A Smart Simulation and Computational Method for Complex Chemical Reactions

Abstract

This invention discloses an intelligent simulation and computation method for complex chemical reactions. Combining multimodal data fusion, deep learning models, and physically constrained networks (PINNs), along with reinforcement learning optimization strategies, it achieves efficient and accurate prediction of reaction processes and intelligent control of reactor parameters. The method first constructs a reaction model using graph neural networks (GNNs) and physically constrained neural networks (PINNs), embedding core equations from chemical kinetics, mass transfer, and fluid dynamics to ensure the prediction results conform to physical laws. Then, based on the proximal strategy optimization (PPO) algorithm, it achieves real-time optimization of reaction conditions, dynamically adjusting operating parameters such as temperature, pressure, and flow rate to maximize reaction conversion, selectivity, and energy efficiency. This invention can be widely applied to chemical reaction systems such as catalytic reactions, multiphase reactors, microchannel reactors, and drug synthesis, significantly improving computational efficiency and accuracy, reducing the computation time of traditional methods, and providing intelligent online optimization and control solutions for industrial chemical reaction processes.

Description

Technical Field

[0001] This invention relates to the fields of chemical reaction engineering and artificial intelligence technology, and in particular to an intelligent simulation and calculation method and its application for complex chemical reactions. Background Technology

[0002] In the field of chemical production, process simulation and precise control of complex chemical reactions such as phase catalysis, polymerization, and redox reactions are the core links to improve reaction conversion rate, product selectivity and production energy efficiency.

[0003] Currently, the mainstream reaction simulation methods in the industry have obvious shortcomings: First, the mechanism modeling method based on physicochemical equations, although it can follow the basic physicochemical laws, has a long modeling cycle and is difficult to adapt to complex multiphase reaction systems; Second, the pure data-driven machine learning modeling method, although it can fit the correlation of multi-source monitoring data, is prone to predictive results that violate physical laws, such as mass non-conservation and abnormal reaction rates, and the model has poor generalization ability.

[0004] Furthermore, the unified integration and feature fusion of multi-source heterogeneous data, such as quantum chemical data, experimental data, CFD simulation data, and process monitoring data, is very difficult, resulting in the input data of deep learning models being of a single dimension and incomplete information, which further reduces the accuracy of reaction simulation and control. Summary of the Invention

[0005] In view of the above problems, the purpose of this invention is to provide an intelligent simulation and calculation method for complex chemical reactions, specifically achieving the following objectives:

[0006] 1. Solve the problem of unifying the processing of multi-source heterogeneous data, and achieve precise integration of quantum chemical data, experimental data, CFD simulation data, and process monitoring data;

[0007] 2. To address the issue of pure data-driven models violating physical laws, the reasonableness of simulation results is ensured through Physically Constrained Neural Networks (PINN).

[0008] 3. To solve the problem of poor dynamic adaptability of traditional control strategies, real-time optimization control of the reaction process is achieved based on reinforcement learning;

[0009] 4. Construct a complete technical chain of "data input - model building - constraint verification - optimization control - feedback closed loop" to improve the accuracy and timeliness of complex chemical reaction simulation and control.

[0010] To achieve the above objectives, this invention provides an intelligent simulation and calculation method for complex chemical reactions, comprising the following steps:

[0011] Step 1, Multimodal Data Acquisition and Data Fusion:

[0012] The multimodal data acquisition module collects four types of multi-source heterogeneous data: quantum chemical data (including but not limited to reactant molecular configuration, chemical bond energy level, and reaction energy barrier), experimental data (including but not limited to reaction yield, component concentration, and catalyst activity in pilot / intermediate-scale stages), CFD simulation data (including but not limited to flow field distribution, mass transfer rate, and temperature gradient within the reactor), and process monitoring data (including but not limited to real-time temperature, pressure, feed flow rate, and catalyst concentration in the industrial reactor).

[0013] Data fusion processing;

[0014] The collected data is preprocessed as follows: outliers are removed using the 3σ criterion to clean the data, z-score standardization is used to unify the units, and a cross-source feature mapping algorithm based on the attention mechanism is used to fuse features, outputting a unified feature dataset with a dimension of 256.

[0015] Data priority calibration rules are set as follows: When there is a deviation between experimental data and CFD simulation data, the simulation data is corrected based on the experimental data. The CFD simulation data is dynamically calibrated based on the confidence level of the experimental data. The confidence level of the experimental data is determined according to the parallel experimental error: when the parallel experimental error is <5%, the weight of the experimental data is 0.7; when the parallel experimental error is in the range of 5%-8%, the weight is 0.5; when the parallel experimental error is >8%, the experimental data needs to be re-collected. The re-collected experiment must meet the following requirements: the number of parallel experimental groups ≥3 (preferably 5 groups) and the parallel error <5%. Otherwise, the experimental equipment needs to be replaced or the experimental scheme optimized before re-collecting. When the parallel error of the experimental data is <3% and the accuracy level of the experimental equipment is 0.1 or higher, the confidence level of the experimental data can be increased to 0.8, corresponding to a calibration weight of 0.2 for the CFD simulation data.

[0016] Step 2, Multi-scale Deep Learning Modeling:

[0017] The deep learning model is a multi-network model composed of GNN, CNN, LSTM, and Transformer. Based on different data types of a unified feature dataset, the output features of each network module are matched with corresponding neural network modules and then mapped to a unified dimension through a 128-dimensional fully connected layer. The division of labor among the modules in the multi-network model is as follows: the GNN module processes the topological structure data of reactant molecules, characterizing intermolecular interactions and reaction paths; the CNN module processes the spatial grid data of the CFD flow field distribution and extracts the flow field gradient features; the LSTM module processes the time series data of the reaction process (such as the curve of component concentration changing with time) and fits the dynamic change law of the reaction; the Transformer module realizes cross-scale feature association and integrates multi-dimensional information of molecular interactions, flow field distribution, and temporal changes.

[0018] Multi-network feature fusion is achieved through a Softmax attention mechanism, with the following weight allocation: molecular topology features 0.3, flow field features 0.25, temporal features 0.25, and cross-scale correlation features 0.2. The output provides preliminary reaction process prediction results (such as temporal predictions of reaction conversion rate and product selectivity) with a time series length of 60 minutes and a dimension of 64. This format can be directly input into the subsequent PINN module for constraint verification.

[0019] Step 3, Physically Constrained Neural Network (PINN) constraint verification:

[0020] Physicochemical constraint equations are embedded in the loss function layer of the deep learning model, specifically including: chemical kinetic equations (reaction rate constraints), mass transfer equations (component diffusion rate constraints), fluid dynamics equations (momentum conservation constraints), and mass / energy conservation equations.

[0021] Constructing a composite loss function And set constraint weights according to the type of reaction system. Constraint weights The value of is determined as follows: 0.3-0.4 for homogeneous reaction systems and 0.5-0.6 for heterogeneous reaction systems. This range has been verified by orthogonal experiments to balance data fitting accuracy and physical constraint rigidity.

[0022] The preliminary prediction results are verified for physical laws. The termination condition for PINN constraint verification is that the mass conservation deviation of the prediction results is ≤0.5%, the momentum conservation deviation is ≤1%, and the reaction rate deviation is ≤2%, while the loss function fluctuation of 100 consecutive iterations is <0.001. If the results violate the conservation law, the model parameters are corrected through backpropagation until the output prediction results conform to physical laws.

[0023] Step 4, Reinforcement learning to optimize control:

[0024] A reinforcement learning control framework is constructed based on the PPO (Proximal Policy Optimization) algorithm: the agent's state space is defined as the real-time monitoring data of the reactor and the PINN prediction results, and the agent's action space is defined as adjustable operating parameters (temperature adjustment range). Flow rate adjustment range Catalyst concentration adjustment range (), where temperature and flow rate are in a continuous action space and are adapted to the optimization logic of the PPO algorithm;

[0025] Design a multi-objective reward function: First, perform maximum-min standardization on each indicator to eliminate dimensional differences, and then perform weighted calculation: Reward = 0.4 × conversion rate + 0.3 × selectivity + 0.3 × energy efficiency - 0.2 × by-product ratio. The weight coefficients are determined by the analytic hierarchy process, where energy efficiency is quantified by the energy consumption per unit product.

[0026] The agent iteratively optimizes the control strategy based on the PPO algorithm. The reinforcement learning strategy iteration is divided into regular iteration with a cycle of 5 minutes and emergency iteration. When the reactor operating condition fluctuates by more than ±2% or the predicted conversion rate deviates from the actual value by more than 3%, the emergency iteration is triggered, and the emergency iteration has a higher priority than the regular iteration.

[0027] Step 5, Industrial Deployment:

[0028] The model trained in the cloud is subjected to knowledge distillation and lightweight processing: a pre-trained large model is used as the teacher model and a lightweight model is used as the student model. The distillation temperature coefficient is set to 1.2, which reduces the number of model parameters to 30% of the original model, making it suitable for industrial computing power.

[0029] The resulting lightweight model was deployed to an industrial edge computing control terminal with 8 cores and 16G specifications to achieve real-time inference of the model with an inference latency of ≤1s.

[0030] Set tiered safety thresholds, divided into warning and emergency levels: the warning level threshold is the temperature deviation from the baseline value. Pressure deviation from the benchmark value This triggers cooling / pressure relief regulation; the emergency threshold is when the temperature exceeds the maximum limit. Pressure > Maximum Limit This triggers an operation that first cuts off the feed and then stops the machine; when multiple thresholds exceed the limit, instructions are executed according to the priority of pressure > temperature > flow rate to ensure the stability and safety of industrial applications.

[0031] Step 6, Closed-loop feedback:

[0032] The reactor operating condition data after adjustment is collected and transmitted back to the multimodal data acquisition module, forming a closed-loop link of "data input-optimized control-operating condition feedback" to realize dynamic adaptive control of the reaction process.

[0033] This invention also provides an intelligent simulation and calculation method for complex chemical reactions, applicable to microchannel reactors or fluidized bed reactors. For microchannel reactors, adding a flow channel microscale effect feature extraction branch to the CNN module in step 2 can reduce the flow field feature extraction error by 20%. For fluidized bed reactors, supplementing the particle aggregation-dispersion topological structure characterization to the GNN module in step 2 can improve the prediction accuracy of aggregation phenomena by 15%. Furthermore, the PPO algorithm is adapted to the adjustment of reactor operating parameters in continuous action space.

[0034] The beneficial effects of this invention are as follows:

[0035] At the data level, multimodal data fusion and priority calibration solve the integration problem of multi-source heterogeneous data, providing comprehensive and reliable input data for the model. Compared with traditional CFD simulation methods, the computational efficiency of this invention is improved by more than 10 times, and the information dimension of the simulation results is improved by more than 30%. Compared with the pure data-driven LSTM model, the average prediction error of this invention is reduced from 15% to less than 5%. The above comparison is based on the ethylene oxide synthesis reaction system. Traditional CFD simulation uses Fluent software with 1 million grids and 5,000 iteration steps, and the calculation time for a single working condition is 24 hours. The method of this invention has a single working condition calculation time of ≤2 hours, which is more than 10 times more efficient. The pure LSTM model has a 6-layer network and 1.2 million parameters. Without physical constraints, the prediction error reaches 15% under a 5% fluctuation in the working condition. The method of this invention, due to the embedding of PINN constraints, keeps the prediction error within 5% and does not violate physical laws.

[0036] At the model level, by embedding physical constraints through PINN, the "physical paradox" of pure data-driven models is avoided, reducing the average error of simulation results to less than 5%, which is far better than the 15% error of traditional machine learning models.

[0037] At the control level: Based on reinforcement learning, multi-objective optimization control realizes dynamic adaptive adjustment of operating parameters. In typical complex reaction systems, it can improve conversion rate by 8%-12%, selectivity by 5%-8%, and production energy efficiency by more than 10%.

[0038] At the application level: The complete closed-loop technology chain can be adapted to a variety of industrial reactors, combining versatility and practicality, and providing a full-process solution for the intelligent control of complex chemical reactions. Attached Figure Description

[0039] Figure 1 This is a flowchart of the intelligent simulation and calculation method for complex chemical reactions according to the present invention;

[0040] Figure 2 This is a schematic diagram of multimodal data fusion in an embodiment of the present invention. Detailed Implementation

[0041] To make the technical means, creative features, objectives and effects of this invention easier to understand, the following embodiments, in conjunction with the accompanying drawings, will specifically illustrate the technical solution of this invention.

[0042] This embodiment illustrates the application of the intelligent simulation and calculation method of the present invention for complex chemical reactions in the ethylene oxide synthesis reaction in a microchannel reactor, including the following steps:

[0043] Step 1, Multimodal Data Acquisition and Data Fusion:

[0044] Collect the following four types of multi-source heterogeneous data:

[0045] Quantum chemical data: Molecular energy levels of ethylene and oxygen, and energy barriers for epoxidation reactions were calculated using Gaussian software.

[0046] Experimental data: The reaction conversion rate and selectivity under different temperatures (180-220℃) and pressures (1.8-2.5MPa) in the pilot plant. The parallel error of this experiment is 3.2%, and the experimental equipment is a 0.05-level high-precision device. The results are in line with the standard of a confidence weight of 0.8.

[0047] CFD simulation data: The flow field distribution and mass transfer rate within the microchannel were simulated using Fluent software.

[0048] Process monitoring data: real-time temperature, pressure, ethylene feed flow rate, and silver catalyst concentration of the industrial microchannel reactor.

[0049] Configure the training environment for reinforcement learning and deep learning models: Use Python, build the model based on the PyTorch framework, set the number of training iterations to 10,000, and the learning rate to 0.001.

[0050] Data fusion processing;

[0051] In this embodiment, the experimental equipment is a 0.05-level high-precision detection device, and the parallel experiment error is 3.2% < 5%, so the overall confidence level of the experimental data is determined to be 0.8. Outliers with errors > 8% in the experimental data are removed using the 3σ criterion. After z-score standardization of all data, an attention mechanism is used to fuse molecular topological features, flow field grid features, and temporal monitoring features into a unified feature vector of dimension 256. Simultaneously, the confidence level of the experimental data is set to 0.8 to complete the calibration of the CFD simulation data. After calibration, the deviation between the CFD data and the experimental data decreases from 12% to 4.5%.

[0052] Step 2, Multi-scale Deep Learning Modeling:

[0053] A unified feature vector was input into the multi-network model, and the output features of each module were unified in dimensionality through a 128-dimensional fully connected layer: the GNN module extracted the adsorption features of the ethylene-silver catalyst, the CNN module extracted the flow field gradient features in the microchannel, the LSTM module fitted the temporal changes of the reaction rate, and the Transformer module integrated the multi-dimensional features. After weighting through the Softmax attention mechanism, a preliminary conversion rate prediction curve with a time length of 60 minutes was output. The average error between the preliminary prediction results and the experimental values ​​was 8.7%.

[0054] Step 3, PINN constraint verification:

[0055] PINN constraint verification: This embodiment involves a heterogeneous catalytic reaction, and constraint weights are set. =0.5), embedding the kinetic equation for ethylene oxide synthesis into the loss function. After 10,000 iterations of training, the average error between the model's predicted conversion rate and the actual value decreased to 3.2%, and there were no predictions that violated mass conservation, thus satisfying the PINN constraint verification termination condition. In this embodiment, after PINN constraint verification, the model's predicted momentum conservation bias was 0.8% (≤1%), the reaction rate bias was 1.5% (≤2%), and the loss function fluctuation over 100 consecutive iterations was 0.0008 (<0.001), satisfying all the PINN constraint verification termination conditions.

[0056] Step 4, Reinforcement learning to optimize control:

[0057] The agent's state space is defined as "real-time temperature / pressure / flow rate + predicted conversion rate". In this embodiment, the process baseline value for the ethylene oxide synthesis reaction is... =200℃ =1.0m / s =0.65g / L, so the action space is "temperature (185-215℃), flow rate (0.8-1.2m / s), catalyst concentration (0.5-0.8g / L)". First, the maximum-minimum standardization of each indicator is performed, and then the PPO algorithm is iterated 5000 times until the reward function converges to a stable value. After the control strategy is issued, the actual conversion rate of the reactor increases from 78% to 87% (an increase of 11.5%, which is in line with the effect range of 8%-12%), the selectivity increases from 85% to 92% (an increase of 8%, which is in line with the effect range of 5%-8%), and the energy consumption per unit product decreases by 11% (in line with the effect range of more than 10%). When the operating condition fluctuates by more than 3%, the system triggers an emergency iteration, which corrects the operating parameters in time and avoids a decline in conversion rate.

[0058] Step 5, Industrial Deployment:

[0059] The trained model is subjected to knowledge distillation and lightweighting, and then deployed to the edge computing control terminal of the microchannel reactor with an inference latency of 0.8s.

[0060] Step 6, Closed-loop feedback and verification:

[0061] The adjusted operating condition data is continuously collected and fed back to the data module, and the model iterates at a regular 5-minute cycle. During 72 hours of continuous operation, the fluctuation range of the operating condition is controlled within ±3%, which is far better than the ±8% fluctuation of traditional PID control. At the same time, when the temperature exceeds the threshold, an emergency stop can be triggered within 0.5 seconds according to priority, verifying the stability and reliability of this method. The 'data input-model modeling-constraint verification-optimized control-feedback closed loop' link constructed by this invention has been verified through 72 hours of continuous operation: multimodal fusion at the data layer improves the integrity of input information by 30%, PINN constraints at the model layer improve prediction accuracy by 60%, PPO algorithm at the control layer improves the operating condition response speed by 50%, and lightweight processing at the industrial deployment layer controls the inference latency to within 1 second. The entire link realizes intelligent management and control of complex chemical reactions.

[0062] In this embodiment, since the reactor operating parameters are continuously adjustable, the PPO algorithm is used for strategy optimization throughout the process, and the DQN algorithm, which is suitable for discrete action spaces, is not used, thus ensuring the adaptability of the algorithm to the application scenario. Furthermore, the index expression is unified, and "reaction efficiency" is clearly defined as a comprehensive index of conversion rate and selectivity, and its improvement effect is completely matched with the numerical range of the technical solution.

[0063] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. An intelligent simulation and calculation method for complex chemical reactions, characterized in that, Includes the following steps: Step 1: Collect four types of multi-source heterogeneous data: quantum chemical data, experimental data, CFD simulation data, and process monitoring data, and perform data fusion processing on the collected data; Step 2: Construct a deep learning model and use the unified feature data obtained after fusion processing for model training; Step 3: Use the physical information neural network to constrain and validate the loss function of the deep learning model; Step 4: Construct a reinforcement learning framework based on the PPO algorithm and perform optimization control; Step 5: Complete the industrial deployment of the model.

2. The intelligent simulation and calculation method for complex chemical reactions as described in claim 1, characterized in that: The specific data fusion processing in step 1 The process includes the following: The collected data is preprocessed: data cleaning, data standardization, and feature fusion, to output a unified feature dataset adapted to deep learning models; Set data priority calibration rules: CFD simulation data is dynamically calibrated based on the confidence level of experimental data. The confidence level of experimental data is determined according to the parallel experimental error: when the parallel experimental error is <5%, the weight of the experimental data is 0.7; when the parallel experimental error is in the range of 5%-8%, the weight is 0.5; when the parallel experimental error is >8%, the experimental data needs to be re-acquired. The re-acquired experiment must meet the requirements of ≥3 parallel experimental groups and parallel error <5%. Otherwise, the experimental equipment needs to be replaced or the experimental plan needs to be optimized before re-acquiring data.

3. The intelligent simulation and calculation method for complex chemical reactions as described in claim 2, characterized in that: The data cleaning uses the 3σ criterion to remove outliers, the data standardization uses the z-score method to achieve dimensional uniformity, and the feature fusion is completed using a cross-source feature mapping algorithm based on an attention mechanism.

4. The intelligent simulation and calculation method for complex chemical reactions as described in claim 1, characterized in that: The deep learning model in step 2 is a multi-network model composed of GNN, CNN, LSTM and Transformer. The output feature dimension of each network is a 128-dimensional fully connected layer. After unifying the dimensions, the preliminary prediction results are output after fusion through the Softmax attention mechanism. The weight allocation of the Softmax attention mechanism is 0.3 for molecular topology features, 0.25 for flow field features, 0.25 for temporal features, and 0.2 for cross-scale correlation features.

5. The intelligent simulation and calculation method for complex chemical reactions as described in claim 4, characterized in that: The division of labor among the networks in the multi-network model is as follows: GNN processes reactant molecular topology data, CNN processes CFD flow field distribution spatial grid data, LSTM processes reaction process time series data, and Transformer realizes cross-scale feature association. The preliminary prediction results are time-series predicted values ​​such as reaction conversion rate and product selectivity with a time series length of 60 min and a dimension of 64.

6. The intelligent simulation and calculation method for complex chemical reactions as described in claim 1, characterized in that: Step 3 details The process includes the following: Physicochemical constraint equations, including chemical kinetic equations, mass transfer equations, fluid dynamics equations, and conservation equations, are embedded in the model loss function layer. Constructing a composite loss function And set constraint weights according to the type of reaction system. For homogeneous reaction systems, the concentration is 0.3-0.4; for heterogeneous reaction systems, the concentration is 0.5-0.

6. The preliminary prediction results are verified for physical laws. The termination condition for PINN constraint verification is that the mass conservation deviation of the prediction results is ≤0.5%, the momentum conservation deviation is ≤1%, and the reaction rate deviation is ≤2%, while the loss function fluctuation of 100 consecutive iterations is <0.

001. If the results violate the conservation law, the model parameters are corrected through backpropagation until the output prediction results conform to physical laws.

7. The intelligent simulation and calculation method for complex chemical reactions as described in claim 1, characterized in that: Step 4 specifically includes the following processes: Constructing a reinforcement learning control framework: The state space of the agent is set as the real-time monitoring data of the reactor and the prediction results of PINN, and the action space of the agent is the adjustable operation parameters, including temperature regulation, flow rate regulation and catalyst concentration regulation. Among them, temperature and flow rate are continuous action spaces and are adapted to the optimization logic of the PPO algorithm. Design a multi-objective reward function: First, perform maximum-min standardization on each indicator to eliminate dimensional differences, and then perform weighted calculation: Reward = 0.4 × conversion rate + 0.3 × selectivity + 0.3 × energy efficiency - 0.2 × by-product ratio. The weight coefficients are determined by the analytic hierarchy process, where energy efficiency is quantified by the energy consumption per unit product. The agent iteratively optimizes the control strategy based on the PPO algorithm. The reinforcement learning strategy iteration is divided into regular iteration with a cycle of 5 minutes and emergency iteration. When the reactor operating condition fluctuates by more than ±2% or the predicted conversion rate deviates from the actual value by more than 3%, the emergency iteration is triggered, and the emergency iteration has a higher priority than the regular iteration.

8. The intelligent simulation and calculation method for complex chemical reactions as described in claim 1, characterized in that: Step 5 specifically includes the following processes: The model is lightened through knowledge distillation: a pre-trained large model is used as the teacher model and a lightweight model is used as the student model. The distillation temperature coefficient is set to 1.2, which reduces the number of model parameters to 30% of the original model. The resulting lightweight model was then deployed to the edge computing module of the industrial reactor control terminal, ensuring that the module inference latency was ≤1s. Set tiered safety thresholds, divided into warning and emergency levels: the warning level threshold is the temperature deviation from the baseline value. Pressure deviation from the benchmark value This triggers cooling / pressure relief regulation; The emergency threshold is a temperature greater than the maximum limit. Pressure > Maximum Limit This triggers an operation that first cuts off the feed and then stops the machine; when multiple thresholds exceed the limit, the command is executed according to the priority of pressure > temperature > flow rate.

9. The intelligent simulation and calculation method for complex chemical reactions as described in claim 1, characterized in that, It also includes a closed-loop feedback step.

10. An intelligent simulation and calculation method for complex chemical reactions as described in any one of claims 1-9, applied to microchannel reactors or fluidized bed reactors.

Images