Modern gamma-ray astronomy utilizes machine learning to tackle three crucial problems: Estimating the type of particle recorded by a telescope, estimating the particle's energy and estimating the particle's spatial origin.

To learn these models, we rely on sophisticated physical simulations that generate large amounts of synthetic training data. As with any simulation, there is a mismatch between reality and simulation that can decrease the performance of our machine learning models in the real world.

Good knowledge of Machine Learning is prerequisite for this Bachelor thesis. This can be illustrated by e.g. very good grades in a Proseminar on the topic of machine learning or successful participation in a Fachprojekt on Data Mining or Machine Learning. 

We want to apply techniques from adversarial machine learning to learn features that are indistinguishable between real-world data and simulated data. To this end, we want to extend the existing neural network architectures that solve the aforementioned three ML-tasks with an adversarial neural network that punishes feature representation that make it easy to identify whether an example is real or simulated.