Sandia National Labs FY20 LDRD Annual Report

FY20 ANNUAL REPORT

The Sandia Fracture Challenge morphed from a materials scientists and engineers’ competition into a collaborative community of researchers refining their techniques for engineering reliable structures made from a variety of materials. (Photo by Randy Montoya)

A new paradigm for failure prediction using 4D materials science and deep learning. Accurate and efficient performance prediction in weapon components is crucial to maintaining a safe and secure nuclear deterrent. Even with rich characterization of the existing microstructure and defect population, simulating the behavior of these parts remains challenging and computationally intensive. This LDRD utilized a 3D convolutional neural network (CNN) to predict mechanical performance based on experimentally measured initial pore structures in additively manufactured tensile samples. High- fidelity simulation results using the Sierra/Solid Mechanics code suite served as training data for the CNN, which classified several metrics of part performance five orders of magnitude faster than traditional finite element analysis. The trained network also remained predictive when the sample geometry and loading condition were varied from what was used in training data. This work is now being extended in a new LDRD to predict failure in conventionally processed metal components using cutting-edge experimental characterization and multiscale modeling. (PI: Kyle Johnson)

(Left) Sample meshes of porous additively manufactured square and cylindrical tensile specimens were generated based on computed tomography measurements. The simulated force- displacement histories of 1,000 microstructure samples (lower middle) were used to train a deep-learning algorithm (upper middle) to classify samples based on peak load predictions with 85% accuracy. The confusion matrix (right) shows predicted labels and true labels for a reserved test set of tensile samples.

53

LABORATORY DIRECTED RESEARCH & DEVELOPMENT