Open Access
Subscription or Fee Access
Maximizing Model Predictive Performance using Minimal Experimental Data with Multiscale Designer
Abstract
Successfully modeling composite material behavior requires good quality experimental data to train models on, but the expense and time required to collect such data is a major bottleneck in practice. We combine Multiscale Designer’s reduced order homogenization technology with a Bayesian framework for parameter inference to fit composite models with just a handful of unnotched tension and compression tests on unidirectional and cross-ply laminates. We demonstrate the predictive ability of the fitted model by comparing model predictions to experimental results that were held out from the fitting procedure. The held-out data consists of several more unnotched and open hole tension and compression experiments on a wide variety of laminates used in practice. This framework furnishes probabilistic predictions of behavior, and accounts for the uncertainty due to natural variability in composite behavior in addition to uncertainty arising from finite replications of the experiments used for model training. Design is an iterative process which cannot begin without some initial estimates of material strength to work with. Our framework allows the design process to begin with computational estimates of material allowables based on a small number of experiments. This can help OEMs begin the design process and test many alternate materials before committing to the time and expense of full material qualification by the traditional experimental procedures.
DOI
10.12783/asc34/31380
10.12783/asc34/31380
Full Text:
PDFRefbacks
- There are currently no refbacks.