Search

CN-121980978-A - Membrane channel biological pollution multi-physical field simulation method based on physical information deep learning

CN121980978ACN 121980978 ACN121980978 ACN 121980978ACN-121980978-A

Abstract

The invention discloses a membrane channel biological pollution multi-physical field simulation method based on physical information deep learning, which comprises the steps of obtaining sampling data and mapping the sampling data into a biological membrane morphology characterization field which is continuously distributed along with space coordinates, constructing a deep neural network model, taking the space coordinates of a membrane channel as input, taking multi-physical-field state variables in the membrane channel as output, establishing a mapping relation from the space coordinates to the physical fields, constructing a physical information driven composite loss function, wherein the composite loss function comprises a physical equation residual error for restraining the physical field to meet a preset physical conservation law and a boundary condition residual error item for restraining boundary conditions, enabling physical field distribution output by the model to simultaneously accord with physical evolution law in the membrane channel and biological membrane pollution characteristics, inputting the space coordinates of a region to be predicted into the trained deep neural network model, and obtaining full-field physical quantity distribution information in the polluted membrane channel.

Inventors

  • LUO JIU
  • QIAO DAN
  • JIANG YANPENG
  • ZHANG HAO

Assignees

  • 苏州大学

Dates

Publication Date
20260505
Application Date
20260409

Claims (9)

  1. 1. A membrane channel biological pollution multi-physical field simulation method based on physical information deep learning is characterized by comprising the following steps: Step 1, acquiring sampling data representing biological pollution distribution characteristics in a membrane channel, and mapping the sampling data into a biological membrane morphology representation field which is continuously distributed along with a space coordinate; step 2, constructing a deep neural network model, taking the space coordinates of the membrane channel as input, taking the state variables of multiple physical fields in the membrane channel as output, and establishing a mapping relation from the space coordinates to the physical fields; step 3, constructing a physical information driven composite loss function, wherein the composite loss function comprises a physical equation residual error used for restraining a physical field to meet a preset physical conservation law and a boundary condition residual error item used for restraining a boundary condition; embedding the biomembrane morphology characterization field into a physical equation residual error term, training a deep neural network model by utilizing a composite loss function, and enabling the physical field distribution output by the model to simultaneously accord with a physical evolution rule and biomembrane pollution characteristics in a membrane channel by minimizing the composite loss function; And 5, inputting the spatial coordinates of the region to be predicted into a trained deep neural network model to obtain the distribution information of the full-field physical quantity in the polluted membrane channel.
  2. 2. The membrane channel biological pollution multi-physical field simulation method based on physical information deep learning as claimed in claim 1, wherein step 1 further comprises smoothing continuous treatment of biological membrane characteristics, namely, a smooth transition function is adopted to convert discrete biological membrane boundaries into continuous physical parameter fields, so that smooth transition of various parameters in a physical equation at a fluid-biological membrane interface is realized, and numerical oscillation caused by abrupt interface parameter changes is avoided.
  3. 3. The membrane channel biological pollution multi-physical field simulation method based on physical information deep learning of claim 1, wherein in step 2, the multi-physical field state variables include a velocity field, a pressure field and a substrate concentration field, and are used for characterizing hydrodynamic behavior and solute mass transfer behavior under the influence of biological pollution.
  4. 4. The membrane channel biological pollution multi-physical field simulation method based on physical information deep learning of claim 1, wherein in step 2, training of a deep neural network model adopts a staged optimization strategy, and first, a global search capacity optimization algorithm is adopted to perform preliminary optimization, so that the model is rapidly positioned to a physical reasonable area, and then, the model is switched to a local accurate locking capacity optimization algorithm to perform fine solution, so that prediction accuracy is improved.
  5. 5. The membrane channel biological pollution multi-physical-field simulation method based on physical information deep learning of claim 1 is characterized in that in step 3, a physical equation residual term comprises a fluid dynamics residual term, a mass transfer and biochemical reaction residual term and a mass conservation constraint term, wherein the fluid dynamics residual term is used for describing fluid motion behaviors of a flow channel region and a biological membrane region in a membrane channel, the mass transfer and biochemical reaction residual term is used for describing a solute transportation process under the driving of a flow field and the nutrient intake and consumption process of the biological membrane, and the mass conservation constraint term is used for forcedly predicting that the physical field meets a fluid continuity equation.
  6. 6. The membrane channel biological pollution multi-physical field simulation method based on physical information deep learning of claim 5, wherein the fluid dynamics residual terms are constructed based on a modified momentum equation introducing a porous medium resistance mechanism, and the unified mathematical description of a clean flow channel region and a biological membrane pollution region is realized by automatically loading different resistance coefficients in different regions.
  7. 7. The membrane channel biological pollution multi-physical field simulation method based on physical information deep learning of claim 5, wherein mass transfer and biochemical reaction residual terms are constructed based on a convection-diffusion-reaction equation coupling biochemical reaction kinetics, and are used for quantitatively describing solute concentration gradient changes in a biological membrane due to microbial metabolic activities, so that bidirectional physical coupling of a fluid field and a concentration field is realized.
  8. 8. The membrane channel biological pollution multi-physical field simulation method based on physical information deep learning according to claim 1, wherein in the step 3, boundary condition residual terms comprise inlet flow rate constraint, outlet pressure constraint, solid wall non-slip constraint and membrane surface flux constraint based on solute mass balance, and the membrane surface flux constraint based on solute mass balance is used for forcing a model prediction result to conform to a real physical operation environment in a membrane channel.
  9. 9. The membrane channel biological pollution multi-physical field simulation method based on physical information deep learning according to claim 1, wherein the method further comprises migration learning, namely after training under standard working conditions to obtain a basic model, when a new pollution scene or geometric configuration is faced, migrating weight parameters of the basic model to the new model as an initialization weight, and then utilizing boundary conditions and physical constraints under the new working conditions to carry out quick fine adjustment so as to realize quick adaptation to multi-scene and multi-configuration working conditions.

Description

Membrane channel biological pollution multi-physical field simulation method based on physical information deep learning Technical Field The invention relates to the technical field of reverse osmosis water treatment, in particular to a membrane channel biological pollution multi-physical field simulation method based on physical information deep learning. Background Reverse osmosis and nanofiltration technology is the mainstream technology of global sea water desalination, industrial pure water preparation and wastewater reuse at present. However, microbial contamination is always the "kari butt" that restricts the long-term stable operation of the membrane system. Microorganisms adhere to the surface of a separator mesh in a membrane channel and grow to form a biological membrane, so that not only can flow channel blockage be caused and system pressure drop be increased sharply (energy consumption is increased), but also serious concentration polarization phenomenon can be caused, and water yield attenuation and desalination rate are reduced. In recent years, deep learning has demonstrated potential in the field of fluid prediction, but its direct application to membrane fouling simulation still faces significant challenges. On the one hand, the traditional convolutional neural network or the cyclic neural network lacks physical constraint, and the predicted result of the convolutional neural network or the cyclic neural network often does not meet the conservation of mass or momentum, so that the predicted value against the common physical knowledge is very easy to generate. At the same time, the pure data driven model requires thousands of high quality CFD or experimental data to train. However, under the actual engineering background of biological pollution, the method is limited by the dynamic evolution of the flow channel geometry and the limitation of multiple physical field measurement means, and the acquisition of large-scale and comprehensive physical field labeling data is extremely high in cost and even difficult to realize under a plurality of extreme working conditions. Disclosure of Invention The invention aims to provide a membrane channel biological pollution multi-physical field simulation method based on physical information deep learning, and provides an intelligent method which does not need massive labeling data, can meet the law of physical conservation, and can improve the speed field, the pressure field and the concentration field distribution in a membrane channel by an order of magnitude under biological pollution at a calculation speed, thereby solving the problems in the background art. The membrane channel biological pollution multi-physical field simulation method based on physical information deep learning comprises the following steps: Step 1, acquiring sampling data representing biological pollution distribution characteristics in a membrane channel, and mapping the sampling data into a biological membrane morphology representation field which is continuously distributed along with a space coordinate; step 2, constructing a deep neural network model, taking the space coordinates of the membrane channel as input, taking the state variables of multiple physical fields in the membrane channel as output, and establishing a mapping relation from the space coordinates to the physical fields; step 3, constructing a physical information driven composite loss function, wherein the composite loss function comprises a physical equation residual error used for restraining a physical field to meet a preset physical conservation law and a boundary condition residual error item used for restraining a boundary condition; embedding the biomembrane morphology characterization field into a physical equation residual error term, training a deep neural network model by utilizing a composite loss function, and enabling the physical field distribution output by the model to simultaneously accord with a physical evolution rule and biomembrane pollution characteristics in a membrane channel by minimizing the composite loss function; And 5, inputting the spatial coordinates of the region to be predicted into a trained deep neural network model to obtain the distribution information of the full-field physical quantity in the polluted membrane channel. Furthermore, step 1 further comprises smoothing and continuous treatment of the biomembrane characteristics, namely, a smooth transition function is adopted to convert a discrete biomembrane boundary into a continuous physical parameter field, so that smooth transition of each parameter in a physical equation at the interface between fluid and biomembrane is realized, and numerical oscillation caused by abrupt change of the interface parameter is avoided. Further, in step2, the multiple physical field state variables include a velocity field, a pressure field, and a substrate concentration field, which are used to characterize the hydrodynamic behavior