Search

CN-121997801-A - Multi-physical-quantity dynamic weighting loss construction method for deep learning flow field prediction

CN121997801ACN 121997801 ACN121997801 ACN 121997801ACN-121997801-A

Abstract

The invention discloses a multi-physical-quantity dynamic weighting loss construction method for deep learning flow field prediction, which is characterized in that based on flow field data obtained by real or numerical simulation, various physical-quantity characteristics such as vortex quantity, shear rate, energy dissipation rate and the like are extracted, unified-scale physical characteristic input is formed through dimensionless treatment, a local weighting function is constructed according to a multi-physical-quantity fusion relation, dynamic updating of a weighting coefficient is realized in a time recurrence mode, the obtained dynamic weight is introduced into a loss function of a deep learning model, and the model can carry out key constraint on error sensitive areas such as vortex shedding areas, shear layers and the like in the training process, so that the local precision and long-term time sequence stability of unsteady flow field prediction are improved. The method has clear structure, can be matched with various deep learning flow field prediction networks to be used, and has good universality and engineering applicability.

Inventors

  • YE PENGCHENG
  • HE YUE
  • HUANG QIAOGAO
  • SHI YAO
  • PAN GUANG
  • YIN YAFEN
  • ZHOU ZIHAN

Assignees

  • 西北工业大学

Dates

Publication Date
20260508
Application Date
20251223

Claims (9)

  1. 1. A method for constructing dynamic weighting loss of multiple physical quantities by deep learning flow field prediction is characterized by comprising the following steps: Step 1, calculating multiple physical quantities; Extracting multiple physical quantity features including vorticity from real flow field data or predicted flow field data in training sample Shear rate Energy dissipation ratio Pressure gradient die The velocity gradient tensor invariant Q, R, the invariant Q is used for describing the relative strong and weak relation between rotation and strain in the local flow field and reflecting the strength characteristics of the vortex structure, and the invariant R is used for describing the third-order topological characteristics of the velocity gradient tensor and reflecting the spatial organization form and evolution trend of the flow structure; A physical feature vector represented by the following formula (2) is formed: wherein x represents a spatial position coordinate and t represents a time variable; step 2, dimensionless and normalization treatment; Dimensionless and normalization of the physical feature vectors is performed, and the physical feature vectors are mapped to a uniform scale according to the formula (2): Wherein the method comprises the steps of Is a normalization operator; representing the original value of the ith physical quantity at the space position x and the time t, wherein i is a physical quantity index and is used for distinguishing different types of physical characteristic quantities; step 3, constructing a multi-physical-quantity fusion weighting function; Constructing a fusion weighting function shown in a formula (3) based on the normalized multi-physical quantity: Wherein, the Is a Sigmoid function; a weight function for each physical quantity in the step 1; is the upper and lower limits of the weight; The function corresponding to the i-th physical quantity is used for describing the influence relation of the physical quantity on the fusion weighting result; Step 4, updating the time self-adaptive weight; and recursively updating the weights in the time dimension according to the formula (4): Wherein, the Is a temporal smoothing factor; dynamic weights for end use; the weight is a spatial fusion weight; the weight of the last time step; Step 5, constructing a dynamic weighting loss function; during the network training process, the dynamic weights are used for deep learning flow field predictive loss according to equation (5): Wherein, the The flow field is predicted for the network, Is a real flow field; Step 6, model training and prediction; the dynamic weighting loss is embedded into the training process of the convolution network, the prediction capability of the high-vorticity coupling region is enhanced by the weight, and the method is used for fast prediction of the time sequence flow field.
  2. 2. The method for constructing the dynamic weighted loss of multiple physical quantities for deep learning flow field prediction according to claim 1, wherein the step 1 is specifically: Vorticity amount The calculation method is shown as a formula (6): where u, v are the components of velocity in the x and y directions, respectively; Shear rate of The calculation is performed as shown in formula (7): in the formula, And Components in different directions in the velocity gradient tensor are respectively represented; is a velocity gradient vector component, where u 1 =u,u 2 =v; To represent spatial coordinates; Is the strain rate tensor; Energy dissipation ratio The calculation is performed as shown in formula (8): in the formula, Is kinematic viscosity; Pressure gradient die The calculation is performed as shown in formula (9): in the formula, Is the pressure; the velocity gradient tensor invariant is calculated as shown in equation (10): wherein i, j, k are sum indexes; this results in a multi-physical quantity set represented by the following formula (11): (11)。
  3. 3. The method for constructing the dynamic weighted loss of multiple physical quantities for deep learning flow field prediction according to claim 2, wherein the kinematic viscosity is 。
  4. 4. The method for constructing the dynamic weighted loss of multiple physical quantities for deep learning flow field prediction according to claim 3, wherein the normalization process of the step 2 is specifically as follows: the normalized form is shown in formula (12): in the formula, 、 Respectively physical quantity 1% And 99% quantile values of (c).
  5. 5. The method for constructing the dynamic weighted loss of multiple physical quantities for deep learning flow field prediction according to claim 1, wherein the convolution-like network is 3D U-Net or V-Net.
  6. 6. An electronic device comprising a processor and a memory, the memory for storing a computer program, the processor for executing the computer program stored by the memory to cause the electronic device to perform the method of any one of claims 1 to 5.
  7. 7. A computer readable storage medium, on which a computer program is stored, which computer program, when being executed by a processor, implements the method according to any one of claims 1 to 5.
  8. 8. A chip comprising a processor for calling and running a computer program from a memory, causing a device on which the chip is mounted to perform the method of any one of claims 1 to 5.
  9. 9. A computer program product comprising a computer storage medium storing a computer program comprising instructions executable by at least one processor, the instructions when executed by the at least one processor implementing the method of any one of claims 1 to 5.

Description

Multi-physical-quantity dynamic weighting loss construction method for deep learning flow field prediction Technical Field The invention belongs to the technical field of hydrodynamics, and particularly relates to a method for constructing multiple physical quantity dynamic weighting loss of deep learning flow field prediction. Background Unsteady flow has important significance in aerospace, ocean engineering, wind energy utilization and automobile engineering. In such flows, complex structures such as vortex shedding, shear enhancement, small scale dissipation enhancement, etc., evolve over time. Although the traditional Computational Fluid Dynamics (CFD) method can obtain accurate flow field information, the calculated amount of the CFD method is rapidly increased along with the grid scale and the time step number, and engineering requirements are often difficult to meet in real-time prediction, rapid design optimization and multi-parameter working condition traversal. Flow field prediction methods based on deep learning gradually become research hotspots due to the efficient nonlinear expression capability. However, the existing methods generally employ a uniform, fixed error weighting scheme, and the assumption of consistency for different regions in the flow field is severely inconsistent with the actual flow characteristics. In areas of intense flow structures such as high vortex areas, shear layers, energy dissipation enhancement areas, and the like, the prediction error is significantly higher than in stationary areas, and the error can propagate cumulatively in time along the vortex street structure. The existing weighting strategy is used for constructing static space weighting based on a single physical quantity, cannot be automatically updated along with the evolution of a flow field, and is difficult to capture the comprehensive influence of multiple types of physical characteristics at the same time. Therefore, a dynamic weighting method which can be used for fusing multiple physical features and adaptively updating in the time dimension is urgently needed, so that the expression capacity of the deep learning model in a local complex region is improved, and the overall stability and physical consistency of long-term flow field prediction are improved. Disclosure of Invention The invention provides a multi-physical-quantity dynamic weighting loss construction method for deep learning flow field prediction, which aims to overcome the defects of the prior art, extracts various physical-quantity characteristics such as vortex quantity, shear rate, energy dissipation rate and the like based on flow field data obtained by real or numerical simulation, forms uniform-scale physical characteristic input through dimensionless treatment, constructs a local weighting function according to a multi-physical-quantity fusion relation, realizes dynamic updating of a weighting coefficient in a time recurrence mode, introduces the obtained dynamic weight into a loss function of a deep learning model, and enables the model to carry out key constraint on error sensitive areas such as vortex shedding areas, shear layers and the like in the training process, thereby improving the local precision and long-term time sequence stability of unsteady flow field prediction. The method has clear structure, can be matched with various deep learning flow field prediction networks to be used, and has good universality and engineering applicability. The technical scheme adopted for solving the technical problems is as follows: Step 1, calculating multiple physical quantities; Extracting multiple physical quantity features including vorticity from real flow field data or predicted flow field data in training sample Shear rateEnergy dissipation ratioPressure gradient dieThe velocity gradient tensor invariant Q, R, the invariant Q is used for describing the relative strong and weak relation between rotation and strain in the local flow field and reflecting the strength characteristics of the vortex structure, and the invariant R is used for describing the third-order topological characteristics of the velocity gradient tensor and reflecting the spatial organization form and evolution trend of the flow structure; A physical feature vector represented by the following formula (2) is formed: wherein x represents a spatial position coordinate and t represents a time variable; step 2, dimensionless and normalization treatment; Dimensionless and normalization of the physical feature vectors is performed, and the physical feature vectors are mapped to a uniform scale according to the formula (2): Wherein the method comprises the steps of Is a normalization operator; representing the original value of the ith physical quantity at the space position x and the time t, wherein i is a physical quantity index and is used for distinguishing different types of physical characteristic quantities; step 3, constructing a multi-physical-quantity fu