Variational Autoencoder for Controllable Synthetic Microstructure Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating synthetic microstructure images of materials struggle to control the generation of microstructures with specific and desired features, due to the lack of clear intuitive meanings of individual dimensions in latent representations.
Innovation Solution
A processor-implemented method using a variational autoencoder defined with a loss function, where the variational autoencoder is trained to learn latent representations of microstructure images, and specific dimensions of these representations are interpreted as physically significant features to generate synthetic microstructure images with desired features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional unsupervised deep generative models (GANs, VAEs) are used to synthesize microstructures, then the realism and statistical equivalence of generated microstructures is improved, but the controllability over specific features is lost due to unclear latent space dimensions
Solution Approach 1:
The latent space is segmented into multiple independent dimensions, where each dimension corresponds to a specific microstructure feature (e.g., grain size, phase distribution). This segmentation allows independent control of each feature by manipulating the corresponding latent dimension while keeping others fixed, resolving the contradiction between statistical equivalence and feature controllability.
Solution Approach 2:
The patent transforms the latent space representation by learning a disentangled encoding where specific parameters (latent dimensions) directly correspond to physical microstructure features. By changing values in specific latent dimensions, one can control desired features while maintaining the statistical properties learned from training data, thus achieving both reliability and ease of operation.
2Ease of operation
If traditional statistical descriptors are used for microstructure synthesis, then the interpretability of generation process is improved, but the realism and statistical equivalence of generated microstructures deteriorates
Solution Approach 1:
The patent replaces traditional mechanical/statistical descriptors with a learned deep generative model that automatically discovers meaningful latent dimensions. The system substitutes manual feature engineering with automatic feature learning, where the neural network learns to represent microstructure features in the latent space, achieving both interpretability and statistical equivalence simultaneously.
3Measurement precision
If experimental characterization is used to determine mapping between manufacturing process and material properties, then the accuracy of property prediction is improved, but the cost and difficulty of the process worsens
Solution Approach 1:
The patent creates synthetic copies of microstructure images through the trained generative model, which can be generated in silico without physical experimentation. These synthetic microstructures maintain statistical equivalence to real microstructures, allowing accurate property prediction and process optimization without the high cost and complexity of experimental characterization.
Solution Approach 2:
The patent replaces expensive, time-consuming experimental characterization with cheap, rapid computational synthesis. The synthetic microstructures serve as disposable virtual samples that can be generated instantly for testing and analysis, eliminating the need for costly physical experiments while maintaining prediction accuracy.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
The disclosure generally relates to methods and systems for generating synthetic microstructure images of a material with desired features. Conventional techniques that make use of unsupervised deep generative models has no control on the generated microstructure images with specific, desired set of features. The present disclosure generates the synthetic microstructure images of the material with desired feature, by using a variational autoencoder defined with a style loss function. In the first step, the variational autoencoder is trained to learn latent representation of microstructure image of the material. In the second step, some of the dimensions of learned latent representation is interpreted as physically significant features. In the third and last step, the latent representation required for getting the desired features is appropriately sampled based on the interpretation to generate the synthetic microstructure images of the material with desired features.