The project, funded by the German Federal Ministry of Education and Research and managed by the University of Hamburg under the leadership of Professor Dr. Gregor Kasieczka, ran from 1 October 2018 to 30 June 2022. It was part of the IDT‑UM consortium and received a total grant of 189 177,57 €. The main goal was to develop new neural‑network architectures for automated learning and to deepen the understanding of fundamental processes in experimental particle physics. Two interrelated research strands were pursued: the study of decision behaviour and statistical properties of neural networks, and the creation of efficient methods for processing and simulating complex data structures.

In the first strand, the focus was on generative models such as Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and Normalising Flows. These models were trained on initial datasets to learn their probability distributions and then used to generate additional samples through rapid sampling. The research demonstrated that the outputs of generative models provide a closer approximation to the underlying distribution than the training data themselves, a result that was extended to particle‑physics‑relevant distributions. By correlating the latent space of surrogate models with physical properties of detector data, the team was able to interpret the models more transparently and identify architectural improvements that enhanced performance. Bayesian techniques were also applied to quantify uncertainties in the generative process, while symmetry properties of the data were exploited to further constrain the models.

The second strand addressed the challenges of representing high‑dimensional detector data. Traditional grid‑based structures were replaced by graph or point‑cloud representations, enabling the application of graph neural networks. Together with other groups in the IDT‑UM consortium, a public dataset was created and a unified graph architecture was demonstrated for its analysis. This effort led to the most accurate simulation to date of electromagnetic and hadronic calorimeters, achieving unprecedented fidelity in reproducing the behaviour of complex detectors. The generative models were used to produce realistic detector responses at a fraction of the computational cost of full Monte‑Carlo simulations, thereby accelerating data analysis pipelines.

The project’s outputs were disseminated through publications in international journals and presentations at workshops and conferences. All software code and relevant data were made publicly available to facilitate reproducibility and further research. In addition to the scientific achievements, the project contributed to the training of bachelor, master, and doctoral students, equipping them with skills that are valuable both academically and industrially.

Overall, the research advanced the application of machine‑learning techniques in particle physics, providing more efficient simulation tools and deeper insights into the statistical behaviour of neural networks. The collaboration between the University of Hamburg, the IDT‑UM consortium, and the funding bodies ensured a comprehensive approach that combined theoretical development, practical implementation, and community engagement.