AI-driven chemical experiment simulation and result prediction system

Abstract

The invention relates to the technical field of chemical experiments, and discloses an AI-driven chemical experiment simulation and result prediction system. According to the system, multi-modal data is collected through an experimental data acquisition module, spectral features are extracted through a spectral feature analysis module, a coupling kinetic model is constructed through a reaction kinetic modeling module, a product distribution probability is predicted through a dynamic result prediction module, reaction conditions are optimized, and reaction parameters are adjusted in real time through an experimental parameter feedback module. In addition, the system also comprises a reaction safety evaluation model which is used for identifying risks and providing a safety protection scheme. The system realizes comprehensive simulation and accurate result prediction of chemical experiments, can effectively optimize reaction conditions, improves reaction efficiency and safety, and has wide application prospects in the fields of chemical experiment research, chemical production and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of chemical experiments, specifically an AI-driven chemical experiment simulation and result prediction system. Background Art

[0002] In the field of chemical experiments, traditional experimental methods face many challenges and are difficult to meet the growing needs of modern chemical research and industrial production. From the perspective of experimental data acquisition, in the past, data was mainly collected by a single type of sensor, with limited data dimensions. For example, when studying solution reactions, only the basic parameter of reactant concentration could be measured, and it was difficult to comprehensively obtain key information such as the temperature distribution of the solution and the dynamic gas diffusion. This makes it impossible for researchers to deeply understand the microscopic changes and complex interaction mechanisms during the reaction process. Although the application of multi-spectral imaging technology and chemical sensor arrays is gradually emerging, there are still deficiencies in the systematicness and coordination of data acquisition, and it is unable to efficiently integrate multi-modal data, resulting in the omission of a large amount of valuable information.

[0003] There are also bottlenecks in the experimental result prediction and analysis. Relying on experience and simple mathematical models for result prediction has poor accuracy and reliability. Take complex organic synthesis reactions as an example. Due to the complex reaction mechanism and numerous influencing factors, traditional prediction methods often cannot take into account the significant impact of small changes in reaction conditions on product distribution, resulting in a large deviation between the prediction result and the actual experimental result. In the face of new chemical reaction systems, traditional methods are even more difficult to quickly give effective predictions and optimization schemes, seriously hindering the R & D process of new chemical products.

[0004] The control accuracy and real-time adjustment ability of experimental conditions also restrict the development of chemical experiments. Traditional experimental devices usually adopt fixed control strategies and cannot dynamically adjust reaction conditions according to real-time changes during the reaction. For example, in terms of temperature control, it is difficult to achieve precise regulation of the temperature gradient of the reaction system. Once a temperature fluctuation occurs during the reaction, it cannot respond in a timely manner, resulting in an unstable reaction environment and affecting reaction efficiency and product quality. In terms of stirring rate control, there is also a lack of means to intelligently adjust according to the reaction process, and it is unable to fully promote the mixing and mass transfer of reactants, restricting the progress of the reaction.

[0005] In addition, with the popularization of the concepts of green chemistry and sustainable development, higher requirements are put forward for the safety, environmental protection and resource utilization efficiency of chemical experiments. However, traditional experimental methods are relatively weak in reaction safety assessment, lacking a real-time monitoring and effective warning mechanism for reaction runaway risks, and are prone to causing safety accidents. In terms of resource utilization, due to the inability to accurately optimize reaction conditions, reactants are often wasted, increasing production costs and having a negative impact on the environment at the same time. Summary of the Invention

[0006] The object of the present invention is to provide an AI-driven chemical experiment simulation and result prediction system to solve the problems raised in the above-mentioned background technology.

[0007] To achieve the above object, the present invention provides the following technical solution: an AI-driven chemical experiment simulation and result prediction system, the system includes: An experimental data acquisition module: used to collect multi-modal data of the experimental process through a chemical sensor array and a multi-spectral imaging device, the multi-modal data includes reactant concentration time-series data, solution temperature distribution images, gas diffusion dynamic video streams, and reaction vessel pressure waveforms; A spectral feature analysis module: uses wavelet transform combined with a convolutional neural network to extract spectral features from the multi-modal data, and generates a multi-dimensional analysis result including substance absorption peak position features, reaction rate correlation features, and spatio-temporal features of the phase change process; A reaction kinetics modeling module: based on the multi-dimensional analysis result, uses partial differential equations to construct a coupled kinetics model of the reactant concentration field and the temperature field, and solves the gradient change of the reaction path through the finite element method; A dynamic result prediction module: based on the coupled kinetics model, uses a temporal convolutional network to predict the product distribution probability under different experimental parameters, and generates a reaction condition optimization scheme; An experimental parameter feedback module: according to the reaction condition optimization scheme, adjusts the temperature gradient and stirring rate parameters of the reaction system in real time through an adaptive control algorithm.

[0008] Preferably, the experimental data acquisition module includes: Measure the time-series data of the ion migration rate at the reaction interface using an electrochemical sensor array, and capture the surface temperature distribution map of the reaction vessel through an infrared thermal imager; Use a high-speed imaging device to record the turbulent feature video stream of the gas evolution process, and synchronously collect the oscillation waveform data of the pressure sensor in the reaction kettle.

[0009] Preferably, the spectral feature analysis module further includes: Use a frequency domain decomposition algorithm to separate the spectral responses of different bands for the multi-spectral imaging data, and combine the attention mechanism to weight and fuse the feature channels; Extract the vortex structure features of the gas diffusion dynamic video stream through a three-dimensional convolution kernel, and construct a correlation matrix between the turbulence intensity and the reaction rate.

[0010] Preferably, the reaction kinetics modeling module further includes: Establish an unsteady mass transfer equation to describe the change of the reactant concentration gradient, and input the temperature field data as a boundary condition; The multi-grid iteration algorithm is adopted to accelerate the solution process of partial differential equations, and a deep neural network surrogate model is constructed based on residual connections.

[0011] Preferably, the dynamic result prediction module further includes: Construct a bidirectional gated recurrent unit network to model the long-term dependencies of the reaction process, extract multi-scale representations of temporal features, generate a perturbation set of reaction parameters through Monte Carlo sampling, and calculate the probability density function of the product distribution.

[0012] Preferably, the dynamic result prediction module further includes: Adopt a generative adversarial network to construct a virtual experimental environment and generate simulation data of the reaction process under extreme conditions; Align the feature distribution differences between the virtual data and the actual observed data through a contrastive learning algorithm.

[0013] Preferably, the experimental parameter feedback module further includes: designing a fuzzy PID controller to adjust the power output of the constant temperature device and dynamically adjusting the control parameters based on the gradient features of the temperature distribution map.

[0014] Preferably, the experimental parameter feedback module further includes: Adopt a reinforcement learning framework to train a parameter adjustment policy network, and define the state space as the multi-modal feature vector of the current reaction system; Generate an optimal control instruction sequence by constraining the balance relationship between energy consumption and reaction efficiency through a reward function.

[0015] Preferably, the system further includes: Construct a reaction safety assessment model, based on the gradient field data of the coupled kinetic model, use a support vector machine classifier to identify the reaction runaway risk threshold; when risk features are detected, trigger an emergency parameter adjustment protocol and generate a safety protection suggestion plan.

[0016] Preferably, the reaction safety assessment model further includes: associating historical accident case data through a knowledge graph to generate a safety protection strategy inference path for the current reaction system.

[0017] Compared with the prior art, the beneficial effects of the present invention are: The AI-driven chemical experiment simulation and result prediction system provided by the present invention has many significant beneficial effects. In terms of experimental data collection, through the chemical sensor array and multi-spectral imaging device, multi-modal data such as the time-series data of reactant concentration, the solution temperature distribution image, the gas diffusion dynamic video stream, and the reaction vessel pressure waveform can be obtained. This comprehensive data collection method greatly enriches the researchers' cognitive dimension of the reaction process. Taking the catalytic reaction experiment as an example, in the past, only the general change of reactant concentration over time could be known. Now, with the help of this system, not only can the details of the concentration change be accurately traced, but also the real-time distribution of temperature in the reaction system can be observed synchronously, the dynamic process of gas diffusion and the pressure fluctuation can be understood. This helps to deeply explore the microscopic mechanism in the reaction process and provides a richer and more accurate data basis for revealing the essential laws of chemical reactions.

[0018] The spectral feature analysis module uses wavelet transform combined with convolutional neural network to extract spectral features and generate multi-dimensional analysis results, which makes the analysis of experimental data more in-depth and accurate. The acquisition of information such as the position characteristics of substance absorption peaks, the reaction rate correlation characteristics, and the spatio-temporal characteristics of the phase change process can help researchers better understand the changes of substances and the reaction process in chemical reactions. When analyzing the synthesis reaction of a certain new material, through these characteristics, the key stages of the reaction can be accurately judged, the trend of the reaction can be predicted in advance, and a strong basis can be provided for timely adjusting the experimental plan.

[0019] The reaction kinetics modeling module constructs a coupled kinetics model based on the multi-dimensional analysis results and uses the finite element method to solve the gradient change of the reaction path. This modeling method can more accurately describe the interaction relationship between the reactant concentration field and the temperature field in the reaction process. In the large-scale reaction simulation in chemical engineering production, this model can help engineers optimize the design of reaction equipment, reasonably arrange the feeding position and temperature distribution of reactants, improve the reaction efficiency, and reduce production costs. At the same time, based on the calculation results of this model, the energy change in the reaction process can be predicted more accurately, providing a scientific basis for the thermal management of the reaction.

[0020] The dynamic result prediction module uses the temporal convolutional network to predict the product distribution probability under different experimental parameters and generates a reaction condition optimization plan. This provides efficient decision-making support for chemical experiments and chemical engineering production. In the drug synthesis experiment, researchers can quickly find the optimal parameter combinations such as reaction temperature, pressure, and reactant ratio according to the prediction results, reduce the number of experimental trials and errors, and greatly shorten the R & D cycle. In industrial production, by optimizing the reaction conditions, the quality and yield of products can be improved, and the market competitiveness of enterprises can be enhanced.

[0021] The experimental parameter feedback module adjusts the temperature gradient and stirring rate parameters of the reaction system in real time according to the reaction condition optimization scheme to ensure that the reaction is always carried out under optimal conditions. This real-time feedback and adjustment mechanism effectively improves the stability and controllability of the reaction. In fine chemical production, even minor changes in reaction conditions can affect product quality. This module can adjust parameters in a timely manner according to the reaction situation, ensuring the consistency of product quality and reducing the scrap rate.

[0022] In addition, the reaction safety assessment model constructed by the system identifies the reaction runaway risk threshold based on the gradient field data of the coupled kinetic model. When risk characteristics are detected, it triggers an emergency parameter adjustment protocol and generates a safety protection suggestion plan. By associating historical accident case data through a knowledge graph, it generates an inference path of safety protection strategies for the current reaction system, which greatly improves the safety of chemical experiments and production processes, effectively avoids safety accidents caused by reaction runaway, and ensures the safety of personnel's lives and the property of enterprises. Brief Description of the Drawings

[0023] Figure 1 is the working principle diagram of the AI-driven chemical experiment simulation and result prediction system described in the present invention; Figure 2 is the refined flowchart of the experimental data acquisition module; Figure 3 is the further expanded working principle diagram of the dynamic result prediction module; Figure 4 is the further expanded working principle diagram of the experimental parameter feedback module. Detailed Embodiments

[0024] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0025] Please refer to Figures 1 - 4 , the present invention provides an AI-driven chemical experiment simulation and result prediction system, and its specific implementation process is as follows: Collect multimodal data during the experiment through the experimental data acquisition module. This module uses a chemical sensor array and a multispectral imaging device to collect multimodal data such as time-series data of reactant concentrations, images of solution temperature distributions, dynamic video streams of gas diffusion, and pressure waveforms of reaction vessels. For example, during a specific chemical reaction experiment, the chemical sensor array monitors the change in reactant concentration over time in real time to generate time-series data of reactant concentrations; the multispectral imaging device takes pictures of the reaction solution to obtain images of solution temperature distributions to visually present the distribution state of the solution temperature during the reaction process; at the same time, relevant equipment is used to record the dynamic video stream of gas diffusion and the pressure waveform of the reaction vessel, thereby comprehensively collecting various types of data during the experiment.

[0026] The spectral feature analysis module processes the collected multimodal data. It uses a method combining wavelet transform and convolutional neural network to extract spectral features, and then generates multi-dimensional analysis results including the position features of substance absorption peaks, reaction rate correlation features, and spatio-temporal features of phase change processes. This process helps to deeply analyze the changes in substances during the reaction process and provides an important basis for subsequent reaction kinetics modeling.

[0027] The reaction kinetics modeling module works based on the multi-dimensional analysis results. It uses partial differential equations to construct a coupled kinetics model of the reactant concentration field and the temperature field to describe the change relationship between the reactant concentration and temperature during the reaction process. Then, the finite element method is used to solve the gradient change of the reaction path to more accurately understand the change trend during the reaction process.

[0028] Then, the dynamic result prediction module makes predictions based on the coupled kinetics model. It uses a temporal convolutional network to predict the product distribution probability under different experimental parameters to provide a prediction basis for the experimental results. At the same time, this module also generates a reaction condition optimization scheme to guide the experimenter to adjust the experimental conditions to obtain better experimental results.

[0029] The experimental parameter feedback module makes real-time adjustments to the reaction system according to the reaction condition optimization scheme. Through an adaptive control algorithm, it makes real-time adjustments to the temperature gradient and stirring rate parameters of the reaction system to ensure that the reaction proceeds under optimized conditions and improve the reaction efficiency and product quality.

[0030] The technical solution of the present invention will be further described in detail below in conjunction with specific embodiments.

[0031] Example 1: In this embodiment, the implementation manner of the experimental data acquisition module is further refined. The experimental data acquisition module specifically includes measuring the time-series data of the ion mobility at the reaction interface using an electrochemical sensor array, and capturing the surface temperature distribution map of the reaction vessel by an infrared thermal imager. Meanwhile, a high-speed camera device is used to record the turbulent feature video stream of the gas evolution process, and the oscillating waveform data of the pressure sensor in the reaction kettle is synchronously acquired.

[0032] For measuring the time-series data of the ion mobility at the reaction interface using an electrochemical sensor array, the electrochemical sensor array is composed of multiple different types of electrochemical sensors, which can accurately measure the migration of ions at the reaction interface. Its working principle is based on the ion conduction characteristics of electrochemistry. When a reaction occurs, ions migrate at the reaction interface, and the electrochemical sensor can detect the change in the electrical signal generated by the ion migration in real time, convert it into ion mobility data, and record it in chronological order to form the time-series data of the ion mobility at the reaction interface. By analyzing these data, the dynamic changes of ions during the reaction can be understood, providing important information for studying the reaction mechanism.

[0033] When using an infrared thermal imager to capture the surface temperature distribution map of the reaction vessel, the infrared thermal imager measures the temperature using the infrared radiation characteristics of the object. During the reaction process of the reaction vessel, due to the thermal effect of the chemical reaction, the surface temperature will change. The infrared thermal imager receives the infrared radiation emitted from the surface of the reaction vessel, converts it into an electrical signal, and then generates the surface temperature distribution map of the reaction vessel through a series of signal processing and algorithm conversions. This distribution map visually shows the temperature differences of each part of the reaction vessel surface in different colors, helping the experimenter clearly observe the temperature distribution during the reaction process, and then analyze the influence of temperature on the reaction.

[0034] When using a high-speed camera device to record the turbulent feature video stream of the gas evolution process, the high-speed camera device shoots the area where the gas evolves during the reaction at a high frame rate. During the gas evolution process, complex turbulent flows will be formed, and the high-speed camera device can capture the instantaneous states of these turbulences and record them as a video stream. By analyzing the video stream, various characteristics of the turbulence, such as the velocity of the turbulence and the vortex structure, can be extracted. These characteristics are of great significance for studying gas diffusion and reaction kinetics. In terms of synchronously acquiring the oscillating waveform data of the pressure sensor in the reaction kettle, the pressure sensor is installed at a suitable position in the reaction kettle. When the pressure in the reaction kettle changes due to the generation or consumption of gas during the reaction process, the pressure sensor will sense the pressure fluctuation and convert it into an electrical signal. These electrical signals are processed to form oscillating waveform data, which reflects the change of pressure with time during the reaction process. By analyzing the pressure oscillation waveform data, information such as the progress of the reaction and the generation or consumption rate of gas can be understood.

[0035] Example 2: This example elaborates in detail the specific implementation details of the spectral feature analysis module. Based on the extraction of spectral features from multi-modal data using wavelet transform combined with convolutional neural network, the spectral feature analysis module also includes separating the spectral responses of different bands from multi-spectral imaging data using a frequency-domain decomposition algorithm, and weighted fusion of feature channels by combining with an attention mechanism. And the vortex structure features of the gas diffusion dynamic video stream are extracted by a three-dimensional convolution kernel, and the correlation matrix between the turbulence intensity and the reaction rate is constructed.

[0036] When separating the spectral responses of different bands from multi-spectral imaging data using a frequency-domain decomposition algorithm, the multi-spectral imaging data contains spectral information of multiple bands. The frequency-domain decomposition algorithm is based on the Fourier transform principle and converts the multi-spectral imaging data from the time domain to the frequency domain. In the frequency domain, the spectral responses of different bands have different frequency characteristics. By setting appropriate frequency filters, the spectral responses of different bands can be separated. Suppose the multi-spectral imaging data is , after Fourier transform to obtain its frequency-domain representation , and then according to the frequency ranges corresponding to different bands , , etc., using band-pass filters , , etc. to filter , that is , , etc., and then through inverse Fourier transform , , etc. to obtain the separated spectral response data of different bands , , etc. In this way, the separation of the spectral responses of different bands is achieved, providing a basis for subsequent feature extraction and analysis.

[0037] When combining with an attention mechanism to weighted fusion of feature channels, the attention mechanism is a method that can make the model pay more attention to important features. In spectral feature analysis, the information contained in the spectral response data of different bands has different degrees of importance for reaction analysis. First, feature extraction is performed on the separated spectral response data of each band to obtain feature vectors , ,..., . Then, the weights of each feature vector are calculated through an attention network , ,..., , and the attention network usually consists of multiple layers of neural networks. Its input is the concatenation of all feature vectors , and the output is the weight vector Finally, the weights are weighted and fused with the feature vectors to obtain the fused feature vectors. In this way, important features are given higher weights, making the fused features better reflect the key information of the reaction.

[0038] When extracting the vortex structure features of the gas diffusion dynamic video stream through a three-dimensional convolution kernel, the gas diffusion dynamic video stream is a three-dimensional data containing information in two dimensions: time and space. The three-dimensional convolution kernel can perform convolution operations on the video stream data simultaneously in time and space. Let the three-dimensional convolution kernel be with a size of where represents the size of the convolution kernel in the time dimension, and and represent the sizes of the convolution kernels in the space dimension. For the video stream data , after the three-dimensional convolution operation, the feature map is obtained. By designing an appropriate three-dimensional convolution kernel, the vortex structure features in the gas diffusion dynamic video stream can be effectively extracted, and these features reflect the turbulence characteristics in the gas diffusion process.

[0039] When constructing the correlation matrix between the turbulence intensity and the reaction rate, the turbulence intensity needs to be calculated first. Based on the extracted vortex structure features, the turbulence intensity can be calculated through a certain algorithm . The reaction rate can be obtained by measuring the change rate of the reactant concentration over time, that is, , where is the reactant concentration and is the time. Then, the turbulence intensities , ,..., at different times are correlated with the corresponding reaction rates , ,..., to construct the correlation matrix . The matrix element represents the correlation relationship between the turbulence intensity at the th time and the reaction rate at the th time, for example . By analyzing this correlation matrix, the internal relationship between the turbulence intensity and the reaction rate can be deeply understood, providing strong support for the study of reaction kinetics.

[0040] Example 3: This embodiment will elaborate on the reaction kinetics modeling module in depth. Based on the multi-dimensional analysis results, the reaction kinetics modeling module constructs a coupled kinetics model of the reactant concentration field and the temperature field using partial differential equations, and on the basis of solving the gradient change of the reaction path by the finite element method, it also includes establishing an unsteady mass transfer equation to describe the change of the reactant concentration gradient and inputting the temperature field data as boundary conditions. Meanwhile, a multi-grid iteration algorithm is used to accelerate the solution process of the partial differential equations, and a deep neural network surrogate model is constructed based on residual connections.

[0041] When establishing the unsteady mass transfer equation to describe the change of the reactant concentration gradient, the unsteady mass transfer equation is established based on the principle of mass conservation. Let the reactant concentration be , the time be , the spatial coordinates be , , , the diffusion coefficient be . The general form of the unsteady mass transfer equation is . This equation describes the diffusion of reactants in space over time, where the left side of the equation represents the rate of change of the reactant concentration with time, and the right side represents the concentration change due to diffusion. In an actual reaction system, the diffusion of reactants is affected by various factors, and these effects can be quantitatively analyzed through this equation. When inputting the temperature field data as boundary conditions, since temperature affects parameters such as the diffusion coefficient of reactants, the temperature field data needs to be incorporated into the unsteady mass transfer equation. Assuming the temperature field data is , the temperature and the diffusion coefficient can be related through a certain functional relationship, for example , where is the reference diffusion coefficient, is the diffusion activation energy, is the gas constant, is the absolute temperature. In this way, the temperature field data participates in the unsteady mass transfer equation by affecting the diffusion coefficient, more accurately describing the concentration change during the reaction process.

[0042] When using the multigrid iteration algorithm to accelerate the solution process of partial differential equations, the multigrid iteration algorithm is an efficient numerical solution method. For partial differential equations, the solution process usually requires discretizing the entire computational domain to form a large linear system of equations. The multigrid iteration algorithm approximates the exact solution of the equation quickly by performing iterative solutions on grids with different resolutions. First, perform a preliminary iteration on the finest grid to obtain an approximate solution. Since there are high-frequency errors on the fine grid, the errors on the fine grid are transferred to the coarse grid through a restriction operator, and the errors are corrected on the coarse grid. Then, the corrected result on the coarse grid is transferred back to the fine grid through an interpolation operator to continue the iteration. In this way, by repeatedly iterating between different grids, the solution process can be effectively accelerated and the computation time can be reduced.

[0043] When constructing a deep neural network surrogate model based on residual connections, the residual connection is a commonly used structure in deep neural networks. Let the input of the deep neural network be , and the output be . The input of a certain layer in the network is , and the output is . The output of the traditional neural network layer is , while in the residual connection, , where is the nonlinear transformation function of this layer. Through this residual connection method, the problem of gradient vanishing in the training process of deep neural networks can be effectively solved, enabling the network to more easily learn complex mapping relationships. When constructing a deep neural network surrogate model, the inputs of the reaction kinetics model (such as reactant concentration, temperature, etc. data) are used as the inputs of the deep neural network, and are trained through a multi-layer residual connection network, so that the output of the network can approximate the output of the reaction kinetics model. In this way, the deep neural network surrogate model can replace the complex reaction kinetics model for rapid calculation and prediction, improving the computational efficiency.

[0044] Example 4: This example details the specific implementation method of the dynamic result prediction module. Based on predicting the product distribution probability under different experimental parameters using a temporal convolutional network and generating a reaction condition optimization scheme based on the coupled kinetics model, the dynamic result prediction module further includes constructing a bidirectional gated recurrent unit network to model the long-term dependencies of the reaction process and extracting multi-scale representations of temporal features. A perturbation set of reaction parameters is generated through Monte Carlo sampling, and the probability density function of the product distribution is calculated.

[0045] When constructing a bidirectional gated recurrent unit network to model the long-term dependencies of the reaction process, the bidirectional gated recurrent unit (Bi-GRU) network consists of a forward GRU and a backward GRU. GRU is a gated recurrent unit that contains a reset gate and the update gate . For the input at time and the hidden state at the previous time , the reset gate , the update gate , where is the sigmoid function, , , , , is the weight matrix. The candidate hidden state , the final hidden state , where represents element-wise multiplication. The forward GRU processes the input sequentially from the start time of the sequence, and the backward GRU processes the input backward from the end time of the sequence. The outputs of the forward GRU and the backward GRU are concatenated to obtain the output of the bidirectional gated recurrent unit network, which can fully capture the long-term dependencies between different times during the reaction process and better understand the dynamic changes of the reaction.

[0046] When extracting the multi-scale representation of the temporal features, based on the bidirectional gated recurrent unit network, different convolutional operations are used to extract the multi-scale temporal features. Convolution kernels of different sizes can be adopted, such as those with sizes of , , . For the output of the bidirectional gated recurrent unit network, convolutional operations are performed on it using convolution kernels of different sizes. Let the convolution kernels be , , , then the features , , after convolution, where represents the convolution operation. Convolution kernels of different sizes can capture features at different time scales. Small convolution kernels can capture the detailed changes at short time scales, and large convolution kernels can capture the trend changes at long time scales, thus obtaining a multi-scale representation of the temporal features and analyzing the reaction process more comprehensively.

[0047] When generating the perturbation set of the reaction parameters through Monte Carlo sampling and calculating the probability density function of the product distribution, Monte Carlo sampling is a method based on random sampling. Assume that there are reaction parameters, which are respectively , ,..., . First, determine the value range , ,..., . Then, a large number of random samplings are carried out within these value ranges to generate sample points of reaction parameters, and each sample point is , . These sample points constitute the perturbation set of reaction parameters. For each sample point, the corresponding product distribution is calculated using the coupled kinetic model. Let the product distribution be . By performing statistical analysis on a large number of , methods such as kernel density estimation are used to calculate the probability density function of the product distribution . In this way, the influence of the uncertainty of reaction parameters on the product distribution can be considered, and the possibility of the product distribution can be predicted more accurately.

[0048] Example 5: This example focuses on the further extended functions of the dynamic result prediction module. In addition to the functions mentioned before, the dynamic result prediction module also includes constructing a virtual experimental environment using a generative adversarial network to generate simulation data of reaction processes under extreme conditions. And the feature distribution differences between the virtual data and the actual observed data are aligned through a contrastive learning algorithm.

[0049] When constructing a virtual experimental environment using a generative adversarial network to generate simulation data of reaction processes under extreme conditions, the generative adversarial network (GAN) consists of a generator and a discriminator . The role of the generator is to generate simulation data according to the input random noise , that is, . The generated simulation data here is the reaction process data under extreme conditions, such as reaction data under conditions like high temperature and high pressure that are difficult to easily achieve in actual experiments. The role of the discriminator is to judge whether the input data is real actual observed data or simulation data generated by the generator, and its output is a probability value indicating the possibility that the input data is real data. During the training process, the generator and the discriminator perform adversarial training. The generator tries to generate more realistic simulation data so that the discriminator is difficult to distinguish between true and false; the discriminator tries to improve its discrimination ability to accurately distinguish between real data and simulation data. Through continuous adversarial training, the generator can generate reaction process simulation data of higher and higher quality under extreme conditions. Assume that the random noise follows a normal distribution , where represents a zero-mean vector, and represents the unit covariance matrix. The generator is a model composed of a multi-layer neural network. It takes the random noise as input and outputs the simulated reaction process data through a series of linear transformations and non-linear activation functions Discriminator It is also a neural network that receives input data (which can be real data or simulated data generated by the generator). After internal calculations, it outputs a probability value between During training, the loss function of the generator aims to minimize the probability that the discriminator misclassifies the generated data as fake data, which can be expressed as ; the loss function of the discriminator aims to maximize the ability to correctly distinguish between real data and generated data, which can be expressed as , where represents the real actual observed data. By alternately optimizing the generator and the discriminator, the generator can finally generate simulated data close to the reaction process under real extreme conditions.

[0050] When aligning the feature distribution differences between virtual data and actual observed data through the contrastive learning algorithm, the core idea of the contrastive learning algorithm is to let the model learn the similarity and difference of the data. First, feature extraction is performed on the virtual data and the actual observed data. Assume that the extracted virtual data features are , and the actual observed data features are . Then, a loss function for contrastive learning is defined. A commonly used contrastive learning loss function is the InfoNCE loss function. The calculation formula of the InfoNCE loss function is , where is the number of data pairs, represents the similarity between the features and , which can usually be calculated using cosine similarity, that is , is a temperature hyperparameter used to adjust the distribution of similarity. The purpose of this loss function is to make the feature similarity of the same pair of virtual data and actual observed data stand out among the feature similarities of all data pairs, thereby narrowing the feature distribution differences between virtual data and actual observed data. During the training process, by continuously adjusting the parameters of the model and minimizing the InfoNCE loss function, the model can better align the feature distributions of virtual data and actual observed data. In this way, when using virtual data for analysis and prediction, it can more effectively combine the actual situation and improve the accuracy and reliability of the prediction.

[0051] Example 6: This embodiment is described in detail around the experimental parameter feedback module. Based on the reaction condition optimization scheme, the experimental parameter feedback module adjusts the temperature gradient and stirring rate parameters of the reaction system in real time through an adaptive control algorithm. It also includes designing a fuzzy PID controller to adjust the power output of the constant temperature device and dynamically adjusting the control parameters based on the gradient characteristics of the temperature distribution map. At the same time, a reinforcement learning framework is used to train the parameter adjustment policy network. The state space is defined as the multi-modal feature vector of the current reaction system, and the balance relationship between energy consumption and reaction efficiency is constrained through a reward function to generate an optimal control instruction sequence.

[0052] When designing a fuzzy PID controller to adjust the power output of the constant temperature device, the fuzzy PID controller combines the advantages of fuzzy logic and PID control. The control law of the PID controller is , where is the output of the controller, that is, the power output of the constant temperature device; is the proportional coefficient, is the integral coefficient, is the differential coefficient; is the error, that is, the difference between the set temperature and the actual measured temperature. In the fuzzy PID controller, according to the gradient characteristics of the temperature distribution map, fuzzy logic is used to , and are dynamically adjusted. First, the gradient characteristics of the temperature distribution map are quantified and fuzzified, and converted into fuzzy language variables, such as "large", "medium", "small", etc. Then, according to the pre-set fuzzy rule base, the adjustment amounts of , and are obtained through fuzzy inference. The fuzzy rule base is established based on experience and experimental data. For example, when the temperature gradient is large and the temperature is higher than the set value, is appropriately increased to accelerate the temperature adjustment speed, and at the same time is decreased to avoid integral saturation, and is increased to suppress the excessive fluctuation of the temperature. Finally, the adjusted , and are substituted into the PID control formula to calculate the power output of the constant temperature device, realizing precise control of the temperature of the reaction system.

[0053] When using a reinforcement learning framework to train the parameter adjustment policy network, reinforcement learning is a method of learning the optimal policy by an agent interacting with the environment and according to the reward signal. In this system, the state space is defined as the multi-modal feature vector of the current reaction system, including information such as reactant concentration, temperature, pressure, etc., denoted as . The action space is the adjustment values of the temperature gradient and stirring rate of the reaction system, denoted as The reward function is used to constrain the balance relationship between energy consumption and reaction efficiency. Let the reward function be . For example, when the reaction efficiency increases and the energy consumption is low, a higher reward is given; when the reaction efficiency decreases or the energy consumption is too high, a lower reward is given. The parameter adjustment policy network outputs a probability distribution of an action according to the current state , that is . During the training process, the agent observes the current state at each time step , selects an action according to the policy network and executes it, then observes the new state and obtains the reward . By continuously interacting with the environment, experience data is accumulated, and the policy network is updated using this data , so that the policy network can learn the optimal policy that maximizes the cumulative reward. Commonly used reinforcement learning algorithms such as the Proximal Policy Optimization algorithm (PPO) update the policy network by optimizing an objective function, and the objective function is , where are the parameters of the policy network are the parameters of the previous update is the number of samples is the advantage function, which represents the degree of advantage of taking the action in the state , the function limits the value of within the range of , is a hyperparameter. Through continuous iterative training, the policy network can generate an optimal control instruction sequence to achieve the optimal adjustment of the temperature gradient and stirring rate of the reaction system, improve the reaction efficiency and reduce the energy consumption.

[0054] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent in such process, method, article or device.

[0055] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-driven chemical experiment simulation and result prediction system, characterized in that, Including: An experimental data acquisition module: used to acquire multi-modal data during the experiment through a chemical sensor array and a multi-spectral imaging device. The multi-modal data includes time-series data of reactant concentration, solution temperature distribution images, gas diffusion dynamic video streams, and reaction vessel pressure waveforms. A spectral feature analysis module: using wavelet transform combined with a convolutional neural network to extract spectral features from the multi-modal data, generating a multi-dimensional analysis result including the position features of substance absorption peaks, reaction rate correlation features, and spatio-temporal features of the phase change process. A reaction kinetics modeling module: based on the multi-dimensional analysis result, using partial differential equations to construct a coupled kinetics model of the reactant concentration field and temperature field, and solving the gradient change of the reaction path through the finite element method. A dynamic result prediction module: based on the coupled kinetics model, using a temporal convolutional network to predict the product distribution probability under different experimental parameters, and generating a reaction condition optimization scheme. An experimental parameter feedback module: according to the reaction condition optimization scheme, adjusting the temperature gradient and stirring rate parameters of the reaction system in real time through an adaptive control algorithm.

2. The chemical experiment simulation and result prediction system according to claim 1, characterized in that The experimental data acquisition module includes: Measuring the time-series data of ion mobility at the reaction interface using an electrochemical sensor array, and capturing the surface temperature distribution map of the reaction vessel through an infrared thermal imager. Using a high-speed imaging device to record the turbulent feature video stream of the gas evolution process, and synchronously collecting the oscillation waveform data of the pressure sensor in the reaction kettle.

3. The chemical experiment simulation and result prediction system according to claim 2, wherein The spectral feature analysis module also includes: Using a frequency-domain decomposition algorithm to separate the spectral responses of different bands for the multi-spectral imaging data, and weighted fusion of feature channels in combination with an attention mechanism. Extracting the vortex structure features of the gas diffusion dynamic video stream through a three-dimensional convolution kernel, and constructing a correlation matrix between turbulent intensity and reaction rate.

4. The chemical experiment simulation and result prediction system according to claim 1, characterized in that, The reaction kinetics modeling module also includes: Establishing an unsteady mass transfer equation to describe the change of reactant concentration gradient, and inputting the temperature field data as boundary conditions. Using a multi-grid iteration algorithm to accelerate the solution process of partial differential equations, and constructing a deep neural network proxy model based on residual connections.

5. The chemical experiment simulation and result prediction system according to claim 1, wherein The dynamic result prediction module also includes: Constructing a bidirectional gated recurrent unit network to model the long-term dependence relationship of the reaction process, extracting multi-scale representations of temporal features, generating a perturbation set of reaction parameters through Monte Carlo sampling, and calculating the probability density function of the product distribution.

6. The chemical experiment simulation and result prediction system according to claim 5, wherein, The dynamic result prediction module also includes: Using a generative adversarial network to construct a virtual experimental environment and generating simulation data of the reaction process under extreme conditions. Aligning the feature distribution differences between virtual data and actual observed data through a contrastive learning algorithm.

7. The chemical experiment simulation and result prediction system according to claim 1, characterized in that, The experimental parameter feedback module also includes: designing a fuzzy PID controller to adjust the power output of the constant temperature device, and dynamically adjusting the control parameters based on the gradient features of the temperature distribution map.

8. The chemical experiment simulation and result prediction system according to claim 7, characterized in that The experimental parameter feedback module also includes: Training a parameter adjustment strategy network using a reinforcement learning framework, and defining the state space as the multi-modal feature vector of the current reaction system. Constraining the balance relationship between energy consumption and reaction efficiency through a reward function, and generating an optimal control instruction sequence.

9. The chemical experiment simulation and result prediction system according to claim 1, characterized in that The system also includes: Build a reaction safety assessment model. Based on the gradient field data of the coupled kinetic model, use a support vector machine classifier to identify the reaction runaway risk threshold; when risk characteristics are detected, trigger an emergency parameter adjustment protocol and generate a safety protection recommendation plan.

10. The chemical experiment simulation and result prediction system according to claim 9, characterized in that The reaction safety assessment model further includes: associating historical accident case data through a knowledge graph to generate an inference path of safety protection strategies for the current reaction system.

Images