Studying the particle collection on single fibers has been an attractive topic in both experimental and numerical fields since 80s. Despite the numerous studies, generating reproduceable data through experimental measurements is still a challenging task. On the other side, the numerical approaches are often too time consuming to reveal particle collection on a microfiber over several minutes or even hours (due to extremely small time intervals in the simulations). Without sacrificing the accuracy, the typical numerical approaches are numerically too expensive to be used for conducting parameter study or optimization purposes. With the help of machine learning concepts, this work is offering an alternative approach to tackle this problem. Computational Fluid Dynamics (CFD) is employed to compute the steady-state flow field around arbitrarily shaped cylinders, which resembles a single fiber with particles collected on it. In order to efficiently generate and simulate various fiber shapes without updating the mesh, an immersed boundary library, ABSFoam, is used. ABSFoam is an in-house C++ library developed for open-source CFD toolbox OpenFOAM®. Using the CFD simulations a dataset is built, which is utilized to train an Autoencoder. The trained neural network is able to predict the two-dimensional flow field (Ux , Uy and p) around arbitrary objects with Reynolds number between 1 and 10. With the data provided using this algorithm, the equation of motion for the particles could be efficiently solved to accurately determine the particle trajectory. Spherical drag and Brownian motion are taken into account as effective forces acting on the particles. By combining the flow field predictor and the particle trajectory solver a surrogate method is developed . The method shows a promising performance regarding both accuracy and reduction in overall calculation time and could be used as an alternative model to mere CFD simulations in the field of particle filtration.