Reese Jones is currently a staff scientist at Sandia National Laboratories in Livermore, CA. He is engaged in materials science and computational physics research with scales ranging from atomic/molecular to the continuum. He has made contributions to multi-scale methods, electrochemical and thermal transport, atomic-level fracture, and contact. Recently he has been developing and applying machine learning methods to provide constitutive models for
complex materials, quantify material uncertainty, and interpret materials imaging for reliability analysis.

PREDICTING FAILURE IN POROUS METALS USING CONVOLUTIONAL NEURAL NETWORKS

Predicting whether defects are critical or not is a high-value task in medicine, materials engineering, and other fields. Tools that augment expert opinion are needed in the current era of high resolution imaging that can reveal an overwhelming number of defects. In particular, porosity is a persistent feature of additively manufactured materials and determines failure locations through complex mechanics that exhibit sensitivity to the initial pore locations. In the case of materials engineering expensive direct numerical simulations are available and can be used to train efficient surrogates. Neural networks, such as the one we have developed, enable more complete analysis of potential outcomes.

In this work, we develop convolutional neural networks as surrogate models for predicting failure
locations. The binary classification problem of categorizing intact/failed voxels is first regularized by recasting it as a regression problem for the continuous damage field subjected to pre-processing transformations. An apparent challenge is the damage fields display a relatively small number of voxels close to failure leading to a form of class imbalance for regression that can cause the optimizer to converge to a poor local minimum. We address this through a re-weighting of the loss function which accounts for the relative frequencies of damage values. Another challenging aspect is the high sensitivity of the outcomes to the porosity field which typically creates multiple regions of high damage competing for failure. This motivates the use of Bayesian neural networks to capture sensitivities in the prediction through uncertainty quantification. We use these uncertainties to rank the likelihood of failure of any particular cluster of porosity in a reliability analysis. Lastly, to aid transferability of the network and reduce the training burden when it is applied to new materials and processes, we are exploring transfer learning techniques.

The MICDE Fall 2022 Seminar Series is open to all.

This seminar is hosted by the Michigan Institute for Computational Discovery & Engineering (MICDE). Dr. Jones will be hosted by Prof. Krishna Garikipati, Professor of Mechanical Engineering and Mathematics and Director of MICDE.

This is an in-person event, Zoom link will only be provided upon request. This seminar will not be recorded.

Graduate Certificate in Computational Discovery and Engineering, and MICDE fellows, please use this form to record your attendance.

Questions? Email [email protected]