Session: 16-01-01: Government Agency Student Poster Competition

Paper Number: 149585

149585 - A Machine Learning Informed Phase Field Damage Model to Simulate Void Nucleation and Growth in Metal Microstructures. 

In mechanical, defense, aerospace, and military applications, structural materials often experience high-impact dynamic loads, such as bullet impacts and space debris collisions. Under these conditions, spallation is a common damage mechanism that involves void nucleation, growth, coalescence, and ultimately material failure. Traditional methods have struggled to capture the combined effects of all factors involved in void nucleation due to the multifaceted nature of the influencing factors. Thus, it is essential to fill this knowledge gap. This work integrates machine learning with phase field modeling into a comprehensive model that is able to predict and simulate void nucleation and growth during the incipient spall of polycrystalline metals under dynamic loads with high accuracy. 

In this work, a hybrid machine learning model is  employed to predict potential void nucleation sites in polycrystalline metal microstructures under dynamic loads. These models consider factors such as grain orientation, grain boundary energy, stress-strain data, crystal plasticity data, and the local environment around void sites. By integrating these diverse data inputs, the machine learning models can more accurately predict where void nucleation is likely to occur.

To enhance the robustness and applicability of these models, they are validated across a broader range of microstructures, including synthetic microstructures generated using Generative Adversarial Networks (GANs). This validation process ensures that the machine learning models can reliably predict void nucleation sites, enhancing their overall robustness and versatility.

Once the machine learning models accurately predict void nucleation sites in polycrystalline metals, a heatmap of potential void nucleation sites is generated. This heatmap informs a phase field model, which simulates the nucleation and growth of voids in polycrystalline materials under dynamic loading. The phase field model, informed by the machine learning predictions, operates on the principle of minimum dissipation potential and incorporates factors such as chemical potential, grain boundary energy, elasticity, and plastic deformation. This integration allows the phase field model to accurately simulate void nucleation and growth, closely aligning with experimental results.

Preliminary results indicate that the machine learning model can predict void nucleation sites with an accuracy of 71%, considering only grain boundary orientations and energy. Recognizing the need for greater accuracy, additional data such as stress-strain and crystal plasticity have been incorporated to enhance the model’s predictive accuracy. This step has been crucial as it allows for a more comprehensive understanding of the factors influencing void nucleation, leading to even higher predictive accuracy. The heatmap fields generated by the improved machine learning model have been used to inform the phase field damage model, which then simulates void nucleation and growth. This model has been validated against experimental results of void nucleation and growth in body-centered cubic (BCC) tantalum under high impact dynamic loads, demonstrating its ability to accurately simulate void nucleation and growth.

Our results demonstrate the potential of the machine learning model to predict void nucleation sites accurately. The integration of heatmap fields from the machine learning model into the phase field model shows promise in accurately simulating void evolution under high-impact conditions. Thus, this  ML-informed phase field modeling framework presents a novel and effective approach to understanding and predicting void nucleation and growth in polycrystalline metals.

Presenting Author: Abhijith Thoopul Anantharanga Iowa State University

Presenting Author Biography: PhD student at Iowa State University in the Department of Aerospace Engineering

Authors:

Abhijith Thoopul Anantharanga Iowa State University
Jackson Plummer Iowa State University
Saryu Fensin Los Alamos National Laboratory
Brandon Runnels Iowa State University