Accurate computational fluid dynamics (CFD) simulations of the internal flow and spray characteristics in automotive injectors are extremely important for modern engine design, but can be computationally expensive. While the turnaround time for engine simulations ranges from one to a few days, it can takes 1-2 weeks to predict the injection conditions for a multi-hole injector. Hence, predicting the internal flowfields in the injector remains a computational bottleneck for engine design.

Argonne has developed a ML-based framework that addresses this computational bottleneck. In this framework, a deep-learning emulator enables fast and accurate predictions of the spatio-temporal flowfields of interest at the exit of the injector orifice, which allows for an efficient coupling between injector and engine simulations. By relating design parameters to the flowfields, the framework serves as a proxy for the actual CFD simulations and significantly accelerates the time-to-solution.

Recent work performed by Argonne researchers has demonstrated the predictive capability and speed‑up offered by the emulation framework. Across the range of operating conditions and fuel properties studied, the emulated and CFD-predicted injection conditions were in good agrement. However, the computational expense in emulating the injection conditions was reduced by a factor of 2 million relative to the traditional CFD approach. Using the emulated injection conditions, spray combustion predictions were found to be in excellent agreement with those from the baseline case, with errors of less than 4% across the range of conditions studied. The emulation framework can be flexibly adapted for different applications, where efficient and accurate predictions of spatio-temporal fields is desired.