We found that likelihood computations in flow-based Boltzmann generators currently are a pressing problem hindering the approach form being scaled to large systems. Thus, our past efforts focused on overcoming this issues by developing a novel graph flow-based generative model (HollowFlow) that increased the sampling speed for these models by several orders of magnitude. More details about our prior success can be found in activity report of the preceding NAISS project (NAISS 2024/22-688). Most importantly, our new method and the corresponding experimental results so far require further experimental testing, especially on larger systems which will increase our requirements of computational resources. We plan to train HollowFlow on larger systems to demonstrate the scalability of our approach. Further, we plan to retrain HollowFlow on some of the small test systems we have used so far to further improve performance. This will be done by trying different underlying equivariant graph neural networks (PaiNN, E3NN) and also by adopting our approach to attention based mechanisms.