|About this Abstract
||7th World Congress on Integrated Computational Materials Engineering (ICME 2023)
||Pushing the Limits of Deep Learning for Synthetic Image Generation of Titanium Alloy Microstructures in Limited Data Regime
||Gowtham Nimmal Haribabu, Jeyapriya T J, Chiranjib Bhattacharya, Bikramjit Basu
|On-Site Speaker (Planned)
||Gowtham Nimmal Haribabu
There is a great interest in material science to predict the microstructure of the product that underwent specific processing conditions. Usually, such information is obtained from trial-and-error experiments. Machine Learning can generate the computationally expensive simulation results almost instantly, without knowing the governing laws. Deep learning algorithms, especially generative adversarial networks (GANs), have demonstrated outstanding performances in synthesizing highly realistic images.
Deep neural networks are hungry for data and need thousands to millions of data for training. In this work, the StyleGANv2-ADA model was explored to generate synthetic microstructural images using few hundreds to thousands of images. Quantitative metrics like Frechet Inception Distance were used to assess the performance of the model. Quantitative Morphometric Analysis (QMA) was also performed to compare the distribution of microstructural features of real and synthetic images. StyleGAN like models will be crucial in establishing structure-process linkage in limited datasets (<1500 images) typical of metallurgical studies.
||Planned: Other (describe below)