CN-121999492-A - Microstructure evolution prediction method based on grain boundary structure constraint

CN121999492ACN 121999492 ACN121999492 ACN 121999492ACN-121999492-A

Abstract

The invention discloses a microstructure evolution prediction method based on grain boundary structure constraint, which comprises the steps of preparing a data set, extracting a grain boundary structural diagram, constructing grain boundary weighted MSE loss, constructing grain boundary consistency loss, constructing global structure similarity loss, constructing a composite loss function, constructing a time sequence prediction model iteration training and parameter optimization, and carrying out model reasoning and output. According to the invention, on the premise of not introducing additional physical parameters and not changing the main structure of the time sequence prediction model, a special grain boundary structure constraint mechanism is constructed through the synergistic effect of automatic extraction of a grain boundary structure diagram, grain boundary weighting MSE loss and grain boundary consistency loss, so that the problems of fuzzy grain boundary structure and poor boundary continuity in the existing microstructure time sequence prediction method are effectively improved, and the overall quality index and the material science semantic fidelity of the predicted image are improved.

Inventors

Zhou Miaolan
MENG WENCHAO
CHEN JIMING
ZHANG YUEFEI
ZHANG XUECHENG

Assignees

浙江大学

Dates

Publication Date: 20260508
Application Date: 20260410

Claims (9)

1. A microstructure evolution prediction method based on grain boundary structure constraints, comprising: step 1, preparing a data set; step 2, extracting a grain boundary structural diagram; step 3, constructing a grain boundary weighted MSE loss; Step 4, constructing grain boundary consistency loss; step 5, constructing a global structure similarity loss; Step 6, constructing a composite loss function; Step 7, performing iterative training and parameter optimization of a time sequence prediction model; and 8, carrying out model reasoning and outputting.
2. The method for predicting microstructure evolution based on grain boundary structural constraints according to claim 1, the method is characterized in that the step 1 comprises the following steps: The method comprises the steps of dividing each original image into a plurality of windows, dividing the row direction into a plurality of sections of row areas according to a preset proportion, dividing the column direction into a plurality of sections of column areas according to a preset proportion, adaptively distributing the sizes of the sections of row areas and the column areas according to the row height, the column width and the number of divided sections of the original image, dividing data sets of all windows obtained by dividing, distributing most of the windows into training sets, distributing the rest of the windows into test sets and verification sets, and extracting image blocks with fixed sizes by adopting preset coincidence degrees for each window.
3. The method for predicting microstructure evolution based on grain boundary structural constraints according to claim 1, the method is characterized in that the step S2 comprises the following steps: For each frame of real image X t in the training set, the verification set and the test set, respectively adopting Sobel gradient operators to calculate horizontal gradient and vertical gradient for RGB three channels, respectively calculating Sobel gradient amplitude for RGB three channels, and then taking the channel mean value as final grain boundary structure gradient amplitude G t ', wherein the calculation formula of the grain boundary structure gradient amplitude G t ' is as follows: Wherein, the A c-th channel representing a t-th frame image, Is the horizontal gradient of the c-th channel of the t-th frame image, Is the vertical gradient of the c channel of the t frame image; G t ' is mapped to the [0,1] range in a linear normalization mode, a high-value region of the grain boundary structure diagram corresponds to an orientation mutation position, a low-value region corresponds to a uniform orientation region in the crystal grain, and a normalization formula is as follows: Wherein, max represents a maximum function, min represents a minimum function, and G t is a real image grain boundary structure diagram.
4. The method for predicting microstructure evolution based on grain boundary structural constraints according to claim 1, the method is characterized in that the step 3 comprises the following steps: First, a weight map W t is calculated from a real image grain boundary structural diagram G t : Wherein λ is a grain boundary weighting coefficient, and then a weighted mean square error L wmse is calculated, with the following calculation formula: Wherein N is the total number of pixels of the image, the subscript i represents the pixel index, all pixels in the image are summed, W t,i is the weight value of the ith pixel of the t-th frame, and the weight map is used for calculating the weight value of the ith pixel of the t-th frame Calculated element by element, X t,i and The values of the ith pixel of the t-th frame of the real image and the predicted image, respectively.
5. The method for predicting microstructure evolution based on grain boundary structural constraints according to claim 1, the method is characterized in that the step 4 comprises the following steps: The first-order absolute difference between the grain boundary structure diagram of the prediction image and the grain boundary structure diagram of the real image is directly minimized, the grain boundary position, the width and the continuity of the prediction image are forced to be consistent with those of the real image, and the calculation formula of the grain boundary consistency loss L edge is as follows: where N is the total number of pixels of the image, i represents the pixel index, summing all pixel positions in the image, The method comprises the steps of extracting a t frame predicted image through a grain boundary structure, and then, constructing a grain boundary structure diagram of an i pixel, wherein G t,i is the grain boundary structure diagram of the i pixel after extracting a t frame real image through the grain boundary structure.
6. The method for predicting microstructure evolution based on grain boundary structural constraints according to claim 1, the method is characterized in that the step 5 comprises the following steps: the core calculation formula of the structural similarity loss SSIM (x, y) in the local window is set as follows: Wherein x, y are the local sub-blocks of the real image and the predicted image in the local window, respectively, mu x ,μ y is the mean value of the predicted image and the real image in the local window, , As a local variance of the values of the local variance, As local covariance, C 1 ,C 2 is a very small stability constant; The global structural similarity penalty L ssim is set to: Where M is the number of local windows in the image, x j ,y j is the local sub-blocks of the real image and the predicted image in the jth local window, respectively, and SSIM (x j ,y j ) is the SSIM value in the jth local window.
7. The method for predicting microstructure evolution based on grain boundary structural constraints according to claim 1, the method is characterized in that the step 6 comprises the following steps: The composite loss function L is set as: The method comprises the steps of obtaining a time sequence prediction model, wherein alpha and beta are weight coefficients of a global structural similarity loss L ssim and a grain boundary consistency loss L edge respectively, L wmse represents grain boundary weighted MSE loss, and training and optimizing the time sequence prediction model through a composite loss function.
8. The method for predicting microstructure evolution based on grain boundary structural constraints according to claim 1, the method is characterized in that the step 7 comprises the following steps: Step 7.1, inputting a plurality of frame images before the sequence into a time sequence prediction model, carrying out forward propagation through a time-space memory unit and a recursion structure of the time sequence prediction model, and generating a plurality of frame prediction images after each frame Simultaneously for the predicted image Extracting the predicted grain boundary structural diagram in the same way as the step 2 ; Step 7.2, back propagation and end-to-end parameter optimization are carried out on the time sequence prediction model by utilizing the calculated composite loss function L; and 7.3, executing the step 7.1 and the step 7.2 through multiple iterations until the time sequence prediction model converges and then preserving parameters.
9. The method for predicting microstructure evolution based on grain boundary structural constraints according to claim 1, wherein the step 8 comprises the following steps: After training, the optimized time sequence prediction model parameters are directly used in the reasoning stage, a plurality of frame images before a new sequence are input, the time sequence prediction model recursively generates future frame prediction output, and finally the future frame image sequence which clearly maintains the grain boundary structure and the orientation distribution is obtained, so that the accurate prediction of microstructure evolution is realized.

Description

Microstructure evolution prediction method based on grain boundary structure constraint Technical Field The invention relates to the field of material science, in particular to a microstructure evolution prediction method based on grain boundary structure constraint. Background With the rapid development of deep learning technology, the time sequence prediction model is increasingly widely applied in the field of material science, and particularly, a time sequence prediction method based on a recurrent neural network has become an important tool for simulating the microstructure evolution of materials. Wherein PredRNN effectively captures dynamic changes of image sequences through a space-time memory structure, and is commonly used for predicting grain growth, phase change and sliding band migration processes of materials such as alloy or carbon steel. The model is excellent in processing small sample Electron Back Scattering Diffraction (EBSD) or Scanning Electron Microscope (SEM) image sequences, and usually adopts a pixel level error loss function (such as mean square error) to optimize the difference between a predicted frame and a real frame, and is practically applied to material microstructure prediction, for example, microstructure space-time evolution prediction is directly realized based on PredRNN. In EBSD image sequence prediction, the grain boundary is used as a direct physical characterization of the grain orientation mutation, and is a core feature for evaluating mechanical properties such as material strength, toughness and the like. While the prior art may utilize Sobel gradient operators to extract image edge information to assist feature learning, it generally relies only on global pixel errors or structural similarity loss (e.g., SSIM) to optimize image quality. In recent years, some improvement methods introduce structural similarity constraints or post-processing filtering to alleviate the boundary blurring problem. In the prior art, although the PredRNN isochronous prediction model achieves better effect in microstructure time sequence prediction, the following disadvantages still exist: The prior art relies primarily on pixel level error loss functions (e.g., mean square error) to optimize the model. Because the pixel ratio of the high-frequency region such as the grain boundary and the slipping zone in the EBSD image is relatively small, the error contribution of the pixel ratio is low in the total loss, and the attention of the model to the key physical feature of the grain boundary orientation mutation in the training process is insufficient. The direct problems generated by the method are that the grain boundary in the predicted image is easy to blur, the boundary is widened, the slip zone is interrupted or the continuity is lost, and the accurate simulation and mechanical property evaluation of the microscopic evolution process of the material are seriously affected. On the other hand, although the prior art can introduce structural similarity loss or extract edge information through the assistance of a Sobel gradient operator, the measures are still mainly optimized on the whole image quality, and the grain boundary structure is not explicitly integrated into a loss function as an independent training constraint target. Therefore, the grain boundary structure of the predicted image cannot be forced to be consistent with the real image in the training stage, and the predicted result is acceptable in the overall quantitative indexes (such as PSNR and SSIM), but has obvious defects in the aspects of grain boundary definition and material semantic preservation, and is difficult to meet the requirement of the microstructure simulation of the precise material. The above-mentioned drawbacks stem from the explicit modeling of the grain boundary physical characteristics by the loss function, which makes it difficult to achieve a clear and continuous preservation of the grain boundary structure in the prior art. Disclosure of Invention The invention aims to provide a microstructure evolution prediction method based on grain boundary structure constraint, so as to solve the problems in the background technology. In order to achieve the above purpose, the present invention provides the following technical solutions: a microstructure evolution prediction method based on grain boundary structural constraints, comprising: step 1, preparing a data set; step 2, extracting a grain boundary structural diagram; step 3, constructing a grain boundary weighted MSE loss; Step 4, constructing grain boundary consistency loss; step 5, constructing a global structure similarity loss; Step 6, constructing a composite loss function; Step 7, performing iterative training and parameter optimization of a time sequence prediction model; and 8, carrying out model reasoning and outputting. Further, the step 1 includes: The method comprises the steps of dividing each original image into a plurality of windows, div