Abstract

Today, process simulation is commonly used for designing and troubleshooting manufacturing processes of composite parts. Physics-based process simulations include a complex set of mathematical models that are solved using numerical methods such as finite elements (FE). These models typically include a large set of parameters and require multiple boundary and initial conditions, many of which are uncertain to some degree. Depending on the amount of uncertainty in the model parameters, boundary and initial conditions, there can be a wide range of possible simulation outcomes. Simulation results can be validated against experimental data, but experimental data carry their own uncertainty (e.g. measurement precision, and calibration error to name a few) and are not always available at the location(s) of interest. In addition, standard approaches to reconciling discrepancies between simulation results and experimental data can be a tedious process of trial and error: making small adjustments to simulation inputs until the outputs are consistent with the observed data. This work demonstrates how a Bayesian framework for data assimilation can be used to systematically update model predictions based on experimental data to enhance prediction accuracy and reduce uncertainty. The updated predictions represent optimal pooling between the simulated and measured data streams as they are combined in proportion to their respective uncertainties. A composite laminate with thermocouples (TC?s) embedded through its center was cured in an autoclave, and the TC data were compared with the results of a probabilistic 1D thermochemical process simulation for the same laminate. Gaussian process regression (GPR) was used to combine model predictions (prior) and TC data (observations) to generate new predictions (posterior) with enhanced accuracy and reduced uncertainty. It is shown that data assimilation can significantly improve simulation results inside the part even in cases where prior model prediction uncertainty is high and TC data are only available at remote points outside the part.