Microchannel gas-liquid two-phase flow pattern prediction system and method based on neural network

Abstract

The invention discloses a micro-channel gas-liquid two-phase flow pattern prediction system and method based on a neural network, and relates to the technical field of computational fluid mechanics and artificial intelligence, and the system comprises a data acquisition module which obtains simulation data of a gas-liquid two-phase flow in a micro-channel under different inlet flow velocity conditions, outputting field value information of the sampling points in each iteration process; the data preprocessing module is responsible for cleaning and smoothing the obtained simulation data; the PINN module is used for learning local physical phenomena in simulation data by utilizing a physical information neural network PINN and ensuring that a prediction result conforms to a known physical rule; the ConvLSTM module is used for capturing the change of a long time span and a large-range space in the simulation data by utilizing a convolutional long short-term memory network ConvLSTM; the information fusion module is used for integrating output results of the PINN module and the ConvLSTM module to form a prediction result; wherein the PINN module and the ConvLSTM module form a deep learning network.

Description

Technical Field

[0001] The present invention relates to the field of computational fluid dynamics and artificial intelligence technology, and in particular to a neural network-based microchannel gas-liquid two-phase flow pattern prediction system and method. Background Art

[0002] In recent years, with the rapid development of microfluidic technology, gas-liquid two-phase flow in microchannels has been widely used in many fields such as chemical engineering, biomedical engineering, and energy conversion. The flow characteristics of gas-liquid two-phase flow in microchannels are very complex, including the generation, merging, and rupture of bubbles and the flow pattern of liquids, which have an important impact on the performance and efficiency of the system. However, due to the variability and nonlinear characteristics of gas-liquid two-phase flow, although traditional numerical simulation methods can provide more accurate results, their computational cost is high and difficult to apply in real time. In addition, although experimental methods can provide intuitive data, their operation is complicated and it is difficult to cover all working conditions.

[0003] In order to overcome the limitations of traditional methods, data-driven methods have gradually become a research hotspot. As a powerful data-driven method, deep learning has achieved remarkable success in image recognition, natural language processing and other fields. In the field of fluid mechanics, deep learning has also shown great potential, especially in dealing with complex nonlinear problems. Physics-Informed Neural Networks (PINN) embed physical equations into neural networks to ensure that the prediction results conform to known physical laws. Convolutional Long Short-Term Memory (ConvLSTM) is particularly suitable for processing spatiotemporal series data and can capture temporal and spatial correlations.

[0004] Although the application of deep learning in fluid mechanics has made some progress, there are still some challenges in the prediction of flow patterns of gas-liquid two-phase flow in microchannels. First, existing methods usually only focus on single-scale problems and cannot effectively deal with multi-scale phenomena. Second, the generalization ability of the model is insufficient and it is difficult to adapt to changes in different inlet flow rates and other operating conditions. In addition, the uncertainty quantification of the prediction results is also an urgent problem to be solved. Therefore, it is particularly important to develop a prediction system and method that can not only utilize physical knowledge but also efficiently process spatiotemporal series data, and can perform multi-scale modeling, adaptive incremental learning, and uncertainty quantification. Summary of the invention

[0005] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is how to develop a prediction system and method that can not only utilize physical knowledge but also efficiently process spatiotemporal series data and perform multi-scale modeling, adaptive incremental learning and uncertainty quantification, so as to not only improve the accuracy and robustness of the prediction but also better meet the needs of practical applications.

[0006] To achieve the above object, the present invention provides a microchannel gas-liquid two-phase flow pattern prediction system based on neural network, comprising: A data acquisition module, wherein the data acquisition module acquires simulation data of gas-liquid two-phase flow in the microchannel under different inlet flow rate conditions, and outputs field value information of the sampling point in each iteration process; A data preprocessing module, the data preprocessing module is connected to the data acquisition module and is responsible for cleaning and smoothing the obtained simulation data; A PINN module, wherein the PINN module is connected to the data preprocessing module and uses a physical information neural network PINN to learn local physical phenomena in the simulation data to ensure that the prediction results conform to known physical laws; A ConvLSTM module, wherein the ConvLSTM module is connected to the PINN module and uses a convolutional long short-term memory network ConvLSTM to capture long-time spans and large-scale spatial changes in the simulation data; An information fusion module, wherein the information fusion module connects the PINN module and the ConvLSTM module, integrates the output results of the PINN module and the ConvLSTM module to form a prediction result; Among them, the PINN module and the ConvLSTM module constitute a deep learning network.

[0007] Furthermore, the system also includes an adaptive incremental learning module, which is connected to the deep learning network, uses an online learning algorithm to update the parameters of the deep learning network, and dynamically adjusts the learning rate and weight update strategy to adapt to changing operating conditions.

[0008] Furthermore, the system also includes an uncertainty quantification module, which is connected to the deep learning network, estimates the uncertainty of the prediction result obtained by the information fusion module through a Bayesian method or a Monte Carlo sampling technique, and provides a credibility estimate of the prediction result.

[0009] Furthermore, the data acquisition module acquires the simulation data by means of Fluent simulation software; the field value information includes coordinates (x, y), static pressure, absolute velocities X Velocity and Y Velocity in xy directions, and gas phase VOF.

[0010] Furthermore, the physical information neural network PINN in the PINN module includes multiple multi-layer perceptrons, which are used to predict stream functions, pressure and level set functions respectively; the physical information neural network PINN is trained using a gradient descent method to minimize a total loss function that includes data fitting errors and penalty terms for violating physical laws.

[0011] Furthermore, the convolutional long short-term memory network ConvLSTM in the ConvLSTM module includes multiple convolutional layers and LSTM units, receives a series of field value sequences at continuous time points as input, each time point corresponds to a two-dimensional flow field snapshot, and updates the weights through the back-propagation algorithm to achieve the prediction of the flow field state at future moments.

[0012] The present invention also provides a method for predicting flow patterns of gas-liquid two-phase flow in a microchannel based on a neural network, comprising the following steps: Step 1, obtain the simulation data of gas-liquid two-phase flow in the microchannel under different inlet flow rate conditions, output the field value information of the sampling point in each iteration process, and store it in a .mat file; Step 2: cleaning and smoothing the obtained simulation data; Step 3, using a physical information neural network PINN to learn local physical phenomena in the simulation data to ensure that the prediction results conform to known physical laws, wherein the physical information neural network PINN is embedded with Navier-Stokes equations and includes multiple multi-layer perceptron (MLP) structures; Step 4: Input the prediction result of step 3 as an initial condition into a convolutional long short-term memory network ConvLSTM, wherein the convolutional long short-term memory network ConvLSTM also receives a sequence of field values ​​at a series of continuous time points, each time point corresponding to a two-dimensional flow field snapshot; the convolutional long short-term memory network ConvLSTM captures the spatiotemporal dependencies within a long time span through convolutional layers and LSTM units, thereby providing a global prediction of flow characteristics, and updates weights through a back-propagation algorithm; Step 5: After completing the training of the physical information neural network PINN and the convolutional long short-term memory network ConvLSTM, the output results of the two models are integrated into the final prediction result; assuming that the output of the physical information neural network PINN is , the output of the convolutional long short-term memory network ConvLSTM is , the prediction result after fusion is , then the fusion formula of the attention mechanism is used: in, is the attention weight.

[0013] Furthermore, the MLP structure in step 3 is specifically expressed as: in, is the MLP structure, For time, is the flow rate, is the inlet flow rate, is the stream function, For pressure, is the level set function.

[0014] Furthermore, in step 3, the training of the physical information neural network PINN is to minimize the total loss function including the data fitting error and the penalty term of violating the physical law by the gradient descent method; the total loss function It is expressed as: in, is the data fitting error, is the penalty term for violating the laws of physics, is the weight coefficient for balancing the two items.

[0015] Furthermore, the update formula of the convolutional long short-term memory network ConvLSTM in step 4 is expressed as: Among them, all input , cell output , hidden state , and the door , , are all three-dimensional tensors, where the second two dimensions are spatial dimensions (rows and columns), are network parameters, is the activation function, is the hyperbolic tangent function, represents the convolution operation, Represents element-wise product.

[0016] The present invention provides a microchannel gas-liquid two-phase flow pattern prediction system and method based on neural network, which has at least the following technical effects: The technical solution provided by the present invention significantly improves the accuracy and robustness of prediction by combining PINN and ConvLSTM, and introducing multi-scale modeling, adaptive incremental learning and uncertainty quantification technology. Specifically, PINN ensures the accuracy of local physical phenomena, while ConvLSTM effectively captures the spatiotemporal dependencies over a long time span. Adaptive incremental learning technology enables the model to gradually adjust parameters when new data arrives, improves its adaptability to new conditions, and enhances the generalization ability and practicality of the model. The multi-scale modeling method can comprehensively capture flow characteristics at different spatial and temporal scales, providing more comprehensive prediction results. The introduction of uncertainty quantification provides confidence estimates for the prediction results, enhancing the reliability and interpretability of the system. Combining these advantages, the technical solution provided by the present invention can not only efficiently handle complex gas-liquid two-phase flow problems, but also adapt to changes under various operating conditions and provide reliable prediction results, and has broad application prospects.

[0017] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A system flow chart of a preferred embodiment of the present invention; Figure 2 for Figure 1 A sample diagram of a neural network combining PINN and ConvLSTM of the illustrated embodiment. DETAILED DESCRIPTION

[0019] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0020] The embodiment of the present invention combines Physics-Informed Neural Networks (PINN) and Convolutional Long Short-Term Memory (ConvLSTM) to improve the accuracy and robustness of the flow pattern prediction of gas-liquid two-phase flow in microchannels through multi-scale modeling, adaptive incremental learning and uncertainty quantification technology. Specifically, the embodiment of the present invention uses PINN to learn detailed physical phenomena in a small section of the channel to ensure that the prediction results conform to known physical laws. At the same time, ConvLSTM is used to capture long time spans and large-scale spatial changes to achieve efficient processing of long time series data. Through multi-scale modeling, flow characteristics can be fully captured at different spatial and temporal scales to improve the accuracy of prediction. In addition, the introduction of adaptive incremental learning technology enables the model to gradually adjust parameters to adapt to different inlet flow rates and changes in other operating conditions, thereby improving the generalization ability of the model. The uncertainty quantification module estimates the uncertainty of the prediction results through techniques such as Bayesian methods or Monte Carlo sampling, provides a credibility estimate of the prediction results, and enhances the reliability and interpretability of the system. Finally, the output results of PINN and ConvLSTM are integrated to form a unified and consistent prediction map, ensuring a smooth transition between different models and improving the overall prediction performance.

[0021] Example 1 like Figure 1 As shown, a microchannel gas-liquid two-phase flow pattern prediction system based on a neural network provided by an embodiment of the present invention includes: The data acquisition module uses Fluent simulation software to obtain simulation data of gas-liquid two-phase flow in the microchannel under different inlet flow rate conditions, and outputs the field value information of the sampling point in each iteration process, where the field value information includes coordinates (x, y), static pressure, absolute velocity X Velocity and Y Velocity in the xy direction, and gas phase VOF; The data preprocessing module is connected to the data acquisition module and is responsible for cleaning and smoothing the obtained simulation data. Specifically, it is responsible for cleaning and preparing the data obtained from the Fluent simulation to ensure its quality and improve the effect of model training. This module first eliminates the noise in the data, corrects inconsistent data, and identifies and deletes outliers. Specifically, it includes steps such as missing value processing, outlier detection and replacement, and data standardization or normalization. In addition, the module will also perform data smoothing, such as using a sliding window-based median filtering method to remove spikes and jump values ​​in the data, thereby improving the consistency and reliability of the data; PINN module, the PINN module is connected to the data preprocessing module, and uses the physical information neural network PINN to learn local physical phenomena in the simulation data to ensure that the prediction results conform to the known physical laws; specifically, this module embeds physical laws such as the Navier-Stokes equation as part of the loss function to ensure that the prediction results conform to the known physical laws. The PINN module consists of multiple multi-layer perceptrons (MLPs), each of which is used to predict the stream function, pressure, and level set function. The module is trained using the gradient descent method or other optimization algorithms to minimize the total loss function that includes data fitting errors and penalty terms for violating physical laws. This not only improves the accuracy of the prediction, but also ensures the physical consistency of the prediction results; The ConvLSTM module is connected to the PINN module and uses the convolutional long short-term memory network ConvLSTM to capture the long-term span and large-scale spatial changes in the simulation data; specifically, this module uses the convolutional long short-term memory network (ConvLSTM) to capture data with long-term spans and large-scale spatial changes, which is particularly suitable for processing spatiotemporal sequence data and can capture temporal and spatial correlations. The ConvLSTM module consists of multiple convolutional layers and LSTM units, and receives a series of field value sequences at continuous time points as input, each time point corresponding to a two-dimensional flow field snapshot. The weights are updated by the back-propagation algorithm, so that the model can accurately predict the flow field state at future moments. The ConvLSTM module can effectively handle the spatiotemporal dependencies within a long span and provide global flow characteristics prediction; among them, the PINN module and the ConvLSTM module form a deep learning network.

[0022] The information fusion module connects the PINN module and the ConvLSTM module, integrates the output results of the PINN module and the ConvLSTM module to form a unified and consistent prediction result, namely the prediction map; specifically, the module uses a specific fusion strategy, such as weighted average or attention-based methods, to balance the influence of the two. The information fusion module ensures a smooth transition between different models and improves the overall prediction performance. The module uses a consistency check and correction algorithm to ensure the consistency and accuracy of the final output, thereby providing a more comprehensive and reliable prediction result; Example 2 On the basis of Example 1, a neural network-based microchannel gas-liquid two-phase flow pattern prediction system provided by an embodiment of the present invention also includes an adaptive incremental learning module, which is connected to a deep learning network, uses an online learning algorithm to update the parameters of the deep learning network, and dynamically adjusts the learning rate and weight update strategy to adapt to changing operating conditions. Specifically, the adaptive incremental learning module enables the system to gradually adjust parameters when new data arrives, thereby improving its adaptability to new conditions. The module uses an online learning algorithm to update the parameters of the existing model online while preventing overfitting of old data. The module introduces technologies such as forgetting factors to dynamically adjust the learning rate and weight update strategy to adapt to changing operating conditions. The adaptive incremental learning module significantly enhances the generalization ability and practicality of the model, enabling it to better cope with changes under various operating conditions; A neural network-based microchannel gas-liquid two-phase flow pattern prediction system provided by an embodiment of the present invention also includes an uncertainty quantification module, which is connected to a deep learning network, estimates the uncertainty of the prediction results obtained by the information fusion module through a Bayesian method or a Monte Carlo sampling technique, and provides a credibility estimate of the prediction results. Specifically, the module quantifies the uncertainty of the prediction results by probabilistically modeling the network weights, or by running the model multiple times and calculating the statistical characteristics of the prediction results. The uncertainty quantification module enhances the reliability and interpretability of the system, helping decision makers better understand the confidence of the prediction results, and thus make more scientific and reliable decisions.

[0023] Example 3 The embodiment of the present invention provides a method for predicting the flow pattern of gas-liquid two-phase flow in a microchannel based on a neural network. First, the simulation data of the gas-liquid two-phase flow in the microchannel is obtained under different inlet flow rate conditions by a data acquisition module, and the data preprocessing module cleans and standardizes it. Then, the PINN model uses the preprocessed data to learn the detailed physical phenomena in the local area, and the output includes the velocity field, the pressure field and the level set function field; these local prediction results are input into the ConvLSTM model as initial conditions to capture the spatiotemporal dependencies within a long span. The information fusion module integrates the output results of PINN and ConvLSTM into the final prediction map through a specific fusion strategy. The adaptive incremental learning module enables the system to gradually adjust parameters as new data arrives, thereby improving its adaptability to new conditions. Finally, the uncertainty quantification module estimates the uncertainty of the prediction results through the Bayesian method or the Monte Carlo sampling technique, and provides a credibility estimate. Through the collaborative work of these modules, the entire system realizes efficient, accurate and robust prediction of the flow pattern of gas-liquid two-phase flow in a microchannel.

[0024] Specifically, the method comprises the following steps: Step 1, obtain the simulation data of gas-liquid two-phase flow in the microchannel under different inlet flow rate conditions, output the field value information of the sampling point in each iteration process, and store it in the .mat file; wherein the field value information includes coordinates (x, y), static pressure, absolute velocity X Velocity and Y Velocity in the xy direction, and gas phase VOF; Step 2: Perform preprocessing such as cleaning and standardization on the obtained simulation data, including smoothing; Step 3: Figure 2 As shown, a physical information neural network PINN is used to learn local physical phenomena in simulation data to ensure that the prediction results conform to known physical laws, wherein the Navier-Stokes equation is embedded in the physical information neural network PINN as part of the loss function and includes multiple multi-layer perceptron (MLP) structures; Step 4: Input the prediction results of step 3 (i.e., velocity field, pressure field, and level set function field) as initial conditions into the Convolutional Long Short-Term Memory Network ConvLSTM. These local prediction results provide a detailed initial state for the ConvLSTM model. The Convolutional Long Short-Term Memory Network ConvLSTM also receives a series of field value sequences at a series of continuous time points, each time point corresponding to a two-dimensional flow field snapshot; the Convolutional Long Short-Term Memory Network ConvLSTM captures the spatiotemporal dependencies over a long time span through convolutional layers and LSTM units, thereby providing global flow characteristics predictions, and updates weights through the back-propagation algorithm, so that the model can accurately predict the flow field state at future moments; In particular, during the training process, the adaptive incremental learning module enables the system to gradually adjust parameters as new data arrives, improving its adaptability to new conditions. When new data arrives, the parameters of the existing model are updated through the online learning algorithm. The module introduces techniques such as the forgetting factor to prevent overfitting of old data. This not only improves the real-time performance of the system, but also enhances its robustness and generalization ability under changing operating conditions; Step 5: After completing the training of the physical information neural network PINN and the convolutional long short-term memory network ConvLSTM, the output results of the two models are integrated into the final prediction result; assuming that the output of the physical information neural network PINN is , the output of the convolutional long short-term memory network ConvLSTM is The prediction result after fusion is , then the fusion formula of the attention mechanism is used: in, is the attention weight.

[0025] In particular, after information fusion, the uncertainty quantification module estimates the uncertainty of the prediction results through Bayesian methods or Monte Carlo sampling techniques to provide credibility estimates. Specifically, the uncertainty of the prediction results is quantified by probabilistic modeling of network weights or by running the model multiple times and calculating the statistical properties of the prediction results.

[0026] Example 4 Based on Example 3, the MLP structure in step 3 is specifically expressed as: in, is the MLP structure, For time, is the flow rate, is the inlet flow rate, is the stream function, For pressure, is the level set function.

[0027] In particular, the training of the physical information neural network PINN is to minimize the total loss function including the data fitting error and the penalty term for violating the laws of physics by the gradient descent method; the total loss function It is expressed as: in, is the data fitting error, is the penalty term for violating the laws of physics, is the weight coefficient for balancing the two items.

[0028] In particular, the update formula of the convolutional long short-term memory network ConvLSTM in step 4 is expressed as: Among them, all input , cell output , hidden state , and the door , , are all three-dimensional tensors, where the second two dimensions are spatial dimensions (rows and columns), are network parameters, is the activation function, is the hyperbolic tangent function, represents the convolution operation, Represents element-wise product.

[0029] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A microchannel gas-liquid two-phase flow pattern prediction system based on neural network, characterized in that: include: A data acquisition module, wherein the data acquisition module acquires simulation data of gas-liquid two-phase flow in the microchannel under different inlet flow rate conditions, and outputs field value information of the sampling point in each iteration process; A data preprocessing module, the data preprocessing module is connected to the data acquisition module and is responsible for cleaning and smoothing the obtained simulation data; A PINN module, wherein the PINN module is connected to the data preprocessing module and uses a physical information neural network PINN to learn local physical phenomena in the simulation data to ensure that the prediction results conform to known physical laws; A ConvLSTM module, wherein the ConvLSTM module is connected to the PINN module and uses a convolutional long short-term memory network ConvLSTM to capture long-time spans and large-scale spatial changes in the simulation data; An information fusion module, wherein the information fusion module connects the PINN module and the ConvLSTM module, integrates the output results of the PINN module and the ConvLSTM module to form a prediction result; Among them, the PINN module and the ConvLSTM module constitute a deep learning network.

2. The neural network-based microchannel gas-liquid two-phase flow pattern prediction system according to claim 1, characterized in that: The system also includes an adaptive incremental learning module, which is connected to the deep learning network, uses an online learning algorithm to update the parameters of the deep learning network, and dynamically adjusts the learning rate and weight update strategy to adapt to changing operating conditions.

3. The neural network-based microchannel gas-liquid two-phase flow pattern prediction system according to claim 1, characterized in that: The system also includes an uncertainty quantification module, which is connected to the deep learning network, estimates the uncertainty of the prediction result obtained by the information fusion module through a Bayesian method or a Monte Carlo sampling technique, and provides a credibility estimate of the prediction result.

4. The neural network-based microchannel gas-liquid two-phase flow pattern prediction system according to claim 1, characterized in that: The data acquisition module acquires the simulation data by means of Fluent simulation software; the field value information includes coordinates (x, y), static pressure, absolute velocities X Velocity and Y Velocity in the xy direction, and gas phase VOF.

5. The neural network-based microchannel gas-liquid two-phase flow pattern prediction system according to claim 1, characterized in that: The physical information neural network PINN in the PINN module includes multiple multi-layer perceptrons, which are used to predict stream function, pressure and level set function respectively; the physical information neural network PINN is trained by gradient descent method to minimize the total loss function including data fitting error and penalty term for violation of physical laws.

6. The neural network-based microchannel gas-liquid two-phase flow pattern prediction system according to claim 1, characterized in that: The convolutional long short-term memory network ConvLSTM in the ConvLSTM module includes multiple convolutional layers and LSTM units, receives a series of field value sequences at continuous time points as input, each time point corresponds to a two-dimensional flow field snapshot, and updates the weights through the back-propagation algorithm to predict the flow field state at future moments.

7. A method for predicting flow pattern of gas-liquid two-phase flow in microchannel based on neural network, characterized in that: The method comprises the following steps: Step 1, obtain the simulation data of gas-liquid two-phase flow in the microchannel under different inlet flow rate conditions, output the field value information of the sampling point in each iteration process, and store it in a .mat file; Step 2: cleaning and smoothing the obtained simulation data; Step 3: Using a physical information neural network PINN to learn local physical phenomena in the simulation data to ensure that the prediction results conform to known physical laws, wherein the Navier-Stokes equations are embedded in the physical information neural network PINN and include multiple multi-layer perceptron structures; Step 4: Input the prediction result of step 3 as an initial condition into a convolutional long short-term memory network ConvLSTM, wherein the convolutional long short-term memory network ConvLSTM also receives a sequence of field values ​​at a series of continuous time points, each time point corresponding to a two-dimensional flow field snapshot; the convolutional long short-term memory network ConvLSTM captures the spatiotemporal dependencies within a long time span through convolutional layers and LSTM units, thereby providing a global prediction of flow characteristics, and updates weights through a back-propagation algorithm; Step 5: After completing the training of the physical information neural network PINN and the convolutional long short-term memory network ConvLSTM, the output results of the two models are integrated into the final prediction result; assuming that the output of the physical information neural network PINN is , the output of the convolutional long short-term memory network ConvLSTM is , the prediction result after fusion is , then the fusion formula of the attention mechanism is used: in, is the attention weight.

8. The method for predicting flow pattern of gas-liquid two-phase flow in a microchannel based on a neural network as claimed in claim 7, characterized in that: The MLP structure in step 3 is specifically expressed as: in, is the MLP structure, For time, is the flow rate, is the inlet flow rate, is the stream function, For pressure, is the level set function.

9. The method for predicting flow pattern of gas-liquid two-phase flow in a microchannel based on a neural network as claimed in claim 8, characterized in that: In step 3, the training of the physical information neural network PINN is to minimize the total loss function including data fitting error and penalty term for violating physical laws by gradient descent method; the total loss function It is expressed as: in, is the data fitting error, is the penalty term for violating the laws of physics, is the weight coefficient for balancing the two items.

10. The method for predicting flow pattern of gas-liquid two-phase flow in microchannel based on neural network according to claim 7, characterized in that: The update formula of the convolutional long short-term memory network ConvLSTM in step 4 is expressed as: Among them, all input , cell output , hidden state , and the door , , are all three-dimensional tensors, and the last two dimensions are spatial dimensions. are network parameters, is the activation function, is the hyperbolic tangent function, represents the convolution operation, Represents element-wise product.

Images