Method for obtaining mechanical parameters of biological membrane in porous medium in situ and application of method

Abstract

The invention discloses a method for acquiring mechanical parameters of a biological membrane in a porous medium in situ and application of the method in deformation simulation and parameter inversion of the biological membrane in the porous medium. The mechanical parameters comprise Young modulus and Poisson's ratio. The method for acquiring the mechanical parameters of the biological membrane in the porous medium in situ comprises the following steps: acquiring deformation related data of the biological membrane in the porous medium, wherein the deformation related data comprises space-time coordinates, flow velocity, pressure and displacement data; using a deep neural network to construct a mapping relationship between the space-time coordinates and the flow velocity, pressure and displacement; and performing deep neural network model training by calculating and minimizing the loss function to obtain the mechanical parameters of the biological membrane in the porous medium. According to the method, the deformation response and the mechanical property of the biological membrane can be rapidly and accurately obtained through the deep learning algorithm on the premise that the original medium structure is not damaged, the method is suitable for multiple fields such as water quality control, bioremediation and soil remediation, and theoretical basis and technical support are provided for related application.

Description

Technical Field

[0001] The present invention relates to the field of microbial research and machine learning simulation, and in particular to a method for in-situ acquisition of mechanical parameters of biofilms in porous media and an application thereof. Background Art

[0002] The study of biofilm deformation simulation is one of the key technologies in understanding biofilm behavior. By deeply understanding the deformation response of biofilm under external forces (for example, forces caused by fluid flow, particle or bubble impact, etc.), it can provide a scientific basis for the application of biofilm in bioremediation, pollutant filtration, and water quality control. The mechanical properties of biofilms, such as Young's modulus and Poisson's ratio, as important parameters of linear elastic materials, have gradually become the focus of research. However, since the mechanical properties of biofilms are affected by many factors, it is currently impossible to directly determine the mechanical properties of biofilms through simple prediction models, and specific experimental methods are required to measure them. In order to better understand and predict the deformation behavior of biofilms in porous media, researchers have proposed some models based on numerical simulation methods such as computational fluid dynamics (CFD) and finite element analysis (FEA). However, these models are usually computationally intensive when dealing with large-scale, multi-scale complex problems, and require a large amount of experimental data for verification.

[0003] Therefore, it is very important to quickly and accurately simulate the deformation of biofilms in porous media and evaluate the mechanical properties of biofilms based on the simulation results.

[0004] The reason for using models is that it is not clear how quickly the mechanical properties of biofilms will change after being removed from their original environment, so ex situ measurements of biofilms have great limitations, such as Figure 1 As shown, under the incompressible assumption, according to the conservation of mass and momentum, the fluid motion in the cavity is described by the following Navier-Stokes equations:

[0005]

[0006] where ρ f is the fluid density, u is the cavity flow velocity, t is the time, σ f is the stress tensor, p is the pressure, μ is the dynamic viscosity, and I is the identity matrix.

[0007] In biological membranes, Darcy's law is used to describe fluid motion:

[0008]

[0009]

[0010] Where u is the flow velocity in the biofilm, p is the pressure, μ is the dynamic viscosity, and κ is the permeability.

[0011] For the biofilm, it is assumed to be a linear elastic material, and the equilibrium equation of the biofilm is:

[0012]

[0013] where ρ s is the biofilm density, v s is the displacement, t is the time, σ s is the stress tensor of the biofilm, f is the volume force on the biofilm, and for linear elastic materials, considering the pore pressure in the biofilm, the relationship between stress and strain is:

[0014] σ s =2μ s ε+λ s tr(ε)I-αpI

[0015] where μ s , s is the Lamé parameter, which can be expressed by Young's modulus (E) and Poisson's ratio (v). ε is the strain tensor of the biofilm, I is the identity matrix, α is the Biot-Willis coefficient, which is assumed to be 1 here, and p is the pressure. Summary of the invention

[0016] The present invention provides a method for in-situ acquiring mechanical parameters of biofilms in porous media and its application, by constructing a mapping relationship between time-space coordinates and flow velocity, pressure, and displacement through a deep neural network, and inverting the mechanical parameters of the biofilm, Young's modulus and Poisson's ratio, as learnable parameters to simulate the deformation of the biofilm, thereby acquiring the Young's modulus and Poisson's ratio of the biofilm in situ.

[0017] [1] A method for in situ acquisition of mechanical parameters of a biofilm in a porous medium, wherein the mechanical parameters include Young's modulus and Poisson's ratio, the method comprising:

[0018] Obtain data related to biofilm deformation in porous media, including time and space coordinates, flow velocity, pressure and biofilm displacement data;

[0019] Use deep neural networks (DNNs) to construct implicit conversion relationships between spatiotemporal coordinates and flow velocity, pressure, and biofilm displacement;

[0020] The deep neural network parameters are determined by calculating and minimizing the loss function L of the deep neural network, and the deep neural network model training is completed at the same time; during the training process, the mechanical parameters of the biofilm in the porous medium are inverted; the loss function L is expressed as follows:

[0021] L=λ data L data +λ physics L physics

[0022]

[0023] Among them, L data is the data loss, L physics is the total physical constraint, λ data and λ physics are the corresponding weight coefficients, u, p, v s Respectively represent the acquired flow rate, pressure, and biofilm displacement data, Denote the deep neural network prediction values ​​of flow velocity, pressure, and biofilm displacement, respectively, and λ u , p , s are the weight coefficients of the corresponding types of data fitting losses, f1, f2, f3, f4, f5 represent the residuals of the differential equations of velocity and stress conditions of velocity field, pressure field, displacement field and fluid-solid boundary, λ1, λ2, λ3, λ4, λ5 are the weight coefficients of the corresponding differential equation residuals;

[0024]

[0025] Among them, ρ f is the fluid density, t represents time, σ f represents the stress tensor of the fluid, Ω f represents the cavity flow domain, T represents the maximum time, κ represents the permeability, μ represents the dynamic viscosity, Ω s represents the biofilm domain, ρ s is the biofilm density, σ s represents the stress tensor of the biofilm, f s is the volume force on the biofilm, represents the boundary between the cavity domain and the biofilm domain, n f The unit external normal vector of the cavity flow domain and the biofilm boundary is from the flow domain to the biofilm domain, n s The unit external normal vector of the biofilm domain, which represents the boundary between the cavity flow domain and the biofilm, points from the biofilm domain to the flow domain; the fluid and biofilm stress tensors are calculated using the following relationship:

[0026]

[0027] Considering the pore pressure in the biofilm,

[0028]

[0029] Where: I is the identity matrix; α is the Biot-Willis coefficient (for example, it can be assumed to be 1, etc.); μ s , s is the Lamé parameter, expressed by Young's modulus E and Poisson's ratio v:

[0030]

[0031] In some embodiments, the input of the deep neural network model is the space-time coordinates (x, y, t), where x and y are the horizontal and vertical coordinates of the corresponding point of the two-dimensional experimental image in the Cartesian coordinate system, and t is time; the output of the deep neural network is the predicted field

[0032] In some embodiments, the method may further include fabrication of a microfluidic chip device and flow experiments;

[0033] The microfluidic chip device is obtained by photolithography and casting (for example, using experimental tools such as a photolithography machine, a glue spreader, a glue drying machine, and a developing device), and can represent a physical model of the spatial structure of a porous medium;

[0034] The flow experiment uses a programmable pumping device to inject bacterial liquid and nutrient solution from the inlet into the microfluidic chip device at a constant flow rate according to actual conditions, and changes the flow rate to obtain biofilm deformation related data.

[0035] In some embodiments, the photolithography is to evenly coat a layer of photoresist of a certain thickness on a substrate, heat and cool it, and then perform ultraviolet exposure. During exposure, a mask is placed between a light source and the photoresist. When ultraviolet light passes through the mask, it selectively irradiates the photoresist, causing the illuminated photoresist to undergo a cross-linking reaction, thereby changing the solubility of this part of the photoresist in a developer, and then the excess photoresist is washed away with the developer.

[0036] In some embodiments, the casting is to cast a high molecular polymer (such as polydimethylsiloxane, etc.) onto the photolithographic substrate, and then peel and slice it after solidification to transfer the pore structure to the microfluidic chip device.

[0037] In some embodiments, the microfluidic chip device may include a glass slide and a microfluidic chip;

[0038] Two holes are punched on the microfluidic chip as the liquid inlet and outlet, and then the glass slide is attached to the microfluidic chip to form a closed pore channel between the glass slide and the microfluidic chip; the cavity formed by the glass slide and the microfluidic chip contains pores and throats, which can represent the pore structure of the porous medium.

[0039] In some embodiments, the flow experiment uses a pumping device to deliver a certain concentration of bacterial liquid, which is left to stand for a period of time (e.g., 2 hours) to allow the microorganisms to fully adhere to the microfluidic chip device. After the biofilm grows, the flow rate is changed, and a set of automatic imaging devices is used to automatically capture the deformation process of the biofilm in the microfluidic chip device, obtain an image of the porous medium with the biofilm, use an integrated microsensor to collect flow rate and pressure data in the microfluidic chip device, and process the image to obtain displacement field data.

[0040] In some embodiments, the method may adopt a deep learning simulation system embedded with a deep neural network model. The deep learning simulation system includes a control terminal and a server for data processing and deep neural network model training. After the server collects biofilm deformation-related data and user-set parameters, it trains the deep neural network model by calculating and minimizing the loss function L, and returns the results to the user; the results include the inverted biofilm mechanical parameters in the porous medium and the predicted flow rate, pressure, biofilm displacement data and related visualization images.

[0041] Further, the control terminal can be used to configure relevant parameters of deep neural network model training, including training parameters (such as the data set used, the number of iterations, etc.) and the ratio of the training set, the test set, and the validation set. The result returned by the receiving server is the flow velocity, pressure, and displacement corresponding to the space-time coordinates (specifically, it can be displayed in the form of images, charts, etc.);

[0042] When the server performs deep neural network model training, it first performs data preprocessing, performs threshold segmentation on the porous medium image with biofilm through RGB value, divides the solid phase, liquid phase and membrane phase, integrates the acquired velocity field, pressure field and displacement field into a data set, and divides it into a training set, a verification set and a test set; the deep neural network model is trained using the training set according to the parameters set by the user through a random or optimized network parameter initialization method, and the trained deep neural network model is sent to the control terminal and stored in a MySQL database together with the training data.

[0043] [2] Application of the method described in [1] to the simulation of biofilm deformation and parameter inversion in porous media. The method of the present invention can be used to simulate the deformation of biofilm in porous media under the action of external forces.

[0044] The present invention can quickly predict the deformation process of biofilm in porous media from limited experimental data, and evaluate its mechanical properties, such as Young's modulus and Poisson's ratio. The advantages of the present invention are: it can quickly and accurately obtain the deformation response and mechanical properties of biofilm through deep learning algorithm without destroying the original medium structure; it can effectively improve the model prediction accuracy with a small amount of experimental data by introducing physical constraint terms in the deep neural network loss function to replace part of the experimental data to guide the deep neural network training; it is applicable to multiple fields such as water quality control, bioremediation, soil remediation, etc., and provides theoretical basis and technical support for related applications.

[0045] The present invention provides a method for rapidly and accurately simulating the deformation of biofilms in porous media. The method can effectively predict the deformation response of biofilms under different fluid mechanics and external forces, thereby providing a scientific basis for applications in the fields of water quality control and bioremediation, and can solve the problem of difficulty in simulating the deformation process of biofilms in porous media under external forces.

[0046] For example, the method of the present invention may include: simulating the deformation process of the biofilm in a porous medium by microfluidic technology, and monitoring the deformation of the biofilm in real time in combination with deep learning technology, so as to obtain the mechanical properties and deformation characteristics of the biofilm in an external environment (such as fluid flow, particle impact, etc.).

[0047] In some cases, the microfluidic chip device of the present invention may include a glass slide, a microfluidic chip, and a plurality of integrated microsensors. The integrated microsensors may be installed at appropriate positions on the glass slide. By changing the injection speed of the nutrient solution, the integrated microsensors can measure the fluid flow rate and pressure during the deformation of the biofilm in the microfluidic chip device, and the biofilm will undergo relevant displacement changes as the injection time passes.

[0048] The present invention fully considers the physical constraints followed by fluids and biofilms in the process of training deep neural networks, and implements this by using the residuals of the constitutive equations in the process model as additional regularization constraints. The residuals of the equations of multiple physical processes constitute the total physical constraint L physics .

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] The present invention can utilize microfluidic technology to simulate the deformation of biofilms in porous media under different flow rates, use a microscopic imaging system and sensors to obtain the velocity field, pressure field and displacement field in the experiment, and upload the images and other data to an analysis system based on deep learning technology. A mapping relationship between input and output is constructed through a deep neural network to obtain biofilm deformation prediction and obtain relevant biofilm mechanical performance parameters.

[0051] The present invention can evaluate the mechanical properties according to the biofilm displacement field without destroying the original biofilm growth process. Based on a deep neural network model with physical constraints, it can predict the displacement field of the biofilm in the pores in a short time when the experiment can only generate less data.

[0052] Therefore, the present invention can simply and quickly evaluate the mechanical properties of biofilms in porous media, evaluate the shear resistance of biofilms in environments such as filter membranes and farmland, predict the evolution of biofilm morphology in natural porous media such as soil, and provide a theoretical basis for the implementation of bioremediation in these natural environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 Schematic diagram of biofilm attachment;

[0054] Figure 2 Schematic diagram of the process and structure of the microfluidic experiment;

[0055] Figure 3 This is a schematic diagram of the deep neural network training optimization control process of the present invention;

[0056] Figure 4 Schematic diagram of the principle of deep neural network model. DETAILED DESCRIPTION

[0057] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0058] 1. Fabrication of microfluidic chip device:

[0059] 1) Prepare the mask: Make the two-dimensional pore model into an optical mask.

[0060] 2) Photolithography development: Use a coating machine to spread a 50μm thick layer of SU-8 2050 photoresist on a single-sided polished silicon wafer. Use a hot plate to pre-bake the photoresist to solidify it, then use a photolithography machine to perform UV exposure, followed by post-bake. After the silicon wafer cools, use a developer to develop it and wash away the excess photoresist. After cleaning the silicon wafer, use trimethylchlorosilane (TMCS) to fumigate the wafer overnight to obtain a chip mold.

[0061] 3) PDMS pouring: Sylgard 184 silicone base material and curing agent were stirred and mixed in a weight ratio of 10:1 to prepare PDMS (polydimethylsiloxane), and after removing bubbles under a vacuum environment, the mixture was poured into a mold and cured at 80° C. for 2 hours.

[0062] 4) Sensor installation: Fix several integrated micro sensors at the corresponding large pore positions of the glass slide. The sensors can collect the flow velocity and pressure at the point.

[0063] 5) Chip bonding: After cooling, the PDMS is peeled off from the mold, sliced ​​and punched, and the glass slide and PDMS are exposed to oxygen plasma for 30 seconds, and then the PDMS is bonded to the glass slide. The chip is then placed in a 70°C oven for ten minutes to make the PDMS and glass slide bond more closely.

[0064] 2. Flow test, see Figure 2 :

[0065] 1) Preparation of bacterial solution: Bacillus subtilis CMCC (B) 63501 is a microorganism used to study biofilms. It was inoculated into 50 ml of nutrient solution (components: tryptone 10 g / L, yeast powder 5 g / L, NaCl 10 g / L). The bacterial solution was cultured at a constant temperature of 27°C and a shaking speed of 180 rpm. On the day of the experiment, the solution was diluted in fresh nutrient solution at a ratio of 1:10.

[0066] 2) Prepare the experimental device: Connect the polyethylene plastic hose to the inlet and outlet of the microfluidic chip after connecting the L-shaped steel needle, fix the chip under the microscope, and select the appropriate magnification. Use a syringe to pass 75 vol% alcohol into the chip to sterilize and remove bubbles, then pass sterile water to wash away excess ethanol, and then pass fresh nutrient solution to create an environment suitable for microbial growth.

[0067] 3) Run the experiment: Inhale the bacterial solution into the syringe, remove the bubbles in the syringe, pass it into the microfluidic chip, and let it stand for 2 hours to allow the microorganisms to fully adhere to the microfluidic channel. Then, inhale fresh nutrient solution with the syringe, fix the syringe on the syringe pump (Harvard Apparatus, Pump 11), set the appropriate flow rate, and then start fluid pumping.

[0068] 4) Image acquisition: After the biofilm has grown for 24 hours, the injection flow rate is changed. During the deformation of the biofilm, the automatic microscopic imaging system captures images of the biofilm deformation at intervals of 1 second. The displacement field data of the biofilm is obtained through the Ncorr software. At the same time, the sensor collects the flow rate and pressure data in the chip and uploads them to the server.

[0069] 3. Model training, see Figure 3 :

[0070] Model training using a deep learning simulation system includes the following steps:

[0071] 1) The user enters the parameters of this training in the control terminal, including the training parameters such as the dataset used, the number of iterations, and the ratio of the training set, test set, and validation set. The input terminal is a graphical user interface developed in Java language, and transmits data with the server using TCP / IP protocol. The terminal can display displacement image data, and the user can exclude image data with poor imaging quality when selecting a dataset.

[0072] 2) The model training server is configured with an open source CentOS system, MySQL database, Tomcat server, Python development environment, etc. The images captured by the microscopic automatic imaging system are first segmented into liquid phase, membrane phase and solid phase by a Python script, and then the Ncorr software is used to obtain the displacement field data from the segmented images and store the data collected by the sensor in the MySQL database.

[0073] 3) The neural network-based deep learning algorithm is written in Python. According to the data set selected by the user from the control terminal, the selected data is obtained through the CentOS Crontab service. The input of the model is the space-time coordinates (x, y, t), and the output is the predicted physical field First, we need to analyze the actual observation data (u, p, v s ) for normalization:

[0074]

[0075] Among them, U is the original measured data to be normalized, X is the original space-time coordinate data; X max / U max and X min / U min are the maximum and minimum values ​​in the data, respectively. is the normalized data. Then, the processed data is Divide into training set, validation set and test set.

[0076] 4) Use the DNN architecture to construct a mapping between spatiotemporal coordinates and the observation field. It consists of multiple linear transformations and nonlinear activation functions. A linear transformation It can be expressed as:

[0077] T l (X l-1 ):=W l X l-1 +b l

[0078] in, It represents the weight of the input from the previous layer l-1 of the neural network to the current layer l. represents the bias of the current lth layer, X0 Represents the initial input of the neural network; plus an activation function The lth layer of the neural network It can be expressed as:

[0079] N l (X l-1 ):=σ(T l (X l-1 ))

[0080] Therefore, a general L-layer DNN neural network can be expressed as:

[0081]

[0082] Where Θ represents all weights W l and bias b l .like Figure 4 As shown, the training set is applied to the network training process to obtain the data fitting loss:

[0083]

[0084] Among them, λ u ,λ p ,λ s are the weight coefficients of the corresponding type of data fitting loss, u, p, v s Respectively represent the actual observed flow velocity, pressure, and displacement data, Represent the neural network prediction values ​​of flow velocity, pressure and displacement respectively.

[0085] 5) Calculate the residual of the differential equation. The residual of the equation can be expressed as:

[0086]

[0087] Among them, ρ f is the fluid density, t represents time, σ f represents the stress tensor of the fluid, Ω f represents the cavity flow domain in porous media, T represents the maximum time, k represents the permeability, μ represents the dynamic viscosity, Ω s represents the biofilm domain in the porous medium, ρ s is the biofilm density, σ s represents the stress tensor of the biofilm, f s is the volume force on the biofilm, represents the boundary between the cavity domain and the biofilm domain, n f The unit external normal vector of the cavity flow domain and the biofilm boundary is from the flow domain to the biofilm domain, n sThe unit external normal vector of the biofilm domain represents the boundary between the cavity domain and the biofilm. The stress tensor of the fluid and the biofilm is calculated by the following relationship:

[0088]

[0089] Considering the pore pressure in the biofilm,

[0090]

[0091] in is the predicted pressure; I is the unit matrix; α is the Biot-Willis coefficient, which is assumed to be 1 here; μ s , s is the Lamé parameter, expressed by Young's modulus E and Poisson's ratio v:

[0092]

[0093] At the same time, the parameters representing the mechanical properties of the biomembrane, Young's modulus and Poisson's ratio, are iteratively updated as trainable parameters of the neural network to achieve a unified solution to the forward and inverse problems.

[0094] The residual of the above control equation As an additional regularization constraint to guide the training of the neural network, it is added to the loss function of the neural network to minimize the residual of the equation so that the output of the neural network satisfies the PDE as much as possible. The total physical mechanism constraint composed of the residual of each equation is:

[0095]

[0096] Among them, λ1, λ2, λ3, λ4 and λ5 are the weight coefficients of the corresponding differential equation residuals;

[0097] Finally, the data fitting loss L data and the PDE residual term L physics The sum of the two parts gives the total loss function:

[0098] L=λ data L data +λ physics L physics

[0099] In the formula, λ data and λ physics are the relevant weight coefficients respectively. Before training, Glorot is used to initialize the model parameters, and the Adam optimization algorithm minimizes the loss function L, so that the neural network prediction meets both the data matching requirements and the PDE description. After multiple iterative training, the prediction model is finally obtained.

[0100] The model training results are sent to the control terminal, and a copy is also stored in the MySQL database. The data returned to the control terminal includes visualization images reflecting the training effect, model prediction accuracy evaluation indicators, mechanical performance parameter inversion values, and network parameters of the trained model.

[0101] The present invention combines microfluidic technology to produce a microscale chip that can represent the pore space structure of the actual porous medium, and relies on the microfluidic chip to carry out a microscopic flow simulation experiment. The processed data of the image in the flow experiment and the sensor data are uploaded to a simulation system based on deep learning, and the neural network model training is completed in the system by minimizing the loss function, and the obtained training model results are returned to the terminal for visual presentation. Therefore, the present invention can quickly and accurately simulate the deformation results of the biofilm in the medium and invert the mechanical parameters of the biofilm through simulation experiments and deep learning algorithms without destroying the original medium, providing a feasible technical reference for the implementation of environmental engineering such as bioremediation and aquatic ecosystem health judgment.

[0102] In addition, it should be understood that after reading the above description of the present invention, those skilled in the art may make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the claims attached to this application.

Claims

1. A method for in situ acquisition of mechanical parameters of biofilm in porous media, wherein the mechanical parameters include Young's modulus and Poisson's ratio, characterized in that: The method comprises: Obtain data related to biofilm deformation in porous media, including time and space coordinates, flow velocity, pressure and biofilm displacement data; Use deep neural networks to construct implicit conversion relationships between spatiotemporal coordinates and flow velocity, pressure, and biofilm displacement; The deep neural network parameters are determined by calculating and minimizing the loss function L of the deep neural network, and the deep neural network model training is completed at the same time; during the training process, the mechanical parameters of the biofilm in the porous medium are inverted; the loss function L is expressed as follows: L=λ data L data +λ physics L physics Among them, L data is the data loss, L physics is the total physical constraint, λ data and λ physics are the corresponding weight coefficients, u, p, v s Respectively represent the acquired flow rate, pressure, and biofilm displacement data, Denote the deep neural network prediction values ​​of flow velocity, pressure, and biofilm displacement, respectively, and λ u , p , s are the weight coefficients of the corresponding types of data fitting losses, f1, f2, f3, f4, f5 represent the residuals of the differential equations of velocity and stress conditions of velocity field, pressure field, displacement field and fluid-solid boundary, λ1, λ2, λ3, λ4, λ5 are the weight coefficients of the corresponding differential equation residuals; Among them, ρ f is the fluid density, t represents time, σ f represents the stress tensor of the fluid, Ω f represents the cavity flow domain, T represents the maximum time, κ represents the permeability, μ represents the dynamic viscosity, Ω s represents the biofilm domain, ρ s is the biofilm density, σ s represents the stress tensor of the biofilm, f s is the volume force on the biofilm, represents the boundary between the cavity domain and the biofilm domain, n f The unit external normal vector of the cavity flow domain and the biofilm boundary is from the flow domain to the biofilm domain, n s The unit external normal vector of the biofilm domain, which represents the boundary between the cavity flow domain and the biofilm, points from the biofilm domain to the flow domain; the fluid and biofilm stress tensors are calculated using the following relationship: Considering the pore pressure in the biofilm, Where I is the identity matrix; α is the Biot-Willis coefficient; μ s , s is the Lamé parameter, expressed by Young's modulus E and Poisson's ratio v:

2. The method according to claim 1, characterized in that The input of the deep neural network model is the space-time coordinates (x, y, t), where x and y are the horizontal and vertical coordinates of the corresponding points of the two-dimensional experimental image in the Cartesian coordinate system, and t is the time; the output of the deep neural network is the predicted field.

3. The method according to claim 1, characterized in that The method also includes the fabrication of a microfluidic chip device and flow experiments; The microfluidic chip device is obtained by photolithography and casting, and can represent a physical model of the spatial structure of the porous medium; The flow experiment uses a programmable pumping device to inject bacterial liquid and nutrient solution from the inlet into the microfluidic chip device at a constant flow rate according to actual conditions, and changes the flow rate to obtain biofilm deformation related data.

4. The method according to claim 3, characterized in that: The photolithography is to evenly coat a layer of photoresist of a certain thickness on a substrate, heat and cool it, and then perform ultraviolet exposure. During exposure, a mask is placed between a light source and the photoresist. When ultraviolet light passes through the mask, it selectively irradiates the photoresist, causing the irradiated photoresist to undergo a cross-linking reaction, changing the solubility of this part of the photoresist in a developer, and then washing away the excess photoresist with the developer; The pouring method is to pour the high molecular polymer onto the photoetched substrate, and then peel and slice it after solidification to transfer the pore structure into the microfluidic chip device.

5. The method according to claim 3, characterized in that: The microfluidic chip device comprises a glass slide and a microfluidic chip; Two holes are punched on the microfluidic chip as the liquid inlet and outlet, and then the glass slide is attached to the microfluidic chip to form a closed pore channel between the glass slide and the microfluidic chip; the cavity formed by the glass slide and the microfluidic chip contains pores and throats, which can represent the pore structure of the porous medium.

6. The method according to claim 3 or 5, characterized in that: The flow experiment uses a pumping device to deliver a certain concentration of bacterial liquid, which is left to stand for a period of time to allow the microorganisms to fully adhere to the microfluidic chip device. After the biofilm grows, the flow rate is changed, and a set of automatic imaging devices is used to automatically capture the deformation process of the biofilm in the microfluidic chip device, obtain an image of the porous medium with the biofilm, use an integrated microsensor to collect flow rate and pressure data in the microfluidic chip device, and process the image to obtain displacement field data.

7. The method according to claim 1, characterized in that The method adopts a deep learning simulation system embedded with a deep neural network model. The deep learning simulation system includes a control terminal and a server for data processing and deep neural network model training. After the server collects biofilm deformation-related data and user-set parameters, it trains the deep neural network model by calculating and minimizing the loss function L, and returns the results to the user; the results include the inverted biofilm mechanical parameters in the porous medium and the predicted flow rate, pressure, biofilm displacement data and related visualization images.

8. The method according to claim 7, characterized in that The control terminal is used to configure the relevant parameters of the deep neural network model training, including the training parameters and the ratio of the training set, the test set and the validation set, and the result returned by the receiving server is the flow velocity, pressure and biofilm displacement corresponding to the time and space coordinates; When the server performs deep neural network model training, it first performs data preprocessing, performs threshold segmentation on the porous medium image with biofilm through RGB value, divides the solid phase, liquid phase and membrane phase, integrates the acquired velocity field, pressure field and displacement field into a data set, and divides it into a training set, a verification set and a test set; the deep neural network model is trained using the training set according to the parameters set by the user through a random or optimized network parameter initialization method, and the trained deep neural network model is sent to the control terminal and stored in a MySQL database together with the training data.

9. Application of the method according to any one of claims 1 to 8 in biofilm deformation simulation and parameter inversion in porous media.

Images