A versatile strategy for hybridizing small experimental and large simulation data: A case for ceramic tape-casting process

[Display omitted] •A deep learning model is developed for efficient process modeling using a limited amount of experimental data.•The model facilitates an under 0.9% discrepancy rate with the actual process by distilling physics-based knowledge from FEM simulation.•The data augmentation based on the deep learning model generates realistic data from FEM simulation and experimental data.•The developed deep learning strategy alleviates the overfitting of learning weights, thereby achieving the high performance of inverse design. In manufacturing industry, finding optimal design parameters for targeted properties has traditionally been guided by trial and error. However, limited data availability to few hundreds sets of experimental data in typical materials processes, the machine-learning capabilities and other data-driven modeling (DDM) techniques are too far from it to be practical. In this study, we show how a versatile design strategy, tightly coupled with physics-based modeling (PBM) data, can be applied to small set of experimental data to improve the optimization of process parameters. Our strategy uses PBM to achieve augmented data that includes essential physics: in other words, the PBM data allows the inverse design model to ‘learn’ physics, indirectly. We demonstrated the accuracy of both forward-prediction and inverse-optimization have been dramatically improved with the help of PBM data, which are relatively cheap and abundant. Furthermore, we found that the inverse model with augmented data can accurately optimize process parameters, even for ones those were not considered in the simulation. Such versatile strategy can be helpful for processes/experiments for the cases where the number of collectable data is limited, which is most of the case in industries.