Search

CN-121978145-A - Automatic detection method and system for defects of 4D-STEM material

CN121978145ACN 121978145 ACN121978145 ACN 121978145ACN-121978145-A

Abstract

The invention relates to the technical field of defect detection, in particular to a method and a system for automatically detecting defects of a 4D-STEM material, wherein the method comprises the steps of obtaining 4D-STEM data and carrying out structural processing to generate low-dimensional characteristics input by a student network and privilege mode data input by a teacher network; an asymmetric teacher-student network is constructed, the teacher network processes privilege modes to obtain physical semantic features, the student network processes low-dimensional features, the student network is pre-trained by a composite loss function formed by DINO distillation and the like, teacher parameters are updated through EMA, the teacher network is removed, then a decoder is connected for fine adjustment, and a network is deployed to input data to be detected and output results. According to the invention, by means of cross-modal distillation, students learn physical priori through a network, the detection precision is improved, the automatic detection of defects such as dislocation is realized, and the efficiency and the robustness are both considered.

Inventors

  • XIE YUJUN
  • SHI PIN
  • ZENG XIAOQIN
  • MAO ZIAN
  • ZHANG HAORAN
  • CHU SHUFEN
  • Lin Shuanglan

Assignees

  • 上海交通大学

Dates

Publication Date
20260505
Application Date
20251224

Claims (8)

  1. 1. An automated detection method for defects of a 4D-STEM material, which is characterized by comprising the following steps: s1, acquiring 4D-STEM data obtained by scanning a material sample, and carrying out structuring treatment to construct a four-dimensional data cube containing information related to scanning positions and diffraction patterns, and respectively generating a low-dimensional potential feature vector input by a student network and privilege mode data input by a teacher network based on the four-dimensional data cube; S2, an asymmetric teacher-student neural network is constructed, wherein the teacher network processes the privilege mode data to obtain physical semantic features, and the student network processes the low-dimensional potential feature vectors to obtain feature representations; Step S3, training a student network by using the low-dimensional potential feature vector and the privilege mode data generated in the step S1 and synchronously updating teacher network parameters through momentum updating in the training process, wherein the teacher network does not back propagate, so that the student network learns physical priori knowledge of the teacher network and aligns feature space; s4, removing a teacher network, accessing a decoder and a segmentation head after a trained student network encoder, and performing supervision and fine adjustment on the student network by adopting marking data; And S5, deploying the finely-adjusted student network, inputting a low-dimensional potential feature vector obtained by preprocessing 4D-STEM data to be detected, and outputting a material defect detection result.
  2. 2. The automated 4D-STEM material defect detection method of claim 1, wherein in step S1, the low-dimensional latent feature vector is obtained by encoding the original diffraction pattern using a pre-trained and frozen-weight variance self-encoder.
  3. 3. The automated 4D-STEM material defect detection method according to claim 1, wherein in step S1, the generating process of the privileged mode data input by the teacher network includes: generating a virtual dark field image under a simulated double-beam condition by calculating and processing 4D-STEM data; Index analysis is carried out on the diffraction patterns, the strongest first N Bragg diffraction peaks are extracted to construct a Bragg point set, and each point in the Bragg point set comprises 6 physical descriptors, namely reciprocal space coordinates qx, qy, miller indexes h, k, l and intensity I.
  4. 4. The automated 4D-STEM material defect detection method of claim 3, wherein the teacher network processing the privileged mode data to obtain physical semantic features comprises: flattening the extracted Bragg point set tensor, and projecting the flattened Bragg point set tensor to a model dimension through a linear layer to obtain point set characteristics; projecting the virtual dark field image to the dimension of the model through the convolution layer to obtain image characteristics; the point set features and the image features are spliced, fused through a linear layer and GeLU activation functions, and input into a teacher transducer encoder to obtain feature representations containing rich physical semantics.
  5. 5. The automated 4D-STEM material defect detection method of claim 1, wherein the composite loss function comprises distillation loss, differential entropy loss, and mask reconstruction loss; The distillation loss calculates cross entropy loss between a teacher network and CLS tokens output by a student network, and the characteristic distribution of the student network and the teacher network is aligned; the differential entropy loss calculates differential entropy of the student network output characteristics and maximizes the differential entropy, and the differential entropy is used for maximizing the distance between different sample characteristics in the same batch and preventing the characteristics from collapsing; The mask reconstruction penalty is used to randomly mask the input to the student network, requiring the network to reconstruct the occluded diffraction pattern features.
  6. 6. The automated 4D-STEM material defect detection method of claim 1, wherein in step S3, the momentum update updates teacher network parameters by exponential moving average.
  7. 7. The automated 4D-STEM material defect detection method according to claim 1, wherein in step S5, the output material defect detection result includes a dislocation segmentation result and a Burgers vector type.
  8. 8. A 4D-STEM material defect automated inspection system, comprising: The data acquisition and preprocessing module is used for acquiring 4D-STEM data obtained by scanning a material sample and carrying out structuring processing, constructing a four-dimensional data cube containing information related to scanning positions and diffraction patterns, and respectively generating a low-dimensional potential feature vector input by a student network and privilege mode data input by a teacher network based on the four-dimensional data cube; the network construction module is used for constructing an asymmetric teacher-student neural network, wherein the teacher network processes the privilege mode data to obtain physical semantic features, and the student network processes the low-dimensional potential feature vectors to obtain feature representations; The network training module trains the student network by using the low-dimensional potential feature vector and the privilege mode data generated in the step S1 and synchronously updates the teacher network parameters through momentum updating in the training process, and the teacher network does not back propagate, so that the student network learns the physical priori knowledge of the teacher network and aligns the feature space; The network fine tuning module removes a teacher network, and accesses a decoder and a segmentation head after a trained student network encoder, and adopts marking data to conduct supervision fine tuning on the student network; the deployment and detection module is used for deploying the finely-tuned student network, inputting the low-dimensional potential feature vector obtained by preprocessing the 4D-STEM data to be detected, and outputting a material defect detection result.

Description

Automatic detection method and system for defects of 4D-STEM material Technical Field The invention relates to the technical field of defect detection, in particular to an automatic defect detection method and system for a 4D-STEM material. Background The first is based on the dual-beam imaging and manual analysis of the traditional electron microscope TEM, manually selects the specific reciprocal lattice vector g through TEM/STEM equipment, generates an image under the dual-beam imaging condition, then manually judges the existence, type and Burgers vector of dislocation according to the g ∙ b=0 invisible criterion, and compares the imaging contrast change under a plurality of g vectors, the scheme relies on expert experience to carry out multiple experimental operation and image comparison, and can only analyze under the specific g vector, thus having the defects of low efficiency, strong subjectivity, needing profound crystallography background, long time consumption for screening g vector and imaging condition, experience and data information loss are depended on the result precision, the diffraction information of other angles is discarded after the dual-beam condition is selected, and ambiguity is easily generated in complex or high-density defect areas due to information deficiency; The second is strain analysis based on traditional geometric phase analysis GPA or central moment CoM, 4D-STEM or HRTEM data is used, a strain field and a displacement field are obtained through Fourier transformation or diffraction spot position deviation calculation, strain field distortion characteristics are analyzed to indirectly position dislocation cores, the scheme belongs to pure physical calculation, low-dimensional physical quantities are directly extracted from the data and then positioned through post-processing, the defects of insufficient precision and robustness, sensitivity to noise, sample thickness variation and local inclination, pseudo strain or artifacts are easily generated near the dislocation cores, long-range tracking capability is lacked, only the local strain field is concerned, continuous dislocation lines or stacking faults across the field are difficult to identify and track, segmentation results are easily broken and discontinuous, calculation resources are wasted, the dimension of the 4D-STEM data is high, and a large amount of hidden nonlinear defect characteristic information can be ignored by using a traditional physical algorithm. Disclosure of Invention In view of the above, the invention aims to provide an automated detection method and system for defects of a 4D-STEM material, which are used for solving the problems that in the prior art, crystal defect analysis depends on manual experience, is low in efficiency and is greatly influenced by human factors, and a traditional physical algorithm is difficult to process high-dimensional data and has poor noise resistance. Based on the above purpose, the invention provides an automatic detection method for defects of a 4D-STEM material, which comprises the following steps: s1, acquiring 4D-STEM data obtained by scanning a material sample, and carrying out structuring treatment to construct a four-dimensional data cube containing information related to scanning positions and diffraction patterns, and respectively generating a low-dimensional potential feature vector input by a student network and privilege mode data input by a teacher network based on the four-dimensional data cube; S2, an asymmetric teacher-student neural network is constructed, wherein the teacher network processes the privilege mode data to obtain physical semantic features, and the student network processes the low-dimensional potential feature vectors to obtain feature representations; Step S3, training a student network by using the low-dimensional potential feature vector and the privilege mode data generated in the step S1 and synchronously updating teacher network parameters through momentum updating in the training process, wherein the teacher network does not back propagate, so that the student network learns physical priori knowledge of the teacher network and aligns feature space; s4, removing a teacher network, accessing a decoder and a segmentation head after a trained student network encoder, and performing supervision and fine adjustment on the student network by adopting marking data; And S5, deploying the finely-adjusted student network, inputting a low-dimensional potential feature vector obtained by preprocessing 4D-STEM data to be detected, and outputting a material defect detection result. Preferably, in step S1, the low-dimensional latent feature vector is obtained by encoding the original diffraction pattern using a pre-trained and frozen-weight variance self-encoder. Preferably, in step S1, the generating process of the privilege mode data input by the teacher network includes: generating a virtual dark field image under a simulated double-beam co