Researchers at CNH Industrial, a major OEM in the off-highway sector, wanted to improve the design of their existing Diesel Exhaust Fluid (DEF) doser in order to increase evaporation rates of liquid DEF and improve overall NOx conversion performance of their SCR system. Traditionally, such a study would involve comprehensive CFD (Computational Fluid Dynamics) simulations, which can be computationally expensive and resource intensive. 

Using a large body of previously run CFD experiments, the researchers trained a ML model that takes as inputs the design parameters of their baseline DEF doser and uses the CFD-simulated results of the evaporation rates of liquid DEF as labeled data for the model. This allows the ML model to act as a stand-in for CFD, allowing much faster turnaround times on experimenting with various combinations of the design parameters of the doser. To converge on parameter values that yield the highest evaporation rates, the ML model was coupled with a Genetic Algorithm (GA) to rapidly scan the large design space for the optimal design parameter values. The GA is designed to model how a species evolves with each generation. It selects the top 50% best performing combinations of inputs, while rejecting the bottom half. The top half is then cross bred to create another full set of inputs and the process is continued until the optimization target is achieved (maximization of evaporation rates over X executions). 

To further increase robustness of this method, a random mutation is introduced every X executions of the GA. The overall result of this combined ML/CFD approach is a much faster convergence on the set of parameters that yield the highest evaporation rates. The converged design parameter values were then validated with a full CFD run and the results were shown to be within 7.5% of the predicted value. Overall improvement over the base doser design was shown to be around 13%, but the real value of this method is in the brevity of computational resources and time taken to get to the optimized design parameters values. Instead of performing a CFD run for every single permutation and combination of inputs, the researchers were able to quickly reject a large body of suboptimal combinations to arrive at an optimized set of values in a fraction of the time. 

Techniques such as these are expected to grow in scale and scope as computational prowess and proficiency in data-driven methods continues to increase across all fields. If you are a researcher or working professional using tools that model complex physical/chemical systems in your current line of work, it may be worth hunting for areas in your project(s) where computationally leaner tricks such as the one discussed here can help reduce computational overhead and developmental costs.