A Deep Learning-Based Method and System for Optimizing Dimethyl Oxalate Catalyst Formulation

Abstract

This invention provides a method and system for optimizing the formulation of dimethyl oxalate catalysts based on deep learning, belonging to the field of formulation optimization technology. The method involves constructing a deep neural network model, employing a multi-task learning strategy to simultaneously optimize catalytic conversion and stability, and utilizing an improved Bayesian optimization algorithm to introduce component compatibility constraints for solving the problem, thereby determining the optimal catalyst formulation. This method can improve the accuracy of catalyst performance prediction, accelerate catalyst optimization iteration, and significantly improve catalytic conversion and service life.

Description

Technical Field

[0001] This invention relates to formulation optimization technology, and more particularly to a method and system for optimizing the formulation of dimethyl oxalate catalysts based on deep learning. Background Technology

[0002] Dimethyl oxalate is an important chemical intermediate widely used in pharmaceuticals, pesticides, dyes, and polymer materials. Optimizing its catalytic synthesis process is crucial for improving product yield, reducing production costs, and minimizing environmental pollution. Traditional optimization of dimethyl oxalate catalyst formulations relies heavily on experimental trial and error and experience accumulation, typically requiring numerous experiments to obtain a suitable catalyst formulation.

[0003] In recent years, with the development of computer technology and data analysis methods, catalyst design methods based on statistical models have been gradually applied. These methods typically employ traditional statistical methods such as response surface methodology and orthogonal experimental design, or simple machine learning algorithms such as linear regression and support vector machines, to establish a model of the relationship between catalyst components and performance. However, these methods have significant limitations when dealing with complex nonlinear systems such as the dimethyl oxalate catalytic reaction.

[0004] The defects and shortcomings of existing technologies are mainly reflected in the following aspects:

[0005] First, traditional catalyst optimization methods struggle to effectively capture the complex synergistic effects between catalyst components, especially when the catalyst contains multiple active components and promoters, making it difficult to accurately model the impact of interactions between components on catalytic performance.

[0006] Secondly, existing optimization methods often only focus on the performance index of catalytic conversion rate, while ignoring the optimization of key factors such as catalyst stability. This results in catalysts with high initial activity but short service life, making it difficult to meet the requirements of long-term operation in actual industrial applications.

[0007] Finally, existing catalyst optimization methods lack a systematic consideration of component compatibility, which may lead to problems such as component phase separation and reduced activity in the actual preparation process of theoretically optimal formulations, thus affecting the actual application effect of the catalyst. Summary of the Invention

[0008] This invention provides a method and system for optimizing the formulation of dimethyl oxalate catalysts based on deep learning, which can solve the problems in the prior art.

[0009] A first aspect of the present invention,

[0010] Historical experimental data of multiple groups of dimethyl oxalate catalysts were obtained, including catalyst group composition data, reaction process parameter data, and catalytic conversion rate data.

[0011] A deep neural network model is constructed, which includes an input layer, multiple hidden layers, and an output layer. The input layer receives catalyst group distribution data and reaction process parameter data, and the output layer outputs predicted catalytic conversion rate data.

[0012] The deep neural network model is trained using a multi-task learning strategy, and the catalytic conversion rate prediction task and the catalyst stability prediction task are optimized simultaneously during the training process. The catalyst stability is characterized by the decay rate of catalytic activity after multiple consecutive reactions.

[0013] Based on the trained deep neural network model, a catalyst formulation optimization objective function is constructed, with catalytic conversion rate and catalyst stability as optimization objectives.

[0014] An improved Bayesian optimization algorithm is used to optimize and solve the objective function. In the Bayesian optimization process, component compatibility constraints are introduced. The sampling points are screened by a pre-established component compatibility knowledge base to ensure that the optimized catalyst formulation components have good compatibility.

[0015] Based on the Bayesian optimization results, the optimal formulation of the dimethyl oxalate catalyst was determined, including the proportions of each component and the corresponding reaction process parameters.

[0016] The deep neural network model includes an input layer, multiple hidden layers, and an output layer. The input layer receives catalyst composition data and reaction process parameter data, and the output layer outputs predicted catalytic conversion rate data, including:

[0017] Multiple input neurons are set in the input layer. The multiple input neurons are used to receive the proportion data of each component of the catalyst. The proportion data of the catalyst components includes the proportion data of the main catalyst, the proportion data of the co-catalyst, the proportion data of the support material, and the reaction process parameter data in the reaction temperature data, reaction pressure data, and reaction time data.

[0018] A residual connection structure is constructed in the multiple hidden layers, and a cross-layer information transmission path is established between adjacent hidden layers through the residual connection structure;

[0019] An output neuron is provided in the output layer, and the output neuron is used to output the predicted catalytic conversion rate data;

[0020] An attention mechanism is used to dynamically weight the input data of the multiple input neurons, and the feature weights are adaptively adjusted according to the degree of influence of different input features on the catalytic conversion rate. The attention weights are obtained by alternating forward and backward propagation optimization.

[0021] The deep neural network model is trained using a multi-task learning strategy, and the catalytic conversion rate prediction task and the catalyst stability prediction task are optimized simultaneously during the training process, including:

[0022] A dual-task neural network structure is constructed, which includes a feature extraction layer, a catalytic conversion rate prediction branch, and a catalyst stability prediction branch, wherein the feature extraction layer is used to extract feature information from the training input data;

[0023] Establish a task weight allocation mechanism to dynamically adjust the weight ratio of the two tasks based on the training loss values ​​of the catalytic conversion rate prediction task and the catalyst stability prediction task.

[0024] A task association module is constructed, which is used to establish a feature mapping relationship between the catalytic conversion prediction branch and the catalyst stability prediction branch, so that the two prediction tasks can share and utilize their respective learning features.

[0025] The dual-task neural network structure is trained using a multi-task joint optimization method, and the network parameters of the feature extraction layer, the catalytic conversion rate prediction branch, and the catalyst stability prediction branch are updated simultaneously.

[0026] When the prediction errors of both the catalytic conversion rate prediction task and the catalyst stability prediction task are less than the preset error threshold, the trained dual-task neural network model is obtained.

[0027] An improved Bayesian optimization algorithm is used to optimize and solve the objective function. The component compatibility constraint is introduced during the Bayesian optimization process, and the sampling points are screened using a pre-established component compatibility knowledge base, including:

[0028] A catalyst component compatibility knowledge base is established, which includes compatibility rule data and component ratio restriction data among the main catalyst, co-catalyst and support material;

[0029] An objective function for catalyst formulation optimization is constructed, wherein the objective function includes a catalytic conversion parameter and a component compatibility parameter calculated based on the compatibility rule data;

[0030] A compatibility constraint function is constructed based on the component compatibility knowledge base. The compatibility constraint function is used to calculate the compatibility score under different component ratios, and the feasibility of the formulation is determined based on the compatibility score. The formulation with a compatibility score greater than a preset threshold is determined to be a feasible formulation.

[0031] The compatibility constraint function is integrated into the Bayesian-optimized Gaussian process regression model to establish an optimization model that considers component compatibility.

[0032] Bayesian optimization search is performed based on the optimization model; the sampling points and their objective function values ​​obtained in each iteration are fed back to the Gaussian process regression model to update the prior information of the optimization model;

[0033] When the number of optimization iterations reaches a preset value or the objective function value converges, the catalyst formulation that satisfies the compatibility constraint and has the optimal objective function value is output as the final optimization result.

[0034] Bayesian optimization search based on the aforementioned optimization model includes:

[0035] In each iteration, a set of candidate sampling points is first generated, and then the compatibility constraint function is used to filter the candidate sampling points to obtain a set of feasible sampling points that meet the compatibility requirements. The sampling point with the largest expected improvement is selected from the set of feasible sampling points for evaluation.

[0036] The proportions of each component include:

[0037] Optimal mass percentage of alkaline oxide main catalyst; optimal mass percentage of metal oxide co-catalyst; optimal mass percentage of porous support material;

[0038] The reaction process parameters include:

[0039] Optimal reaction temperature range; optimal reaction pressure range; optimal space velocity range.

[0040] A second aspect of the present invention provides a deep learning-based system for optimizing the formulation of dimethyl oxalate catalysts, comprising:

[0041] The first unit is used to acquire historical experimental data of multiple groups of dimethyl oxalate catalysts, including catalyst group distribution data, reaction process parameter data, and catalytic conversion rate data.

[0042] The second unit is used to construct a deep neural network model, which includes an input layer, multiple hidden layers and an output layer. The input layer receives catalyst group distribution data and reaction process parameter data, and the output layer outputs predicted catalytic conversion rate data.

[0043] The third unit is used to train the deep neural network model using a multi-task learning strategy, and simultaneously optimize the catalytic conversion rate prediction task and the catalyst stability prediction task during the training process, wherein the catalyst stability is characterized by the decay rate of catalytic activity after multiple consecutive reactions.

[0044] The fourth unit is used to construct a catalyst formulation optimization objective function based on the trained deep neural network model, with catalytic conversion rate and catalyst stability as optimization objectives.

[0045] The fifth unit is used to optimize and solve the objective function using an improved Bayesian optimization algorithm. In the Bayesian optimization process, component compatibility constraints are introduced, and sampling points are screened through a pre-established component compatibility knowledge base to ensure that the optimized catalyst formulation components have good compatibility.

[0046] The sixth unit is used to determine the optimal formulation of the dimethyl oxalate catalyst based on the Bayesian optimization results, including the proportions of each component and the corresponding reaction process parameters.

[0047] A third aspect of the present invention,

[0048] An electronic device is provided, comprising:

[0049] processor;

[0050] Memory used to store processor-executable instructions;

[0051] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0052] Fourth aspect of the embodiments of the present invention,

[0053] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0054] The beneficial effects of this application are as follows: Beneficial effects

[0055] 1. By applying deep learning models, the catalytic conversion rate and stability of dimethyl oxalate catalysts can be effectively predicted, thereby improving the performance of the catalysts.

[0056] 2. The multi-task learning strategy enables simultaneous optimization of catalytic conversion rate and catalyst stability, thereby improving the overall reliability and lifespan of the catalyst.

[0057] 3. An improved Bayesian optimization algorithm and component compatibility constraints are introduced to ensure that the optimized catalyst formulation has good compatibility, thereby improving the actual application effect of the catalyst. Attached Figure Description

[0058] Figure 1 This is a schematic flowchart of the deep learning-based method for optimizing the formulation of dimethyl oxalate catalyst, as described in an embodiment of the present invention.

[0059] Figure 2 This is a flowchart illustrating the training process of the dual-task neural network model according to an embodiment of the present invention.

[0060] Figure 3 This is a complete flowchart of catalyst formulation optimization in an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0063] Figure 1 This is a schematic flowchart of the deep learning-based method for optimizing the formulation of dimethyl oxalate catalyst, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0064] Historical experimental data of multiple groups of dimethyl oxalate catalysts were obtained, including catalyst group composition data, reaction process parameter data, and catalytic conversion rate data.

[0065] A deep neural network model is constructed, which includes an input layer, multiple hidden layers, and an output layer. The input layer receives catalyst group distribution data and reaction process parameter data, and the output layer outputs predicted catalytic conversion rate data.

[0066] The deep neural network model is trained using a multi-task learning strategy, and the catalytic conversion rate prediction task and the catalyst stability prediction task are optimized simultaneously during the training process. The catalyst stability is characterized by the decay rate of catalytic activity after multiple consecutive reactions.

[0067] Based on the trained deep neural network model, a catalyst formulation optimization objective function is constructed, with catalytic conversion rate and catalyst stability as optimization objectives.

[0068] An improved Bayesian optimization algorithm is used to optimize and solve the objective function. In the Bayesian optimization process, component compatibility constraints are introduced. The sampling points are screened by a pre-established component compatibility knowledge base to ensure that the optimized catalyst formulation components have good compatibility.

[0069] Based on the Bayesian optimization results, the optimal formulation of the dimethyl oxalate catalyst was determined, including the proportions of each component and the corresponding reaction process parameters.

[0070] In one optional embodiment, the deep neural network model includes an input layer, multiple hidden layers, and an output layer, wherein the input layer receives catalyst composition data and reaction process parameter data, and the output layer outputs predicted catalytic conversion rate data, including:

[0071] Multiple input neurons are set in the input layer. The multiple input neurons are used to receive the proportion data of each component of the catalyst. The proportion data of the catalyst components includes the proportion data of the main catalyst, the proportion data of the co-catalyst, the proportion data of the support material, and the reaction process parameter data in the reaction temperature data, reaction pressure data, and reaction time data.

[0072] A residual connection structure is constructed in the multiple hidden layers, and a cross-layer information transmission path is established between adjacent hidden layers through the residual connection structure;

[0073] An output neuron is provided in the output layer, and the output neuron is used to output the predicted catalytic conversion rate data;

[0074] An attention mechanism is used to dynamically weight the input data of the multiple input neurons, and the feature weights are adaptively adjusted according to the degree of influence of different input features on the catalytic conversion rate. The attention weights are obtained by alternating forward and backward propagation optimization.

[0075] A multi-layer neural network architecture was constructed, consisting of an input layer, four hidden layers, and an output layer. The input layer has 15 input neurons, each receiving different types of input data: 8 for the main catalyst composition data (such as the mass percentage of metal elements like Ni, Mo, W, V, Co, Cu, Fe, and Pt), 4 for the co-catalyst composition data (such as the mass percentage of rare earth elements like La, Ce, Pr, and Nd), 1 for the support material composition data (such as the mass percentage of γ-Al₂O₃, SiO₂, and ZrO₂), and 2 for reaction process parameters (reaction temperature and reaction pressure).

[0076] All input data are normalized to map data of different dimensions to the interval [0,1]. For example, if the Ni content in the catalyst is 15% and the maximum possible content is 30%, the normalized value is 0.5; if the reaction temperature is 350℃, the minimum possible temperature is 200℃ and the maximum possible temperature is 500℃, the normalized value is also 0.5.

[0077] The first hidden layer contains 64 neurons, the second hidden layer contains 128 neurons, the third hidden layer contains 128 neurons, and the fourth hidden layer contains 64 neurons. Residual connections are constructed between adjacent hidden layers. Specifically, the output of the first hidden layer is not only passed to the second hidden layer as input, but also directly added to the input of the third hidden layer via a residual connection; similarly, the output of the second hidden layer is not only passed to the third hidden layer, but also directly added to the input of the fourth hidden layer via a residual connection. This cross-layer information transfer path effectively alleviates the gradient vanishing problem during deep neural network training, improving the model's training efficiency and prediction accuracy.

[0078] An attention mechanism is introduced to dynamically weight each input feature. The implementation steps of the attention mechanism are as follows: First, an attention coefficient is generated for each input feature, with initial values ​​set to be equal; then, during model training, these coefficients are adjusted through network learning. For example, when it is found that Ni content has a significant impact on catalytic conversion, its corresponding attention coefficient will increase, while for parameters with smaller impact, their attention coefficients will decrease accordingly. Taking actual data as an example, after sufficient training, the attention coefficient of Ni in the main catalyst may reach 0.23, while the attention coefficient of the support material may only be 0.08, indicating that Ni content is a key factor affecting the catalytic effect in this catalytic reaction.

[0079] The output layer has one output neuron to output the predicted catalytic conversion rate data, with a value ranging from 0 to 1, representing a conversion rate of 0% to 100%. A sigmoid activation function is used to ensure that the output value is within the valid range.

[0080] The network training employs batch gradient descent, selecting 32 samples per batch. Mean squared error (MSE) is used as the loss function to measure the difference between predicted and true values. The training process consists of two phases: forward propagation and backpropagation. In the forward propagation phase, input data passes through each hidden layer and finally reaches the output layer to obtain the prediction result. In the backpropagation phase, the loss is calculated based on the difference between the prediction result and the true value, and propagated backward along the network structure to update the weight parameters and attention coefficients of neurons in each layer. To prevent overfitting, a Dropout layer is added after each hidden layer, with a dropout rate set to 0.3.

[0081] The initial learning rate for model training was set to 0.001, and a learning rate decay strategy was adopted. When the model performance did not improve significantly for 5 consecutive training epochs, the learning rate was reduced to 0.5. The training process lasted for 100 epochs, or was terminated early when the loss on the validation set no longer decreased for 10 consecutive epochs.

[0082] In practical applications, 500 sets of catalytic experimental data were used for model training and validation, with 80% used for training and 20% for validation. Each set of data included catalyst composition information (e.g., Ni:Mo:W = 10:5:3, La content of co-catalyst 2%, γ-Al2O3 support content 80%), reaction process parameters (temperature 350℃, pressure 3MPa, reaction time 2h), and the corresponding catalytic conversion rate (e.g., 87.5%).

[0083] After complete training, the model's average prediction error on the validation set is less than 3%, accurately predicting catalytic conversion rates under different catalyst formulations and process conditions. For example, when the input main catalyst Ni:Mo:W ratio is 12:4:2, the co-catalyst Ce content is 1.5%, the support material γ-Al2O3 content is 82%, the reaction temperature is 370℃, and the reaction pressure is 3.2MPa, the model predicts a catalytic conversion rate of 91.2%, while the actual measured conversion rate is 90.8%, with a prediction error of only 0.4%.

[0084] This deep neural network model can significantly reduce the trial-and-error costs in the catalyst development process, accelerate the catalyst formulation optimization process, and provide an effective decision support tool for industrial catalyst design.

[0085] In one optional implementation, the deep neural network model is trained using a multi-task learning strategy, and the catalytic conversion rate prediction task and the catalyst stability prediction task are optimized simultaneously during the training process, including:

[0086] A dual-task neural network structure is constructed, which includes a feature extraction layer, a catalytic conversion rate prediction branch, and a catalyst stability prediction branch, wherein the feature extraction layer is used to extract feature information from the training input data;

[0087] Establish a task weight allocation mechanism to dynamically adjust the weight ratio of the two tasks based on the training loss values ​​of the catalytic conversion rate prediction task and the catalyst stability prediction task.

[0088] A task association module is constructed, which is used to establish a feature mapping relationship between the catalytic conversion prediction branch and the catalyst stability prediction branch, so that the two prediction tasks can share and utilize their respective learning features.

[0089] The dual-task neural network structure is trained using a multi-task joint optimization method, and the network parameters of the feature extraction layer, the catalytic conversion rate prediction branch, and the catalyst stability prediction branch are updated simultaneously.

[0090] When the prediction errors of both the catalytic conversion rate prediction task and the catalyst stability prediction task are less than the preset error threshold, the trained dual-task neural network model is obtained.

[0091] This layer is used to extract feature information from input data. Input data can include the chemical composition, physical properties, and catalytic reaction conditions of the catalyst. The feature extraction layer can employ a convolutional neural network (CNN) or a fully connected layer, depending on the type of input data. Through this layer, the input data undergoes multiple processing steps to extract a high-dimensional feature representation.

[0092] This branch receives feature information from the feature extraction layer and focuses on predicting the catalytic conversion rate. It can use a fully connected layer and ultimately outputs a scalar value representing the predicted catalytic conversion rate.

[0093] Similar to the catalytic conversion prediction branch, this branch also receives the output of the feature extraction layer and focuses on predicting catalyst stability. This branch also uses a fully connected layer and outputs a scalar value representing the predicted catalyst stability.

[0094] A dynamic task weight allocation mechanism is established to address the training loss values ​​of the catalytic conversion rate prediction and catalyst stability prediction tasks. The core idea of ​​this mechanism is to automatically adjust the weight ratio between the two tasks based on the current loss value of each task. For example, when the loss value of the catalytic conversion rate prediction task is high, its weight is increased to give it more attention during training; conversely, the weight is decreased. This mechanism can be implemented by monitoring the loss value of each task and adjusting the weights according to a preset strategy.

[0095] To achieve feature sharing between the catalytic conversion prediction branch and the catalyst stability prediction branch, a task association module is constructed. The main function of this module is to establish a feature mapping relationship between the two branches, allowing the learned features of one branch to influence the learning process of the other. This can be implemented using shared layers or cross-connection layers, enabling the two branches to mutually promote each other in feature learning.

[0096] During training, a multi-task joint optimization approach is used to update the network parameters of all layers in the dual-task neural network structure. The specific steps are as follows:

[0097] Initialize network parameters, and set the learning rate and optimizer (such as Adam or SGD).

[0098] In each training round, the input training data passes through the feature extraction layer to obtain feature representations of catalytic conversion rate and catalyst stability, respectively.

[0099] Calculate the loss values ​​for the two tasks and dynamically adjust the weights according to the task weight allocation mechanism.

[0100] By using the backpropagation algorithm, the network parameters are updated to ensure that the parameters of the feature extraction layer, the catalytic conversion prediction branch, and the catalyst stability prediction branch are all optimized.

[0101] The training process continues until the prediction errors for both the catalytic conversion rate prediction task and the catalyst stability prediction task are less than the preset error threshold. At this point, the dual-task neural network model can be considered to have completed training and possesses good predictive performance.

[0102] To verify the effectiveness of the above method, an experiment can be conducted using a set of real catalyst data. Assume there are 1000 sample data points, each containing information such as catalyst composition, physical properties, and reaction conditions. After training with the aforementioned dual-task neural network structure, the resulting model shows root mean square errors (RMSE) below a preset threshold (e.g., 0.05) in both catalytic conversion rate and catalyst stability predictions, indicating that the model has good predictive ability.

[0103] Figure 2 Here is a flowchart of the training process for the dual-task neural network model according to an embodiment of the present invention:

[0104] This method first constructs a dual-task neural network structure, including a feature extraction layer and two prediction branches (catalytic conversion rate prediction branch and catalyst stability prediction branch). The feature extraction layer is responsible for extracting key feature information from the training input data. Based on this, a task weight allocation mechanism is established, dynamically adjusting the task weight ratio by monitoring the training loss values ​​of the two prediction tasks in real time to ensure the balance of the training process. Simultaneously, a task association module is designed to establish the feature mapping relationship between the two prediction branches, realizing feature sharing and complementary learning. During training, a multi-task joint optimization strategy is adopted to synchronously update the parameters of each layer of the network, including the parameters of the feature extraction layer and the two prediction branches. The system observes the prediction error values ​​of the catalytic conversion rate prediction task and the catalyst stability prediction task. When the prediction errors of both tasks are below a preset threshold, training is considered complete, and the trained dual-task neural network model is finally output. The entire method achieves high-precision joint prediction of catalytic conversion rate and catalyst stability through mechanisms such as feature sharing, dynamic weight adjustment, and task association.

[0105] In one optional implementation, an improved Bayesian optimization algorithm is used to optimize and solve the objective function. The component compatibility constraint is introduced during the Bayesian optimization process, and the sampling points are screened using a pre-established component compatibility knowledge base, including:

[0106] A catalyst component compatibility knowledge base is established, which includes compatibility rule data and component ratio restriction data among the main catalyst, co-catalyst and support material;

[0107] An objective function for catalyst formulation optimization is constructed, wherein the objective function includes a catalytic conversion parameter and a component compatibility parameter calculated based on the compatibility rule data;

[0108] A compatibility constraint function is constructed based on the component compatibility knowledge base. The compatibility constraint function is used to calculate the compatibility score under different component ratios, and the feasibility of the formulation is determined based on the compatibility score. The formulation with a compatibility score greater than a preset threshold is determined to be a feasible formulation.

[0109] The compatibility constraint function is integrated into the Bayesian-optimized Gaussian process regression model to establish an optimization model that considers component compatibility.

[0110] Bayesian optimization search is performed based on the optimization model; the sampling points and their objective function values ​​obtained in each iteration are fed back to the Gaussian process regression model to update the prior information of the optimization model;

[0111] When the number of optimization iterations reaches a preset value or the objective function value converges, the catalyst formulation that satisfies the compatibility constraint and has the optimal objective function value is output as the final optimization result.

[0112] A catalyst component compatibility knowledge base needs to be constructed. This knowledge base should include compatibility rules and component ratio limitations data among the main catalyst, co-catalyst, and support material. Compatibility rules data can be collected through literature reviews, experimental data, and expert opinions to form a systematic database. For example, if a certain main catalyst performs well with a specific co-catalyst in a certain ratio, but may lead to a decrease in catalytic effect in other ratios, this information needs to be clearly recorded in the knowledge base.

[0113] An objective function needs to be constructed, which should include the catalytic conversion parameter and the component compatibility parameter calculated based on compatibility rules data from a component compatibility knowledge base. The catalytic conversion parameter can be obtained experimentally, while the component compatibility parameter is evaluated based on the corresponding compatibility scores in the knowledge base. The design of the objective function should ensure that it comprehensively considers both the performance and compatibility of the catalyst.

[0114] Based on the component compatibility knowledge base, a compatibility constraint function needs to be constructed. This function is used to calculate the compatibility score under different component ratios and to determine the feasibility of the formulation based on this score. Specifically, when the compatibility score of a formulation is higher than a preset threshold, the formulation is determined to be feasible. The calculation of the compatibility score can be based on historical data and experimental results to ensure its accuracy and reliability.

[0115] By integrating the compatibility constraint function into a Bayesian-optimized Gaussian process regression model, an optimization model considering component compatibility is established. This model can dynamically adjust the evaluation of sampling points during the optimization process, ensuring that the selected formulation not only performs well in catalytic conversion but also meets compatibility requirements.

[0116] Based on the established optimization model, a Bayesian optimization search is performed. In each iteration, the current sampling point and its objective function value are evaluated, and the results are fed back into the Gaussian process regression model to update the prior information of the optimization model. The performance of the sampling points should be recorded in each iteration for subsequent analysis and adjustment.

[0117] When the number of optimization iterations reaches a preset value or the objective function value converges, the catalyst formulation that satisfies the compatibility constraints and has the optimal objective function value is output as the final optimization result. At this point, to ensure the feasibility and practicality of the final result, small-scale experiments may be needed to verify its performance.

[0118] In the catalyst formulation optimization process, the main catalyst was A, the co-catalyst was B, and the support material was C. Experimental data showed that the optimal ratio of A to B was 3:1, achieving a catalytic conversion rate of 85%. The compatibility knowledge base recorded a compatibility score of 0.9 for A and B, and 0.8 for C. According to the compatibility constraint function, a formulation is considered feasible if the overall compatibility score is greater than 0.75. Finally, using a Bayesian optimization algorithm, the optimal formulation was determined to be a 3:1:1 ratio of A, B, and C, achieving a catalytic conversion rate of 88% and satisfying all compatibility constraints.

[0119] Figure 3 Here is a complete flowchart of the catalyst formulation optimization for embodiments of the present invention:

[0120] This method first establishes a comprehensive catalyst component compatibility knowledge base, systematically including compatibility rules and component ratio constraints among the main catalyst, co-catalyst, and support materials, providing fundamental data support for subsequent optimization. Based on this, an objective function for formulation optimization is constructed, which comprehensively considers the catalytic conversion parameter and the component compatibility parameters calculated based on the compatibility rules. Subsequently, a compatibility constraint function is constructed based on the established component compatibility knowledge base. This function is used to evaluate the compatibility level of different component ratio schemes and sets a scoring threshold to determine the feasibility of the formulation. Next, the compatibility constraint function is integrated into a Bayesian-optimized Gaussian process regression model, forming a complete optimization model considering component compatibility. During the optimization process, the system performs iterative search using a Bayesian optimization algorithm. Each iteration first generates candidate sampling points, then uses the compatibility constraint function to screen out feasible sampling points that meet the requirements, and selects the point with the largest expected improvement for evaluation. Finally, when the optimization iteration reaches a preset number or the objective function value converges, the catalyst formulation that simultaneously satisfies the compatibility constraints and is optimal in the objective function is output as the final optimization result. The entire method achieves catalyst formulation optimization that satisfies both component compatibility requirements and excellent performance through the organic combination of compatibility constraints and Bayesian optimization.

[0121] In one optional implementation, performing Bayesian optimization search based on the optimization model includes:

[0122] In each iteration, a set of candidate sampling points is first generated, and then the compatibility constraint function is used to filter the candidate sampling points to obtain a set of feasible sampling points that meet the compatibility requirements. The sampling point with the largest expected improvement is selected from the set of feasible sampling points for evaluation.

[0123] At the start of each iteration, the system generates a set of candidate sampling points using a pre-defined sampling strategy. These candidate points can be selected based on previous evaluation results and current model predictions, with the aim of covering the entire search space and ensuring diversity and representativeness.

[0124] A compatibility constraint function is applied to the generated set of candidate sampling points. This function evaluates whether each candidate point meets predetermined compatibility conditions. These compatibility conditions can be based on physical limitations, resource constraints, or other business logic requirements. Through this process, a set of feasible sampling points that meet the compatibility requirements is selected.

[0125] After obtaining the set of feasible sampling points, the system calculates the expected improvement for each feasible point. The expected improvement refers to the degree to which the model performance may improve after evaluation at that point. This process, based on historical data and the current model's predictions, aims to identify the most promising sampling points.

[0126] Select the sampling points with the greatest expected improvement from the set of feasible sampling points for practical evaluation. The evaluation process may include physical experiments, simulation tests, or other data collection methods. The selection is based on ensuring that the most valuable information is obtained in each iteration to drive the optimization process.

[0127] After the evaluation is completed, the newly acquired data is fed back into the optimization model to update the model parameters. This step is the core of Bayesian optimization, ensuring that the model can learn new information and continuously improve its predictive ability.

[0128] Latin hypercube sampling (LHS) or other random sampling methods are used to ensure that candidate points are uniformly distributed in multidimensional space.

[0129] When designing compatibility constraint functions, heuristic algorithms or rule engines can be used to define compatibility conditions based on specific application scenarios. For example, in material selection, it may be necessary to consider the physical properties and cost constraints of the materials.

[0130] By constructing a Gaussian process regression model, historical data is used to predict the performance of each feasible point and calculate its expected improvement. This process does not involve mathematical formulas, but rather describes the input-output relationship of the model.

[0131] Suppose that in a certain optimization problem, the characteristics of candidate sampling points include temperature, pressure, and time. Through the above steps, the generated candidate points might be (100°C, 1.0MPa, 30min), (120°C, 1.5MPa, 45min), (110°C, 1.2MPa, 60min), etc. After screening using a compatibility constraint function, the final feasible points are (100°C, 1.0MPa, 30min) and (110°C, 1.2MPa, 60min). Calculating the expected improvement for these two points, assuming the expected improvement for (100°C, 1.0MPa, 30min) is 5% and the expected improvement for (110°C, 1.2MPa, 60min) is 8%, then the latter is selected for evaluation.

[0132] In one optional implementation, the proportions of the components include:

[0133] Optimal mass percentage of alkaline oxide main catalyst; optimal mass percentage of metal oxide co-catalyst; optimal mass percentage of porous support material;

[0134] The reaction process parameters include:

[0135] Optimal reaction temperature range; optimal reaction pressure range; optimal space velocity range.

[0136] Choosing a suitable basic oxide as the main catalyst, a mass percentage range of 15% to 25% is generally recommended. Within this range, the catalyst's activity and selectivity achieve an optimal balance. In practice, sodium hydroxide or potassium hydroxide can be used, and reactivity tests should be conducted to ensure that the catalyst exhibits optimal performance under the desired reaction conditions.

[0137] The mass percentage of the co-catalyst is generally between 5% and 15%, with the specific choice depending on the reaction system. Commonly used metal oxides include aluminum oxide and zinc oxide. Determining the optimal ratio of co-catalyst through experiments can effectively improve the reaction rate and product yield.

[0138] The proportion of porous support material should be between 30% and 50% to ensure catalyst dispersion and reactant contact efficiency. Activated carbon or silica gel is recommended as the support, with surface modification to improve its adsorption capacity and reactivity.

[0139] The optimal temperature range for the reaction should be set between 200°C and 300°C. Experiments were conducted using a gradual heating method, recording the reaction rate and product distribution to determine the optimal reaction temperature. Within this temperature range, the activation energy of the reactants is effectively overcome, and the reaction rate is significantly increased.

[0140] The optimal range of reaction pressure is 0.5 MPa to 1.5 MPa. By adjusting the pressure inside the reactor and observing the reactant conversion and product selectivity, the most suitable pressure conditions were ultimately determined. Under high pressure conditions, the density of the gaseous reactants increases, and the reaction rate also increases accordingly.

[0141] The space velocity should be controlled between 100 h⁻¹ and 300 h⁻¹. By adjusting the gas flow rate inside the reactor, the contact time between the reactants and the catalyst should be ensured to avoid a decrease in conversion rate caused by premature outflow of reactants from the reactor. The reaction effect at different space velocities should be monitored during the experiment to determine the optimal space velocity.

[0142] Sodium hydroxide was used as the main catalyst, with a mass percentage of 20%, a reaction temperature of 250°C, a pressure of 1.0 MPa, and a space velocity of 200 h⁻¹. The final product conversion rate was 85%.

[0143] Using aluminum oxide as a co-catalyst at a mass percentage of 10%, the reaction temperature was 230°C, the pressure was 0.8 MPa, the space velocity was 150 h⁻¹, and the product selectivity reached 90%.

[0144] A second aspect of the present invention provides a deep learning-based system for optimizing the formulation of dimethyl oxalate catalysts, comprising:

[0145] The first unit is used to acquire historical experimental data of multiple groups of dimethyl oxalate catalysts, including catalyst group distribution data, reaction process parameter data, and catalytic conversion rate data.

[0146] The second unit is used to construct a deep neural network model, which includes an input layer, multiple hidden layers and an output layer. The input layer receives catalyst group distribution data and reaction process parameter data, and the output layer outputs predicted catalytic conversion rate data.

[0147] The third unit is used to train the deep neural network model using a multi-task learning strategy, and simultaneously optimize the catalytic conversion rate prediction task and the catalyst stability prediction task during the training process, wherein the catalyst stability is characterized by the decay rate of catalytic activity after multiple consecutive reactions.

[0148] The fourth unit is used to construct a catalyst formulation optimization objective function based on the trained deep neural network model, with catalytic conversion rate and catalyst stability as optimization objectives.

[0149] The fifth unit is used to optimize and solve the objective function using an improved Bayesian optimization algorithm. In the Bayesian optimization process, component compatibility constraints are introduced, and sampling points are screened through a pre-established component compatibility knowledge base to ensure that the optimized catalyst formulation components have good compatibility.

[0150] The sixth unit is used to determine the optimal formulation of the dimethyl oxalate catalyst based on the Bayesian optimization results, including the proportions of each component and the corresponding reaction process parameters.

[0151] A third aspect of the present invention,

[0152] An electronic device is provided, comprising:

[0153] processor;

[0154] Memory used to store processor-executable instructions;

[0155] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0156] Fourth aspect of the embodiments of the present invention,

[0157] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0158] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing the formulation of dimethyl oxalate catalyst based on deep learning, characterized in that, include: Historical experimental data of multiple groups of dimethyl oxalate catalysts were obtained, including catalyst group composition data, reaction process parameter data, and catalytic conversion rate data. A deep neural network model is constructed, which includes an input layer, multiple hidden layers, and an output layer. The input layer receives catalyst group distribution data and reaction process parameter data, and the output layer outputs predicted catalytic conversion rate data. The deep neural network model is trained using a multi-task learning strategy, and the catalytic conversion rate prediction task and the catalyst stability prediction task are optimized simultaneously during the training process. The catalyst stability is characterized by the decay rate of catalytic activity after multiple consecutive reactions. Based on the trained deep neural network model, a catalyst formulation optimization objective function is constructed, with catalytic conversion rate and catalyst stability as optimization objectives. An improved Bayesian optimization algorithm is used to optimize and solve the objective function. In the Bayesian optimization process, component compatibility constraints are introduced. The sampling points are screened by a pre-established component compatibility knowledge base to ensure that the optimized catalyst formulation components have good compatibility. Based on the Bayesian optimization results, the optimal formulation of the dimethyl oxalate catalyst was determined, including the proportions of each component and the corresponding reaction process parameters.

2. The method according to claim 1, characterized in that, The deep neural network model includes an input layer, multiple hidden layers, and an output layer. The input layer receives catalyst composition data and reaction process parameter data, and the output layer outputs predicted catalytic conversion rate data, including: Multiple input neurons are set in the input layer. The multiple input neurons are used to receive the proportion data of each component of the catalyst. The proportion data of the catalyst components includes the proportion data of the main catalyst, the proportion data of the co-catalyst, the proportion data of the support material, and the reaction process parameter data in the reaction temperature data, reaction pressure data, and reaction time data. A residual connection structure is constructed in the multiple hidden layers, and a cross-layer information transmission path is established between adjacent hidden layers through the residual connection structure; An output neuron is provided in the output layer, and the output neuron is used to output the predicted catalytic conversion rate data; An attention mechanism is used to dynamically weight the input data of the multiple input neurons, and the feature weights are adaptively adjusted according to the degree of influence of different input features on the catalytic conversion rate. The attention weights are obtained by alternating forward propagation and back propagation.

3. The method according to claim 1, characterized in that, The deep neural network model is trained using a multi-task learning strategy, and the catalytic conversion rate prediction task and the catalyst stability prediction task are optimized simultaneously during the training process, including: A dual-task neural network structure is constructed, which includes a feature extraction layer, a catalytic conversion rate prediction branch, and a catalyst stability prediction branch, wherein the feature extraction layer is used to extract feature information from the training input data; Establish a task weight allocation mechanism to dynamically adjust the weight ratio of the two tasks based on the training loss values ​​of the catalytic conversion rate prediction task and the catalyst stability prediction task. A task association module is constructed, which is used to establish a feature mapping relationship between the catalytic conversion prediction branch and the catalyst stability prediction branch, so that the two prediction tasks can share and utilize their respective learning features. The dual-task neural network structure is trained using a multi-task joint optimization method, and the network parameters of the feature extraction layer, the catalytic conversion rate prediction branch, and the catalyst stability prediction branch are updated simultaneously. When the prediction errors of both the catalytic conversion rate prediction task and the catalyst stability prediction task are less than the preset error threshold, the trained dual-task neural network model is obtained.

4. The method according to claim 1, characterized in that, An improved Bayesian optimization algorithm is used to optimize and solve the objective function. The component compatibility constraint is introduced during the Bayesian optimization process, and the sampling points are screened using a pre-established component compatibility knowledge base, including: A catalyst component compatibility knowledge base is established, which includes compatibility rule data and component ratio restriction data among the main catalyst, co-catalyst and support material; An objective function for catalyst formulation optimization is constructed, wherein the objective function includes a catalytic conversion parameter and a component compatibility parameter calculated based on the compatibility rule data; A compatibility constraint function is constructed based on the component compatibility knowledge base. The compatibility constraint function is used to calculate the compatibility score under different component ratios, and the feasibility of the formulation is determined based on the compatibility score. The formulation with a compatibility score greater than a preset threshold is determined to be a feasible formulation. The compatibility constraint function is integrated into the Bayesian-optimized Gaussian process regression model to establish an optimization model that considers component compatibility. Bayesian optimization search is performed based on the optimization model; the sampling points and their objective function values ​​obtained in each iteration are fed back to the Gaussian process regression model to update the prior information of the optimization model; When the number of optimization iterations reaches a preset value or the objective function value converges, the catalyst formulation that satisfies the compatibility constraint and has the optimal objective function value is output as the final optimization result.

5. The method according to claim 4, characterized in that, Bayesian optimization search based on the aforementioned optimization model includes: In each iteration, a set of candidate sampling points is first generated, and then the compatibility constraint function is used to filter the candidate sampling points to obtain a set of feasible sampling points that meet the compatibility requirements. The sampling point with the largest expected improvement is selected from the set of feasible sampling points for evaluation.

6. The method according to claim 1, characterized in that, The proportions of each component include: Optimal mass percentage of alkaline oxide main catalyst; optimal mass percentage of metal oxide co-catalyst; optimal mass percentage of porous support material; The reaction process parameters include: Optimal reaction temperature range; optimal reaction pressure range; optimal space velocity range.

7. A deep learning-based dimethyl oxalate catalyst formulation optimization system, used to implement the method as described in any one of claims 1-6, characterized in that, include: The first unit is used to acquire historical experimental data of multiple groups of dimethyl oxalate catalysts, including catalyst group distribution data, reaction process parameter data, and catalytic conversion rate data. The second unit is used to construct a deep neural network model, which includes an input layer, multiple hidden layers and an output layer. The input layer receives catalyst group distribution data and reaction process parameter data, and the output layer outputs predicted catalytic conversion rate data. The third unit is used to train the deep neural network model using a multi-task learning strategy, and simultaneously optimize the catalytic conversion rate prediction task and the catalyst stability prediction task during the training process, wherein the catalyst stability is characterized by the decay rate of catalytic activity after multiple consecutive reactions. The fourth unit is used to construct a catalyst formulation optimization objective function based on the trained deep neural network model, with catalytic conversion rate and catalyst stability as optimization objectives. The fifth unit is used to optimize and solve the objective function using an improved Bayesian optimization algorithm. In the Bayesian optimization process, component compatibility constraints are introduced, and sampling points are screened through a pre-established component compatibility knowledge base to ensure that the optimized catalyst formulation components have good compatibility. The sixth unit is used to determine the optimal formulation of the dimethyl oxalate catalyst based on the Bayesian optimization results, including the proportions of each component and the corresponding reaction process parameters.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images