| About this Abstract |
| Meeting |
MS&T24: Materials Science & Technology
|
| Symposium
|
Integrated Computational Materials Engineering for Physics-Based Machine Learning Models
|
| Presentation Title |
Utilizing Convex Neural Networks to Predict the Yield Surfaces of Polycrystalline Samples with Complex Crystallographic Textures |
| Author(s) |
Matt Kasemer, Lloyd van Wees, Mark Obstalecki, Paul Shade |
| On-Site Speaker (Planned) |
Matt Kasemer |
| Abstract Scope |
Classical yield functions and more modern analytical functions are unable to capture the generalized yield behavior of metals with complex crystallographic textures, which tend to exhibit pronounced anisotropy. As a result, engineering designs often require generous factors of safety. While crystal plasticity finite element (CPFE) modeling is able to capture this anisotropic yield behavior, it is computationally expensive. In this talk, we present the use of partially input convex neural networks (pICNN) to develop a constitutive model trained by crystal plasticity finite element simulations. In particular, we will describe an adaptive algorithm for the generation of training data as informed by the shape of yield surfaces to reduce the time for both the generation of training data as well as pICNN training. pICNN predictions will be compared against baseline large-data CPFE results, which serve as ground-truth. We will discuss errors in both training and test datasets, limitations, and future extensibility. |