Ensemble Predictions of Material Behavior for ICMSE

摘要

Crystal plasticity studies are dominated by a range of deterministic models of varying complexity ranging from spatially resolved physics-based models incorporating known physical laws to less complex models based on empirically observed dynamics. The increased use of such models to study material behavior requires a better understanding of associated uncertainty from various sources, and the important task of selecting the most appropriate model under uncertainty. It is well known that model complexity is not necessarily indicative of predictive power, as model performance is also highly dependent on the information provided by available calibration data. The information conveyed by available experimental data is often not sufficient to confidently calibrate complex physics-based models, so their choice over a simpler model may introduce bias in the estimation under a data-poor scenario. The objective of this work is twofold. The first stage consists of uncertainty quantification (UQ) via a Bayesian inferential framework within the problem of model comparison. The goal is to understand when increased predictive fidelity justifies implementation of a more complex physics-based model that is estimated from limited data over a simpler phenomenological model. The second stage is the development of an ensemble prediction system (EPS), which combines predictions from multiple models into an average prediction in lieu of choosing the single best model. This technique has been successfully applied in other disciplines to enhance the predictive ability from a single selected model. This technique assigns weights to predictions from different models according to the expected fidelity of model performance. The two phases of model selection and model combination are complementary in this setting. The results from the model selection framework provide important information for constructing a realistic EPS, by informing weight assignment. In summary, taking a statistical perspective on crystal plasticity modeling will aid in quantifying the degree of confidence in model predictions, and inform model selection while enhancing prediction.