Session: 16-01-01: Poster Session: NSF-Funded Research (Grad & Undergrad)

Paper Number: 99314

99314 - Compact Representation and Secure Sharing of Metal Microstructure Data 

Neural networks and other machine learning algorithms are now commonly employed to characterize metal microstructure by taking advantage of the large volumes of microstructural data generated through modern SEMs and optical microscopes. However, the use of large volumes of high-resolution microstructural images poses challenges for the storage, and transfer of data. Also, the regions that should be given more focus during the microscopy stage are not known. Hence, the goal of this study is to obtain a securely shareable compact representation of the metal microstructure and to identify the pixel regions which are important for compact representation and secure sharing.

A convolutional autoencoder (CAE) that constructs a compact representation through the reconstruction process is employed in this study. A CAE with 50,353 network parameters was configured and tested for its generalizing ability. The configured CAE consists of four alternating convolution and max-pooling layers for the encoder and four alternating layers of upsampling and convolution layers for the decoder. A standard kernel filter size of 3 × 3 with a stride of 1 is used for all the convolution operations. A window size of 2 ×2 with stride 1 is used for max-pooling which results in the reduction of the input size by 2 along both dimensions.  The feature maps resulting from convolutions are activated through ‘ReLU’ for all the intermediate layers. For the final convolution layer, the ‘Linear’ activation function is used. The CAE is trained with the microstructural image input acquired from three grades of dual-phase structural steels: A36, A572, and A992. The designed architecture of CAE encodes an image input of size 256×256×1 to latent features of size 16×16×8 and achieves a high image compression ratio of 32 without loss of important information. The significance of this result is that input images are reconstructed from just 3.5% amount of original data without losing any essential information.

The important regions of the metal microstructure for reconstruction are identified through an in-house developed model agnostic sensitivity analysis. The sensitivity analysis is implemented by a complex derivative approximation approach. The implementation involves the construction of a complex neural network with convolution, pooling, and upsampling layers which can handle complex-valued data. Finally, saliency maps that show the pixel relevance are generated for three grades of dual-phase structural steels. The saliency maps consistently indicate pearlite regions (secondary phase region) and grain boundaries to be critical for microstructure image reconstruction. In addition, since the decoder part of the trained CAE is necessary to reconstruct the original input from the compact representation, the decoder can be used as the key for secure sharing of the microstructural data.  The sensitivity analysis framework introduced in the study can be extended to microstructural characterization and can also be applied to other materials.

Presenting Author: Dharanidharan Arumugam North Dakota State University

Presenting Author Biography: Dharanidharan Arumugam is currently a graduate research student in the Department of Civil, Construction & Environmental Engineering, North Dakota State University. He works in the Damage in Materials and Structures lab with Dr. Ravikiran. His research interests include applications of machine learning and deep learning algorithms in mechanics, explainable AI, and structural optimization. He received his Master’s degree in structural engineering from the Indian Institute of Technology, Madras, and his Bachelor’s degree in civil engineering from Anna University, Chennai.

Authors:

Dharanidharan Arumugam North Dakota State University
Ravi Kiran North Dakota State University