Session: Rising Stars of Mechanical Engineering Celebration & Showcase

Paper Number: 148670

148670 - A Long Short Term Memory Network-Based Surrogate Model for Predicting Ductile Failure 

Data-driven surrogate modeling using machine learning and deep learning emulators is being extensively explored in various domains to partially replace or supplement high-fidelity simulations. The reduced-order surrogate models are developed using the data extracted from high-fidelity simulations such as finite element analysis, finite discrete element method, etc. The reduced-order models are traditionally built using machine learning emulators to perform relatively less complex tasks such as regression, classification, and dimensional reduction. However, with the availability of big data and an exponential increase in the computational power, artificial deep neural networks with millions of trainable parameters are being increasingly used to perform complex tasks with history dependencies which include playing Go, autonomous driving, and simulating the response of highly non-linear dynamic systems. However, very limited progress was made in ductile fracture mechanics.

 Ductile fracture is driven by plastic deformation and depends on the stress-state history of the material. Simulating history-dependent phenomena such as fracture of ductile metals is a complex phenomenon to learn for the reduced-order surrogate models. However, with the developments in the artificial neural networks such as, recurrent neural networks (RNN)), it may be possible to capture history-dependent fracture phenomena. In particular, long short-term memory (LSTM) networks have been successfully used to predict fracture in brittle materials. The present study aims to configure and train a geometry-specific reduced-order surrogate model based on LSTM and bidirectional LSTM (Bi-LSTM) to predict the overall load-displacement behavior of cylindrically notched test specimens made of ASTM A992 structural steel.

After the configuration of its architecture, the LSTM based surrogate model was trained using high-fidelity finite element analyses of 42 notched cylindrical test specimens that are capable of producing various high stress triaxialities. The micromechanics-based Gurson-Tvergaard-Needleman (GTN) model was used as a constitutive model in the finite element analyses. The load-displacement data from finite element analysis of each of the test specimens was discretized into 500 loading steps, with the initial n load steps serving as input to the trained network. These initial steps were utilized to predict the (n + 1) load step. Subsequently, the model iteratively predicted p load steps based on prior predictions. The trained model was validated using 10 axisymmetric notched test specimens subjected to uniaxial tension with significantly different dimensions than that of the specimens in the training dataset. The accuracy of the model predictions, computational time savings when compared to the traditional approach, and limitations associated with the proposed surrogate model, especially the geometrical restrictions will be illustrated through examples.

Presenting Author: Ravi Yellavajjala Arizona State University

Presenting Author Biography: Dr. Ravi Kiran Yellavajjala is an associate professor in the School of Sustainable Engineering and Built Environment at ASU and is the Principal Investigator (P.I.) for the DAMS lab. He received his Ph.D. in structural engineering from the University of Notre Dame in 2015. Among other awards, he received the 2022 AASHTO High-value research award, the 2021 NSF CAREER award, the 2014 O.H. Ammann structural engineering fellowship, and stood runner-up in the 2014 Computational Mechanics poster competition. Dr. Ravi’s research interests lie in the broad areas of damage mechanics and Artificial Intelligence with applications in civil and transportation infrastructure. He co-authored 55+ journal articles, presented his research work in over 60 international conference venues, and delivered 17 invited talks. He currently serves as a P.I. on several externally funded research projects. He is an ad hoc reviewer for over 40+ journals and has served as an invited reviewer for several funding agencies. He is an associate editor for the ASCE Journal of Materials in Civil Engineering, an editorial board member for the Journal of Infrastructure Preservation and Resilience, and served as Guest Editor for a couple of special issues devoted to steel structures research (Advances in Structural Steel Research, Sustainable, and Resilient Steel Structures). He is a member of the American Institute of Steel Construction (AISC) and the American Society of Civil Engineers (ASCE). He is a licensed Professional Engineer in Minnesota (License # 57807).

Authors:

Surajit Dey Arizona State University
Ravi Yellavajjala Arizona State University