Abstract
We investigate how to set the inlet temperature, and arrange a set of vehicle parts inside a paint curing oven, so as to maximize a non-convex, non-linear objective function. Standard methods for solving this kind of problem require a large number of objective function evaluations, each of which depends on a computationally expensive (minutes/hours) CFD simulation.
We replace the CFD solver with machine learning surrogates that can approximate the data required for an objective function evaluation extremely quickly (sub-second). We develop i) simulation surrogates that produce simulations that are structurally identical to their CFD-generated counterparts, and ii) objective function surrogates that learn an objective function directly.
We consider elementary learners (simple neural networks, non-linear regressions, Gaussian processes) and develop various techniques to use and combine them to solve single- and multi-criteria optimization problems.
We combine our surrogates in a configuration resembling a stack ensemble, and use it to solve the optimization problem at greatly reduced computational cost. We are thus able to explore multiple local maxima, and obtain solutions with higher objective function values than with traditional methods. Finally, we propose an approach that allows practitioners to throttle the computational effort until a satisfactory solution quality is achieved.