds one-to-one with the number of catalyst performance indicators to be predicted. These indicators include catalytic reaction conversion rate, target product selectivity, catalytic stability, and lifetime. The output layer directly outputs the predicted results of each performance indicator. During forward propagation computation, all calibrated input features are first input to the corresponding neurons in the input layer. The input layer neurons then transmit the received feature values ​​completely to each neuron in the first hidden layer. Each neuron in the first hidden layer first receives the feature values ​​transmitted by all the input layer neurons, then multiplies each feature value by the connection strength parameter corresponding to that neuron, sums the results, and adds the sum to the neuron's own bias parameter. The resulting sum is then processed by the ReLU activation function to generate the output value of that hidden layer neuron. Subsequently, the output values ​​of all neurons in the first hidden layer are... The output value serves as the input data for the second hidden layer. Each neuron in the second hidden layer repeats the above calculation process: first, it calculates the sum of the products of the input value and the corresponding connection strength parameter; then, it adds a bias parameter and processes the data through the ReLU activation function to generate the output value of the neuron in that layer. The same logic is then followed to pass the data layer by layer until it reaches the output layer. Similarly, the output layer neurons first calculate the sum of the products of the output value of the previous hidden layer and the corresponding connection strength parameter; then, it adds a bias parameter and outputs the result directly without processing it through the activation function. This output result is the predicted value of each performance index of the catalyst, which is the network output result obtained from this forward propagation calculation.

[0061] Based on the network output, the difference between the network output and the true label is calculated to obtain the output error. Then, based on the output error, the error contribution of each layer is calculated backwards from the output layer to obtain the gradient information of the parameters of each layer. Specifically, the network output is the predicted value of each performance index of the catalyst, and the true label is the accurate value of the corresponding performance index of the catalyst actually measured through rigorous experiments in the local dataset. Each predicted value has a unique corresponding true label. When calculating the output error, firstly, for each performance index, the predicted value of the index is compared with the corresponding true label value one by one, and the difference of each comparison group is calculated. Then, each difference is squared to obtain the squared difference of each comparison group. Next, sum the squared differences corresponding to all performance indicators, and finally divide the sum by the total number of performance indicators. The average value obtained is the output error. This error directly reflects the degree of deviation between the network prediction result and the actual experimental result. The smaller the error, the higher the prediction accuracy. When calculating the error contribution of each layer in reverse, we start from the output layer and work backward step by step. First, we calculate the degree of influence of the output error on the output value of the neurons in the output layer. Since the output layer does not use an activation function, its activation function derivative is 1. Therefore, the degree of influence of the output error on the output value of each neuron in the output layer is equal to the output error itself. Then, based on this degree of influence, we deduce the magnitude and direction of the influence of the output error on each bias parameter of the output layer. The error contribution of the parameters is equal to the degree of influence of the output error on the output value of the corresponding neuron. Simultaneously, combining the connection strength parameter between the output layer and the penultimate hidden layer, the degree of influence of the output error on the output value of the output layer neurons is multiplied by the corresponding connection strength parameter to obtain the magnitude and direction of the output error's influence on that connection strength parameter. After calculating the error contribution of the output layer, the error contribution of the output layer is transposed according to the connection strength parameter between the output layer and the penultimate hidden layer, and then multiplied by the derivative of the activation function of the penultimate hidden layer to obtain the degree of contribution of the output error to the output value of each neuron in the penultimate hidden layer. Based on this contribution, the degree of contribution of the penultimate hidden layer is further derived. The error contribution of each bias parameter in a layer, and the error contribution of each connection strength parameter between that layer and the penultimate hidden layer, are calculated sequentially from the output layer forward and backward. Each layer calculates its error contribution using the error contribution of the previous layer, the transpose of the connection strength parameters between the two layers, and the derivative of the activation function of the current layer. This process continues until the error contribution is passed to the first hidden layer directly connected to the input layer, completing the calculation of the error contribution for all layers. Finally, the gradient information corresponding to all connection strength parameters and bias parameters of each layer is obtained. The magnitude of the gradient information indicates the magnitude of the parameter's influence on the output error, and the sign of the gradient information indicates the direction of the parameter's influence on the output error.

[0062] Based on the gradient information of each layer's parameters, the adjustment amounts of the connection strength parameters and bias parameters are calculated. The network parameters are gradually adjusted through multiple rounds of update operations, and the update amounts of the connection strength parameters and the adjustment values ​​of the bias parameters are recorded and integrated into local model parameter adjustment information. Specifically, this includes: first, setting the initial learning rate for the deep neural network model training. The initial learning rate is determined based on the feature complexity and network structure of the catalyst data, with a value ranging from 0.001 to 0.01. Simultaneously, a learning rate decay strategy is set, reducing the learning rate to 0.9 times its original value every 100 training rounds. Dynamically adjusting the learning rate ensures stable convergence of the parameter adjustment process, avoiding oscillations or slow convergence. When calculating the adjustment amount of the connection strength parameters, each connection strength parameter is compared with... The gradient information is multiplied by the learning rate of the current round. The product is the adjustment value of the connection strength parameter. The adjustment direction is opposite to the direction indicated by the gradient information. Specifically, if the gradient information is positive, it means that the current value of the parameter is too large, resulting in an increase in output error, and the connection strength parameter needs to be reduced accordingly. If the gradient information is negative, it means that the current value of the parameter is too small, resulting in an increase in output error, and the connection strength parameter needs to be increased accordingly. When calculating the adjustment value of the bias parameter, the same calculation method is used as for the connection strength parameter. The gradient information corresponding to each bias parameter is multiplied by the learning rate of the current round to obtain the adjustment value of the bias parameter. The adjustment direction is also opposite to the direction indicated by the gradient information. This completes the first round. After parameter adjustment, the adjusted connection strength and bias parameters are updated in the deep neural network. The forward propagation calculation process is executed again, and the updated network parameters are used to re-predict the local dataset to obtain a new network output. Then, the error calculation and gradient derivation process is repeated. Based on the new network output and the true labels, a new output error is calculated, and new gradient information is derived layer by layer from the output layer. Then, based on the new gradient information and the learning rate of the current round, the next round of parameter adjustment is performed. This process is repeated continuously for multiple rounds of parameter update operations. Each round includes four stages: forward propagation prediction, error calculation, gradient derivation, and parameter adjustment. During training, the trend of output error changes is monitored in real time. Once the error decreases below a preset error threshold (typically 0.005, set based on the accuracy requirements of catalyst performance prediction) or the training epochs reach a preset maximum of 1000 epochs, the parameter update operation stops. During each parameter update epoch, the current value, adjustment amount, adjusted value, and adjustment direction of each connection strength parameter are recorded in detail. Simultaneously, the current value, adjustment value, adjusted value, and adjustment direction of each bias parameter are also recorded to ensure complete recording of every parameter change. After training stops, the parameter update records from all epochs are categorized and summarized according to network layer and parameter type, integrating them to form the local model parameter adjustment information specific to this distributed node.

[0063] By preprocessing the input features using data calibration parameters, all raw catalyst data is stored locally on the local node, eliminating the need to upload to a centralized platform or share with other nodes. This fundamentally avoids the risk of leakage of sensitive data such as catalyst formulations and process parameters, fully protecting data privacy and security, and preventing interference from invalid data.

[0064] In a preferred embodiment of the present invention, based on local model parameter adjustment information, differential privacy is used to encrypt the parameter adjustment information to obtain privacy-protected parameter information, which may include:

[0065] In this embodiment of the invention, the numerical distribution range of the local model parameter adjustment information is analyzed, and the global sensitivity of the parameters is determined based on the numerical distribution range. Specifically, the local model parameter adjustment information is the full optimization data of the connection strength parameters and bias parameters completely recorded by each distributed node in the multi-round training of the deep neural network. Specifically, it covers the single update amount, cumulative update amount, and adjustment direction of each connection strength parameter in each round of training, as well as the single adjustment value, cumulative adjustment value, and adjustment direction of each bias parameter. These data are directly related to the characteristic patterns of the local catalyst data. For example, differences in data such as the content of metal active components and reaction temperature will be directly reflected in the parameter adjustment range. When analyzing the numerical distribution range, the local model parameter adjustment information is first finely classified according to parameter type, splitting it into a connection strength parameter adjustment dataset and a bias parameter adjustment dataset. The two datasets correspond to optimization information in different dimensions. For the connection strength parameter adjustment dataset, the update records of all parameters are traversed round by round, and the update amount of each connection strength parameter in each round is recorded. The maximum positive value and the minimum negative value of all values ​​in the dataset are counted, and the difference between the maximum positive value and the minimum negative value is calculated to obtain the numerical range of the connection strength parameter adjustment. At the same time, the absolute value of the update amount of each parameter is calculated, and the value with the largest absolute value is selected and recorded as the extreme adjustment value of the connection strength parameter.

[0066] The same analysis process was performed on the bias parameter adjustment dataset. The adjustment value of each bias parameter was recorded round by round, and the maximum positive value, minimum negative value, and corresponding numerical range of the data set were analyzed. The absolute value of each adjustment value was calculated, and the extreme adjustment values ​​of the bias parameters were selected. During this process, special attention was paid to the adjustment of parameters related to the core sensitive data of the catalyst, such as model parameters corresponding to formulation components and key process parameters. The distribution characteristics of their adjustment values ​​needed to be separately labeled to ensure that the determination of global sensitivity could cover the parameter fluctuations related to the sensitive data. When determining global sensitivity, the analysis results of the two datasets were combined. First, the numerical range of the connection strength parameter and the numerical range of the bias parameter were compared, and the larger value was selected as the candidate sensitivity. Then, the extreme adjustment values ​​of the connection strength parameter and the extreme adjustment values ​​of the bias parameter were compared, and the larger value was selected as the other candidate sensitivity. Finally, considering the heterogeneity of the catalyst data, if the parameter adjustment fluctuations in different rounds were large, or the adjustment magnitude of the parameters related to the sensitive data was prominent, the larger value of the two candidate sensitivities was selected as the final global sensitivity.

[0067] Based on global sensitivity and a preset privacy budget parameter, the scale parameter of Laplace noise is calculated. Specifically, the privacy budget parameter is a core parameter jointly negotiated by each distributed node based on the privacy sensitivity level of its local catalyst data. During the negotiation process, the sensitivity of the data type is fully considered. If the local data contains highly sensitive information such as the core catalyst formulation or proprietary process parameters, the privacy budget parameter is set to a smaller value to enhance privacy protection. If the local data mainly consists of routine performance test results and general process parameters, the privacy budget parameter can be appropriately increased to retain more valid parameter information while ensuring privacy. The range of the privacy budget parameter must be combined with industry privacy protection standards to ensure compliance without affecting the subsequent fusion effect of the global model. When calculating the scale parameter of Laplace noise, the core calculation logic of differential privacy technology is strictly followed, using the determined global sensitivity as the numerator and the negotiated privacy budget parameter as the scalar value. The denominator is calculated by performing a division operation. Specifically, the global sensitivity value is divided by the privacy budget parameter value, and the quotient is the scale parameter of the Laplacian noise. For example, if the global sensitivity is a certain value and the privacy budget parameter is another specific value, the former value is divided by the latter value, and the result is the scale parameter. After the calculation, the rationality of the scale parameter needs to be verified. If the scale parameter is too large, it indicates that the noise injection may be too large, which will cause the effective features in the local model parameter adjustment information to be overly masked, affecting the global model fusion degree. If the scale parameter is too small, the noise injection is insufficient and cannot effectively mask sensitive information, posing a risk of privacy leakage. During verification, sensitive points of the catalyst data are considered, such as whether the noise corresponding to the scale parameter can mask key information such as formula ratio and process parameters. If there is an unreasonable situation, the privacy budget parameter needs to be readjusted and recalculated until the scale parameter achieves a balance between privacy protection and parameter usability.

[0068] Based on the scale parameter, a random noise sequence conforming to a Laplace distribution is generated. This random noise sequence is then superimposed with local model parameter adjustment information to obtain the parameter information after noise superposition. Specifically, this involves: first, constructing a Laplace distribution rule based on the calculated scale parameter. The position parameter of this rule is fixed at 0 to ensure that the generated random noise is symmetrically distributed around zero, preventing the noise from shifting in a specific direction and thus distorting the overall characteristics of the original parameter adjustment information. The probability density characteristics of the Laplace distribution rule are determined by the scale parameter; the larger the scale parameter value, the more dispersed the probability density distribution of the rule, the higher the probability that the generated noise value is far from zero, and the wider the noise fluctuation range. The smaller the noise level, the more concentrated the probability density distribution of the regularity, the closer the generated noise value is to zero, and the narrower the noise fluctuation range, which fully meets the core requirements of differential privacy protection for the randomness and rationality of noise. Next, the total number of local model parameter adjustment information is counted, counting sequentially according to network layers, including the update amount of connection strength parameters and the number of bias parameter adjustment values ​​at all layers, ensuring the accuracy of the statistical results. This determines the length of the random noise sequence, ensuring that the length of the random noise sequence is completely consistent with the total number of local model parameter adjustment information, guaranteeing that each parameter adjustment value corresponds to an independent noise value, achieving a one-to-one masking effect. The random noise sequence is generated using an encrypted pseudo-random number generator. During the generation process, each noise number... The values ​​are obtained from the constructed Laplace distribution rule through random sampling. The sampling process strictly follows the probability density characteristics of the model, that is, noise values ​​close to zero have a high sampling probability, and noise values ​​far from zero have a low sampling probability. After generating a complete random noise sequence, the sequence needs to be validated by calculating the mean and variance of all noise values ​​in the sequence. The mean must be infinitely close to 0, and the variance must be proportional to the square of the scale parameter. If the validation results do not meet the requirements, the random noise sequence is regenerated until it meets the Laplace distribution requirements. Then, the parameter and noise superposition process is started. Before superposition, the alignment rules of the parameters and noise need to be defined and arranged according to the layer order of the deep neural network, that is, the connection strengths of all input layers are arranged first. The parameters are updated, then all bias parameter adjustments in the input layer are arranged, followed by all connection strength parameter updates in the hidden layer, then all bias parameter adjustments in the hidden layer, and finally all connection strength parameter updates and bias parameter adjustments in the output layer. The random noise sequence is arranged in the same order. After alignment, they are aligned and superimposed one by one. That is, the first parameter adjustment value in the local model parameter adjustment information is added to the first noise value in the random noise sequence, the second parameter adjustment value is added to the second noise value, and so on, until all parameter adjustment values ​​have been added to their corresponding noise values. The positive and negative attributes of the values ​​are preserved during the addition process, and no sign correction is performed.By using this method of sequential alignment and superposition, each original parameter adjustment value is added with independent and distributed random noise, ultimately resulting in a noise-superimposed value for each parameter adjustment value. All these noise-superimposed values ​​together constitute the parameter information after noise superposition.

[0069] The parameter information after noise superposition is standardized to obtain standardized parameter information. The standardized parameter information is then verified to meet differential privacy protection requirements, resulting in privacy-preserving parameter information. Specifically, this involves: first, determining a reasonable range of parameter values ​​in the catalyst modeling scenario; then, fine-tuning this range based on the current deep neural network structure and training requirements to ultimately form a reasonable range of parameter values. This ensures that the parameters can participate normally in subsequent global model fusion and training within this range, without causing difficulties in global model convergence or a decrease in prediction accuracy due to abnormal values. Next, each parameter value after noise superposition is checked individually, comparing each value with the determined reasonable range. The system checks whether the value is within a reasonable range. If a parameter value exceeds the upper limit of the reasonable range, the parameter value is adjusted to the upper limit; if a parameter value is below the lower limit, the parameter value is adjusted to the lower limit; if the parameter value is within the reasonable range, the parameter value remains unchanged without any adjustment. Through this targeted adjustment method, all parameter values ​​after noise superposition are standardized to a reasonable range, resulting in standardized parameter information. This effectively avoids some parameter values ​​exceeding the valid range due to noise superposition. After standardization, the differential privacy protection requirement verification process is initiated. The verification process strictly follows the core definition of differential privacy. First, any two... Two adjacent catalyst datasets are selected, meaning one dataset contains only one more or one less valid catalyst sample than the other, and all other samples are identical. For these two adjacent datasets, numerical distribution range analysis, global sensitivity determination, scale parameter calculation, random noise sequence generation, overlay processing, and standardization are performed sequentially to obtain the corresponding standardized parameter information for the two sets. Next, the degree of difference between these two sets of standardized parameter information is calculated. Specifically, corresponding parameters in the two sets of standardized parameter information are paired, the absolute difference of each pair of parameters is calculated, and the maximum value among all absolute differences is selected as the difference between the two sets of standardized parameter information. The degree of difference index is then used to determine whether it meets two conditions. The first condition is that the degree of difference index does not exceed the power of a preset privacy budget parameter with the natural constant as the base. The second condition is that the probability of the degree of difference index occurring does not exceed a preset probability threshold. This probability threshold is set according to the security standards for Catalyst data privacy protection, referencing the common thresholds for data privacy protection in the industry, and adjusted in conjunction with the sensitivity level of Catalyst data to ensure sufficient privacy protection strength. If the degree of difference index of two sets of standardized parameter information simultaneously meets the above two conditions, it means that the standardized parameter information meets the differential privacy protection requirements, and the standardized parameter information is the privacy-protected parameter information.If the degree of difference index does not meet any of the conditions, the privacy budget parameter is readjusted. This typically involves appropriately reducing the privacy budget parameter, then recalculating the scale parameter of the Laplace noise based on the adjusted privacy budget parameter. A new random noise sequence is then generated and superimposed on the original parameter adjustment information. Numerical range standardization is then performed, followed by another verification of the differential privacy protection requirements. This process of adjustment, calculation, superposition, standardization, and verification is repeated until the obtained standardized parameter information simultaneously satisfies both degree of difference conditions. This ensures that the final parameter information not only meets the differential privacy protection requirements but also has the usability to participate in global model fusion.

[0070] By combining privacy budget parameters to calculate the scale parameters of Laplace noise, the noise intensity can be flexibly adjusted according to the privacy requirements of catalyst development, achieving a precise balance between privacy protection and data availability, and meeting the privacy protection needs of different distributed nodes.

[0071] In a preferred embodiment of the present invention, based on privacy-preserving parameter information, when fusing multi-node parameter adjustment information, the contribution of each node is comprehensively considered, and a global collaborative model is constructed by obtaining complex nonlinear relationships between data through a deep neural network. This can include:

[0072] In this embodiment of the invention, based on the privacy-protected parameter information of each node, the amount of local training data and the evaluation results of the local model for each node are statistically analyzed. Specifically, this includes: firstly, statistical analysis of the amount of local training data for each node is conducted. The statistical scope is strictly limited to all catalyst sample data actually used by that node in local training, covering compositional and process characteristic data such as the content of the active metal components of the catalyst, the type and ratio of the support, reaction temperature, reaction pressure, and reaction time, as well as performance test data such as catalytic activity, selectivity, and stability. Before the statistical analysis, a triple validity check is performed. The first check is the data integrity, ensuring that no key fields of each sample are missing. The first step is to fully record core information such as the content of active components, reaction temperature, and performance test results. The second step is to verify data consistency by checking whether there are logical contradictions between different test indicators of the same catalyst sample. For example, the selectivity of a highly active catalyst should not show an extremely low abnormality. The third step is to verify data credibility by removing samples whose performance test results fluctuate beyond the industry's normal error range. For example, if the difference between the activity test value of a catalyst and the average activity of similar samples exceeds 3 times the standard deviation, it is judged as an abnormal sample. After verification, the total number of remaining valid samples is counted, which is the amount of local training data for that node. At the same time, the effective proportion of data in each feature dimension is recorded.

[0073] Next, the local model evaluation results are statistically analyzed. The evaluation object is the original local training network of each node after local training, before privacy encryption. The evaluation process uses a local validation set specifically reserved for that node. The validation set samples must not overlap with the training set samples and must also meet the above triple validity verification criteria to ensure the objectivity of the evaluation results. The evaluation indicators selected are three core indicators in the catalyst modeling scenario: prediction accuracy, performance error rate, and goodness of fit. The prediction accuracy is calculated by dividing the number of samples in the local training network that match the predicted performance of all catalysts in the validation set with the actual performance by the total number of samples in the validation set. The matching criterion is the difference between the predicted value and the actual performance. The relative error of the value does not exceed 15% as recognized in the industry; the performance error rate is calculated by summing the absolute differences between the predicted performance values ​​and the true performance values ​​of all validation samples, and dividing by the total number of validation samples, reflecting the overall prediction deviation level; the goodness of fit is calculated by comparing the linear correlation between the predicted values ​​and the true values ​​of the validation samples, quantifying the network's ability to capture data trends, with a value range between 0 and 1, the closer to 1, the better the fit. These three indicators comprehensively reflect the fitting effect, prediction reliability, and data trend capture ability of the locally trained network. After the statistics are completed, the amount of local training data for each node, the results of the three evaluation indicators, and the data validity verification report are summarized and recorded.

[0074] Based on the data volume and evaluation results, the contribution ratio of each node in parameter fusion is calculated. Then, based on the contribution ratio, the privacy-protected parameter information of each node is proportionally allocated and fused to obtain the initial global parameter values. Specifically, this includes: first, initiating the AI-assisted collaborative negotiation process for each distributed node to determine the weight ratio of data volume and evaluation results in the contribution calculation. The negotiation process needs to be combined with the actual needs of catalyst development. Next, the contribution score of each node is calculated in three steps: the first step is to calculate the data volume score, which is obtained by dividing the local training data volume of the node by the sum of the local training data volumes of all distributed nodes, and then multiplying the data volume ratio by the data volume weight to obtain the data volume score of that node; the second step is to calculate the evaluation result score, first calculating the contribution score of the node... The comprehensive evaluation value of a node is calculated by multiplying the prediction accuracy by 0.5, adding the reciprocal of the performance error rate multiplied by 0.3, and adding the goodness of fit multiplied by 0.2. This weighting emphasizes the core role of prediction accuracy while also considering error control and trend capture capabilities. Next, the comprehensive evaluation value of the node is divided by the sum of the comprehensive evaluation values ​​of all distributed nodes to obtain the evaluation result percentage. This percentage is then multiplied by the evaluation result weight to obtain the node's evaluation score. The third step is to calculate the contribution ratio by adding the node's data volume score to its evaluation score. The sum of the contribution ratios of all nodes must strictly equal 100%. If there are minor deviations, adjustments are made proportionally to meet the requirement.

[0075] Then, parameter fusion is performed. Before fusion, AI-driven precise alignment of parameter positions is required. According to the hierarchical order and neuron number of the deep neural network, the correspondence of privacy-preserving parameters of each node is established. For example, the connection strength parameter of the 2nd neuron in the input layer and the 5th neuron in the hidden layer of node A needs to be paired one by one with the connection strength parameters of the same position of all nodes such as node B and node C. The bias parameters are also paired according to the rules of hierarchical and neuron numbering to ensure that parameters with the same function and the same position correspond precisely and avoid mismatch that may cause fusion failure. During fusion, a weighted average method is used. A unified calculation logic is executed for each paired parameter position. The privacy-preserving parameter value of node 1 at that position is multiplied by the contribution ratio of node 1, plus the privacy-preserving parameter value of node 2 at that position multiplied by the contribution ratio of node 2, and then the product of the parameter value of the corresponding position of all other nodes and their respective contribution ratios is added in turn. All the sums are summed to obtain the initial global parameter value of that position. After all parameter positions are fused according to this logic, they are integrated to form a complete initial parameter set of the global deep neural network.

[0076] Based on the initial values ​​of global parameters, the initial parameters of the global deep neural network are configured. Then, based on these initial parameters, the input data undergoes multi-layer nonlinear transformations through the global deep neural network to extract high-order feature representations between the data. Specifically, this includes: firstly, building a global deep neural network structure adapted to the complex characteristics of catalyst data. The structure design strictly adheres to the needs of catalyst R&D scenarios, with the number of neurons in the input layer completely consistent with the catalyst feature dimensions, covering key feature dimensions such as active component content, support ratio, reaction temperature, reaction pressure, reaction time, catalyst specific surface area, and pore size distribution, ensuring that all core factors affecting catalyst performance can be captured by the network; the hidden layers are set to four layers, and the number of layers is determined... The basis is the nonlinear relationship complexity between catalyst features and performance. The number of neurons in the first hidden layer is twice that of the input layer, aiming to initially expand the feature dimension and capture simple nonlinear relationships. The number of neurons in the second hidden layer is consistent with that in the first layer, further deepening feature extraction. The number of neurons in the third hidden layer is half that of the first layer, achieving feature dimension compression and focusing on key information. The number of neurons in the fourth hidden layer is the same as that in the third layer, strengthening the integration of higher-order features. The number of neurons in the output layer corresponds one-to-one with the catalyst performance prediction target. If it is necessary to predict the three core performances of catalytic activity, selectivity, and long-term stability at the same time, the output layer is set with three neurons, each corresponding to one performance indicator.

[0077] The nonlinear transformation rules for each layer of the network were specifically adapted. A nonlinear transformation rule with gradient propagation optimization characteristics was used between the input layer and the first hidden layer. This rule effectively alleviates the gradient vanishing problem, ensures the gradient propagation efficiency of deep networks, and is suitable for the large range of some features in the catalyst data. A symmetrical output nonlinear transformation rule was used between the first and second hidden layers. Its output range exhibits a symmetrical distribution and possesses strong nonlinear expressive power, capable of capturing complex interactions between catalyst features. A negative interval gradient-preserving nonlinear transformation rule was used between the second and third hidden layers. By preserving small gradients in negative intervals, this rule addresses the neuron inactivation problem that some transformation rules might cause, ensuring the network's ability to capture niche features of the catalyst. A composite nonlinear transformation rule was used between the third and fourth hidden layers, combining the advantages of gradient optimization and probabilistic transformation to further enhance the flexibility of feature nonlinear transformation. A numerical range-adaptive transformation rule was used between the fourth hidden layer and the output layer to ensure that the output results cover the actual numerical range of catalyst performance, meeting the prediction accuracy requirements.

[0078] Next, initial parameter configuration is performed. Following the order of layer, parameter type, and position, the obtained initial global parameter values ​​are loaded one by one into the corresponding positions of the network. First, all connection strength parameters between the input layer and the first hidden layer are configured, then the bias parameters of all neurons in the first hidden layer are configured. This process is repeated sequentially, configuring the connection strength parameters between the first and second hidden layers, the second and third hidden layers, the third and fourth hidden layers, and the fourth hidden layer and the output layer, as well as the bias parameters of neurons in each hidden and output layer. During configuration, parameter values ​​are checked in real-time to ensure they are within a reasonable range. If any parameters exceed the range, the fusion process is immediately backtracked to investigate the problem and ensure accurate loading of initial parameters. Then, multi-layer nonlinear transformations and higher-order feature extraction are performed, using standardized input data. The multi-node catalyst integrated dataset is processed. This dataset integrates high-quality feature data from all distributed nodes after validity verification and has been standardized in terms of dimensions. The data is first fed into the input layer and then passed to the first hidden layer without transformation. Each neuron in the first hidden layer performs a weighted summation operation on the input features. The output value of the corresponding neuron in the input layer is multiplied by the connection strength parameter between that neuron and the current hidden layer neuron. All the product results are then added together, and finally the bias parameter of the current hidden layer neuron is added to obtain the linear combination result. Subsequently, the linear combination result is fed into a nonlinear transformation rule with gradient propagation optimization characteristics. Through the nonlinear transformation of this rule, the basic features that have a significant impact on catalyst performance are screened out, and the feature output of the first hidden layer is obtained.

[0079] The feature output is passed to the second hidden layer, where it is processed by weighted summation and symmetric output nonlinear transformation rules to further explore the interaction relationships between features and extract more complex intermediate features. The intermediate features are passed to the third hidden layer, where they are transformed by weighted summation and negative interval gradient preservation nonlinear transformation rules to solve the problem of some niche features being suppressed and retain more potential effective information. The features output from the third layer are passed to the fourth hidden layer, where they are deeply integrated and optimized by weighted summation and nonlinear transformation of composite nonlinear transformation rules to finally form a high-order feature representation.

[0080] Based on higher-order feature representations, the difference between the output of a global deep neural network and the true value is calculated, and the parameter configuration of the global deep neural network is adjusted according to the difference. Specifically, this involves: first, obtaining the true value corresponding to the higher-order feature representation. The true value is the actual performance data of the catalyst sample obtained through experimental testing, including activity, selectivity, proportion, stability, and duration, ensuring a one-to-one correspondence between the true value and the prediction target of the network output layer. Then, the difference between the output and the true value is calculated using a comprehensive error index. The calculation method is to divide the sum of the absolute differences between the network output value and the corresponding true value of all samples by the total number of samples to obtain the mean absolute error. Simultaneously, the sum of the squared differences between the output value and the true value of each sample is divided by the total number of samples to obtain the mean squared error. These two error indices comprehensively reflect the degree of deviation in the network prediction. Next, the network parameters are adjusted according to the difference. The adjustment process proceeds backward layer by layer from the output layer. First, the error contribution of the output layer is calculated, and the comprehensive error is then divided according to the output layer... The connection strength of neurons is proportionally allocated to obtain the error term for each neuron. Based on this error term, the connection strength parameter between the output layer and the third hidden layer is adjusted. The adjustment direction is as follows: if the error term is positive, the corresponding connection strength parameter is appropriately decreased; if it is negative, the corresponding connection strength parameter is appropriately increased. At the same time, the bias parameter of the output layer neurons is adjusted to make the error term approach zero. Then, the error term of the output layer is propagated back to the third hidden layer. Combined with the derivative of the activation function of this layer, the error contribution of the third hidden layer is calculated. Similarly, the connection strength parameter between the third hidden layer and the second hidden layer and the bias parameter of the third hidden layer neurons are adjusted. Following the same logic, the error is propagated back in reverse, and the connection strength parameters between the second and first hidden layers, the first hidden layer and the input layer, and the bias parameters of each hidden layer neuron are adjusted respectively. The adjustment range of the parameters for each layer is positively correlated with the magnitude of the error contribution. The larger the error contribution, the larger the adjustment range, ensuring that the parameter adjustment is targeted at solving the prediction bias problem.

[0081] The parameter adjustment process is repeated until the prediction of the global deep neural network reaches the preset requirements, and a globally collaborative model with the adjustments completed is constructed. Specifically, the preset requirements include two core conditions: first, the mean absolute error and mean square error in the comprehensive error index are both within a preset reasonable range. This range is determined based on the prediction needs in the catalyst research and development field, ensuring that the prediction results meet the reference requirements for actual catalyst screening; second, in three consecutive iterations, the change in the comprehensive error index is less than a preset threshold, indicating that the network parameters have stabilized and further adjustments cannot significantly improve performance. After each parameter adjustment, the multi-node comprehensive dataset is divided into a training subset and a validation subset according to a preset ratio. The training subset is used to perform forward and backward propagation again to update the parameters, and the validation subset is used to perform forward and backward propagation again. The prediction performance of the network is validated, a comprehensive error index is calculated, and it is determined whether the preset requirements are met. If the preset requirements are not met, a multi-layer nonlinear transformation is performed again based on the adjusted parameters to extract higher-order features, and then difference calculation and parameter adjustment are performed. This process is repeated iteratively. During the iteration process, the comprehensive error index and parameter adjustment status of each iteration are recorded in real time. If the error index first decreases and then increases, the current iteration direction is stopped in time, and the parameter adjustment range is adjusted to avoid overfitting. When the comprehensive error index meets the preset requirements during the iteration process, parameter adjustment is stopped. At this time, the parameter configuration of the network reaches the optimal state, including the connection strength parameters between each layer and the bias parameters of all neurons, which are stable. The global deep neural network in this state is determined as the final global collaborative model.

[0082] A global deep neural network adapted to the characteristics of catalyst data is constructed. High-order features are extracted through multi-layer nonlinear transformations, effectively capturing the complex correlation between low-dimensional features and performance of catalysts. Based on error back-adjustment of parameters and iterative optimization, the stability and accuracy of the global collaborative model are ensured.

[0083] In a preferred embodiment of the present invention, predicting and screening catalyst performance based on a global collaborative model to obtain adjusted catalyst screening results may include:

[0084] In this embodiment of the invention, the characteristic data of the catalyst samples to be screened are input into a global collaborative model to obtain the performance prediction values ​​of each catalyst sample. Specifically, the characteristic data of the catalyst samples to be screened covers the core composition characteristics, structural characterization characteristics, and process adaptation characteristics of the catalyst. The core composition characteristics include the type and content of the active components, the type and ratio of the support, and the addition ratio of the co-catalyst, etc. The structural characterization characteristics include the specific surface area, pore size distribution, crystal structure, and grain size of the catalyst, etc. The process adaptation characteristics include key process parameters such as reaction temperature, reaction pressure, and reaction time suitable for the sample. Before inputting the data into the global collaborative model, these characteristic data need to undergo preprocessing operations consistent with the training phase of the global collaborative model. First, the completeness of the characteristic data is checked to ensure that all key characteristic fields of each sample are not missing. If there are missing fields, they are supplemented by the characteristic average of similar samples. After supplementation, the rationality of the data is checked again. Next, the dimensions are standardized. Referring to the standardization rules used during the training of the global collaborative model, the same mean and standard deviation as the training data are used to adjust each feature data. For example, the active component content data is offset and scaled according to the mean and standard deviation of the training data to ensure that the dimensions of the data to be screened are consistent with the training data of the global collaborative model, avoiding prediction bias due to differences in dimensions. After preprocessing, the feature data of each sample to be screened are arranged into a data sequence according to a preset order. The order is completely consistent with the feature dimensions corresponding to the neurons in the input layer of the global collaborative model to ensure that the features are accurately matched with the input nodes of the global collaborative model. Then, the arranged data sequence is input into the global collaborative model one by one. After receiving the data, the global collaborative model passes it to each hidden layer through the input layer. Through multi-layer nonlinear transformation, the features are deeply processed to extract the complex correlation information between features. Finally, the corresponding performance prediction value is output through the output layer.

[0085] Based on the predicted performance values, all catalyst samples are ranked from high to low to generate a ranking list. Specifically, this includes: first, determining the calculation rules for the comprehensive performance score. These rules need to take into account the actual needs of catalyst research and development, comprehensively consider the importance of three core performance indicators, and have each distributed node jointly negotiate and determine the importance weight coefficient of each indicator. Among them, catalytic activity, as the core performance indicator, has the highest importance weight coefficient, followed by selectivity, and then long-term stability. The sum of the importance weight coefficients of the three indicators is 100%. The comprehensive performance score is calculated by multiplying the predicted catalytic activity value of each sample by the importance ratio coefficient of catalytic activity, adding the predicted selectivity value multiplied by the importance ratio coefficient of selectivity, and adding the predicted long-term stability value multiplied by the importance ratio coefficient of long-term stability. This weighted calculation yields the comprehensive performance score for each sample, ensuring that the score comprehensively reflects the overall performance level of the sample. After calculation, the comprehensive performance scores of all catalyst samples to be screened are collected and sorted in descending order of score. If two or more samples have the same comprehensive performance score, the predicted catalytic activity values ​​are further compared, with samples having higher predicted catalytic activity values ​​ranked higher. If the predicted catalytic activity values ​​are still the same, the predicted selectivity values ​​are compared, with samples having higher predicted selectivity values ​​ranked higher. If the predicted selectivity values ​​are also the same, the predicted long-term stability values ​​are compared, with samples having higher predicted long-term stability values ​​ranked higher. This ensures that the ranking results are unique and based on clear evidence. After ranking, a detailed ranking list is generated, containing each sample's unique identifier, predicted performance values, comprehensive performance score, and ranking. The list also records the original characteristic data of the samples, including complete information on core component characteristics, structural characterization characteristics, and process adaptability characteristics.

[0086] Based on the ranking list, catalyst samples with predicted values ​​greater than preset performance thresholds are selected to form preliminary screening results. Specifically, the determination of preset performance thresholds needs to consider the performance range of known effective catalysts according to industry standards and the target requirements of catalyst research and development. It is set in two levels: the first level is the individual performance threshold, namely the catalytic activity threshold, selectivity threshold, and long-term stability threshold. Each threshold is set with reference to the lowest value of the corresponding performance index of known effective catalysts, while also considering the minimum requirements for each performance aspect in the research and development goals, ensuring that the selected samples meet the basic application requirements in terms of individual performance. The second level is the comprehensive performance threshold, which is determined by considering the performance range of known effective catalysts. The average comprehensive performance score of the catalyst is determined in conjunction with the improvement target of the research and development to ensure that the overall performance of the screened samples reaches the expected level. In the screening process, each sample in the ranking list is first checked for individual performance. The predicted value of catalytic activity is checked to see if it is greater than the catalytic activity threshold, the predicted value of selectivity is greater than the selectivity threshold, and the predicted value of long-term stability is greater than the long-term stability threshold. Only samples that meet the threshold requirements for all three individual performances are allowed to proceed to the next step of comprehensive performance verification. Then, the samples that pass the individual verification are checked for comprehensive performance to see if their comprehensive performance score is greater than the comprehensive performance threshold. If it is greater, the sample is included in the preliminary screening results.

[0087] The preliminary screening results were compared with a dataset of known effective catalysts to verify the reliability of the screening results and obtain comparative analysis results. Specifically, the comparative analysis was conducted from two core dimensions: feature similarity and performance consistency. For feature similarity comparison, for each sample in the preliminary screening results, multiple samples with the most similar features were selected from the dataset of known effective catalysts as comparison benchmarks. The selection criterion was minimizing the sum of differences in core component features and structural characterization features. The feature differences between the sample to be screened and each comparison benchmark sample were calculated by taking the absolute value of the numerical difference for each corresponding feature dimension and then summing the absolute differences of all feature dimensions to obtain the total difference value. The smaller the total difference value, the higher the feature similarity. For performance consistency comparison, the predicted values ​​of various performance indicators of the sample to be screened were compared with the actual performance data of the comparison benchmark samples. The difference between the predicted value and the actual value of each performance indicator was calculated. The percentage of difference is calculated by dividing the difference by the actual performance value of the benchmark sample to obtain the percentage of difference for each performance item. Simultaneously, the percentage of difference in the overall performance score is calculated to comprehensively assess the consistency of performance predictions. Based on the comparison results, three reliability levels are assigned: Level 1 is highly reliable, with the total difference value of characteristics in a low range and the percentage of difference for each performance item within a reasonable range, indicating that the sample to be screened is highly compatible with known effective catalysts in terms of characteristics and performance; Level 2 is basically reliable, with the total difference value of characteristics in a moderate range and the percentage of difference for core performance indicators within a reasonable range, while minor performance indicators show slight deviations that do not affect overall application; Level 3 is unreliable, with a large total difference value of characteristics or the percentage of difference for multiple performance indicators exceeding the reasonable range, indicating that the prediction results may be biased. Detailed records of the comparative analysis results for each sample are required, including detailed identification feature differences, detailed performance differences, and the final reliability level.

[0088] Based on the comparative analysis results, the preliminary screening results were adjusted to obtain the final catalyst screening results. Specifically, the adjustment process was conducted based on the reliability levels derived from the comparative analysis using AI intelligent decision-making logic, while also considering the actual application requirements of catalyst development. Samples deemed highly reliable were directly retained in the screening results without adjustment, as their predicted performance closely matched known effective catalysts and possessed high practical application value. For samples deemed basically reliable, the source of performance differences was further investigated. If the differences stemmed from minor variations in secondary characteristics and did not affect the core application scenario, the sample was retained, and the specific details of the performance differences and their potential impact were noted in the results. If the differences might lead to specific... Performance fluctuations in application scenarios are evaluated in conjunction with R&D goals. If the application scenario is a core requirement, the sample is removed; if it is not a core requirement, it is retained, and targeted process adjustment suggestions are proposed. For samples deemed unreliable, if they exist, they are corrected and re-predicted, screened, and compared. If they are still unreliable after ruling out the above issues, they are removed from the initial screening results, and the reasons for removal are recorded, including specific dimensions with excessively large feature differences or specific indicators with excessive performance prediction deviations. After adjustment, the retained samples are re-verified to check whether the sample information is complete, whether the performance prediction values ​​all meet the preset thresholds, and whether the comparative analysis results are accurate, ensuring no omissions or misjudgments. Finally, the final catalyst screening results are compiled.

[0089] A tiered, preset performance threshold system is used for initial screening, ensuring that individual core performance indicators of the catalysts meet the standards while also considering the overall performance level. This effectively filters out samples that do not meet the basic requirements, improving screening efficiency. Multi-dimensional comparative analysis with known effective catalyst datasets verifies the reliability of the screening results and reduces the risk of misjudgment due to prediction bias.

[0090] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0091] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0092] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the 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 distributed catalyst data collaborative modeling and privacy-preserving screening system based on AI learning, characterized in that, include: The acquisition module is used to acquire catalyst-related data from multiple distributed data sources, and to clean and standardize the catalyst-related data to obtain a standardized dataset, including the basic properties of the catalyst, preparation process parameters, and performance test data. The generation module is used to analyze the data distribution characteristics based on the standardized dataset, determine the core reference set, and generate two sets of reference dimensions based on the core reference set to determine the scope of feature analysis. A validation sample set is selected both inside and outside the feature analysis scope. The distribution density of the validation samples in the feature space is calculated to identify high-density regions as key evolutionary nodes, constructing data association paths and generating data calibration parameters. This includes calculating and analyzing the local density and relative distance of data points in the standardized dataset to identify data points with high local density in the data space. Based on these high-density data points, high-density regions in the data distribution are determined, and the center points of each high-density region are calculated to form a core reference set. Using the core reference set as the analysis object, principal component analysis is used to extract the main direction of data change, resulting in two mutually orthogonal principal component vectors. These two mutually orthogonal principal component vectors are defined as the first and second reference dimensions, respectively, to determine the scope of feature analysis. The first reference dimension is the principal direction with the largest data variance, used to characterize the most important distribution feature of the data. The second reference dimension is the principal direction with the second largest data variance and is orthogonal to the first reference dimension, used to characterize the secondary distribution feature of the data. Sample points are randomly selected within and outside the feature analysis scope according to a preset ratio to form an initial validation sample set. Kernel density estimation is performed on the feature space of the initial validation sample set, constructed using the reference dimensions, to obtain the probability density estimate for each sample point. The probability density estimate is compared with a preset density threshold, and all sample points greater than the threshold are selected. The regions where the sample points are located are marked as key evolution nodes. Based on the key evolution nodes, the distance between any two nodes is calculated as the connection cost. Connections between nodes are established in order of increasing connection cost, ensuring that no loops are formed, until all nodes are connected and the overall connection cost is minimized, forming a data association path. The density change rate between adjacent key evolution nodes is calculated along the data association path to obtain the path density gradient. Based on the path density gradient, data calibration parameters for adjusting and calibrating the data distribution are generated. The training module is used to calibrate parameters using data, deploy deep neural networks on each distributed node for local training, and generate local model parameter adjustment information. The protection module is used to encrypt the parameter adjustment information based on the local model parameter adjustment information using differential privacy to obtain the privacy-protected parameter information; The module is used to comprehensively consider the contribution of each node when fusing multi-node parameter adjustment information based on privacy-preserving parameter information, and to build a global collaborative model by obtaining complex nonlinear relationships between data through deep neural networks. The screening module is used to predict and screen catalyst performance based on the global collaborative model, and obtain adjusted catalyst screening results.

2. The AI-based distributed catalyst data collaborative modeling and privacy-preserving screening system according to claim 1, characterized in that, By using data to calibrate parameters, deep neural networks are deployed on each distributed node for local training, generating local model parameter adjustment information, including: The data calibration parameters are used as input layer preprocessing coefficients and applied to the input layer of the deep neural network of each distributed node to calibrate the input features and obtain calibrated input features. Based on the calibrated input features, the forward propagation calculation of the deep neural network is performed on each distributed node using the local dataset to obtain the network output results. Based on the network output, the difference between the network output and the real label is calculated to obtain the output error. Based on the output error, the error contribution of each layer is calculated backward from the output layer to obtain the gradient information of the parameters of each layer. Based on the gradient information of each layer's parameters, the adjustment amounts of the connection strength parameters and bias parameters are calculated. The network parameters are gradually adjusted through multiple rounds of update operations, and the update amounts of the connection strength parameters and the adjustment values ​​of the bias parameters are recorded and integrated into local model parameter adjustment information.

3. The AI-based distributed catalyst data collaborative modeling and privacy-preserving screening system according to claim 2, characterized in that, Based on local model parameter adjustment information, differential privacy is used to encrypt the parameter adjustment information, resulting in privacy-protected parameter information, including: The numerical distribution range of local model parameter adjustment information is analyzed, and the global sensitivity of the parameters is determined based on the numerical distribution range. Based on global sensitivity and combined with preset privacy budget parameters, the scale parameters of Laplace noise are calculated. Based on the scale parameter, a random noise sequence conforming to the Laplace distribution is generated, and the random noise sequence is superimposed with the local model parameter adjustment information to obtain the parameter information after noise superposition. The parameter information after noise superposition is standardized by numerical range processing to obtain standardized parameter information. The standardized parameter information is then verified to meet the differential privacy protection requirements to obtain privacy-protected parameter information.

4. The AI-based distributed catalyst data collaborative modeling and privacy-preserving screening system according to claim 3, characterized in that, Based on privacy-preserving parameter information, the contribution of each node is comprehensively considered when fusing multi-node parameter adjustment information. A global collaborative model is constructed by acquiring complex nonlinear relationships between data through a deep neural network, including: Based on the privacy-preserving parameter information of each node, the amount of local training data and the evaluation results of the local model for each node are statistically analyzed. Based on the amount of data and the evaluation results, the contribution ratio of each node in parameter fusion is calculated, and based on the contribution ratio, the privacy-protected parameter information of each node is proportionally allocated and fused to obtain the initial value of the global parameters. Based on the initial values ​​of global parameters, the initial parameters of the global deep neural network are configured, and based on the initial parameters, the input data is subjected to multi-layer nonlinear transformation through the global deep neural network to extract high-order feature representations between data. Based on higher-order feature representation, the difference between the output of the global deep neural network and the true value is calculated, and the parameter configuration of the global deep neural network is adjusted according to the difference. Repeat the parameter adjustment process until the prediction of the global deep neural network reaches the preset requirements, and build the adjusted global collaborative model.

5. The AI-based distributed catalyst data collaborative modeling and privacy-preserving screening system according to claim 4, characterized in that, Based on the global collaborative model, catalyst performance is predicted and screened, resulting in adjusted catalyst screening results, including: The characteristic data of the catalyst samples to be screened are input into the global collaborative model to obtain the performance prediction values ​​of each catalyst sample. Based on the performance prediction values, all catalyst samples are sorted from high to low performance to generate a sorted list. Based on the sorted list, catalyst samples with predicted values ​​greater than the preset performance threshold are selected to form preliminary screening results. The preliminary screening results were compared and analyzed with known effective catalyst datasets to verify the reliability of the screening results and obtain the comparative analysis results. Based on the comparative analysis results, the preliminary screening results were adjusted to obtain the final catalyst screening results.

6. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the system as described in any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the system as described in any one of claims 1 to 5.

Images