3D Microstructure Reconstruction via Convolutional Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for reconstructing 3D microstructures from 2D images are inefficient due to excessive energy consumption and loss of image quality, and lack accuracy as they do not utilize convolution neural networks, leading to discrepancies between generated and original shapes.
Innovation Solution
A method using a convolution neural network to generate 3D microstructures by configuring an initial 3D structure, obtaining cross-sectional images, generating feature maps, calculating a loss value, and applying it to a back-propagation algorithm to update the structure, thereby refining the 3D microstructure based on gradients from multiple directions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If many cross-sections of the new material are obtained to reconstruct a 3D structure, then the accuracy of the 3D structure is improved, but excessive energy consumption and loss of image quality occur
Solution Approach 1:
The patent applies preliminary action by using a neural network to predict and generate the 3D microstructure from a single 2D image before any physical cross-sectioning is performed. The neural network is trained beforehand on datasets of 2D images and corresponding 3D structures, enabling it to directly reconstruct 3D microstructures without requiring multiple physical sections, thereby eliminating the energy consumption associated with obtaining many cross-sections while maintaining reconstruction accuracy
2Measurement precision
If many cross-sections are obtained to reconstruct a 3D structure, then the accuracy of the 3D structure is improved, but loss of image quality occurs
Solution Approach 1:
The neural network performs preliminary reconstruction from a single high-quality 2D image before any physical sectioning that would degrade image quality. By predicting the 3D structure algorithmically rather than physically sectioning multiple times, the original image quality is preserved while still achieving accurate 3D reconstruction through the learned patterns in the neural network
3Ease of manufacture
If probability determination method is used to reconstruct 3D microstructure, then the process is simplified, but accuracy decreases due to shape discrepancy between generated and original particles
Solution Approach 1:
The patent replaces the probabilistic determination method with a deterministic neural network-based approach. The neural network, trained on actual 2D-3D image pairs, deterministically reconstructs 3D microstructures that accurately preserve particle shapes by learning the transformation patterns from 2D projections to 3D structures, eliminating the shape discrepancies that occur with probability-based methods while maintaining process simplicity through the automated neural network pipeline
Data Source
AI summary
A method of generating a 3D microstructure using a neural network includes configuring an initial 3D microstructure; obtaining a plurality of cross-sectional images by disassembling the initial 3D microstructure in at least one direction of the initial 3D microstructure; obtaining first output feature maps with respect to at least one layer that constitutes the neural network by inputting each of the cross-sectional images to the neural network; obtaining second output feature maps with respect to at least one layer by inputting a 2D original image to the neural network; generating a 3D gradient by applying a loss value to a back-propagation algorithm after calculating the loss value by comparing the first output feature maps with the second output feature maps; and generating a final 3D microstructure based on the 2D original image by reflecting the 3D gradient to the initial 3D microstructure.