Benchmarking of Neural Network Methodologies for Piston Thermal Model Calibration

Design of internal combustion (IC) engine pistons is dependent on accurate prediction of the temperature field in the component. Experimental temperature measurements can be taken but are costly and typically limited to a few select locations. High-fidelity computer simulations can be used to predict the temperature at any number of locations within the model, but the models must be calibrated for the predictions to be accurate. The largest barrier to calibration of piston thermal models is estimating the backside boundary conditions, as there is not much literature available for these boundary conditions. Bayesian model calibration is a common choice for model calibration in literature, but little research is available applying this method to piston thermal models. Neural networks have been shown in literature to be effective for calibration of piston thermal models. In this work, Bayesian model calibration will be compared to two neural network-based calibration methodologies for piston thermal models. The models were compared for both computation time and error across three different data densities. Each data set represents an increasing density of steady-state temperature measurement locations. The results show that the error between the methods is largely consistent across the different data densities, with each model having similar error to the others at each calibration case. On the other hand, computation time highlights the advantage of the neural network methodologies over the Bayesian methodology. At the lowest data density, the Bayesian model calibration methodology had the fastest computation time but only by a few minutes. As the data density increased, the Bayesian model calibration method became hours slower than the Neural network methods, up to 4673.3% slower at the highest data density. Both neural networks-based approaches and the Bayesian model calibration methodology are effective at calibrating at low data densities but for higher data densities, the Bayesian model calibration becomes too computationally expensive.