Physics-based simulation tools for spray painting exist but are not fast enough to be useful for automatic optimization of the spray painting process.
The results of spray painting depend primarily on the process parameters, the path of the paint applicator and the geometry of the target object. These factors affect the air flow and the electrostatic field and are hard to incorporate in approximate simulation tools.
This thesis proposes a novel approach based on combining fast, approximate simulations with machine learning based error correction.
The proposed approach is to create a height profile of the target geometry from the local coordinate system of the paint applicator. This height field captures most parameters that affect the simulation error and can be used as an input for machine learning regression algorithms. These algorithms are then trained to estimate the painting error. The training is performed with a set of samples from common painting scenarios that are generated beforehand.
Creating a training set for dynamic simulations is time-consuming. Static simulations can sufficiently approximate dynamic simulations and are therefore used for training. This drastically improves the time to create training sets and reduces training time for the machine learning models.
Linear regression, tree-based regression models and support vector regression are compared on benchmarking problems and especially tree-based regression methods show promising prediction accuracy and are able to reduce the projection error more than 40% on real world benchmarks.
Tree-based models are also the fastest algorithms among the compared regression models.
Finally, a way to integrate the proposed method into the simulation framework is presented. The results are investigated for different artificial and real world painting scenarios.