In recent years, significant advancements have been made in solving N-Body Problems. The N-Body Solver proposed by Gonçalves, Ferrari, Boekholt, and Portegies Zwart introduces an innovative approach of decomposing the complex problem into multiple two-body sub-problems. By solving these sub-problems analytically and then combining their results, the method demonstrates a neural network-like computational strategy.

In this thesis we work with such a neural network framework for N-Body problem solving. We divide the complex system into manageable sub-problems that can be systematically integrated. A critical component of this approach is the Kepler Solver map, which attempts to resolve individual time-steps recursively. However, we identified this step as a potential performance bottleneck.

To address this limitation, we explore replacing the recursive Kepler Solver with machine learning models. Our investigation aims to determine whether these models can effectively handle large numbers of time-steps while maintaining computational efficiency. Additionally, we examine the potential of these models to solve inverse problems within N-Body systems, demonstrating the versatility of our approach.