The presence of rigid body particles in fluid flows can affect the fluid properties in non-trivial ways. In creeping flows, the effect of the particles can be modeled efficiently by the mobility problem, a system of linear equations that relate external forces acting on the particles to resulting velocities of each individual particle. The mobility problem is typically complicated by two main factors. First, lubrication effects between closely interacting particles requires high resolution in the discretization, which results in costly evaluation. Secondly, long range effects between particles result in a dense system which is costly to evaluate efficiently. There have been many different approaches to solving the mobility problem, such as Stokesian dynamics in its many forms, the method of fundamental solutions, boundary integral methods and the multi blob method. These methods typically achieve near-linear complexity with respect to the number of particles, and are orders of magnitude faster than naive methods. On the other hand, the mobility problem still remains a major bottleneck in simulations of fluids containing particles, and recent efforts have therefore been made to develop methods that sacrifice accuracy to become computationally feasible. In particular, deep learning has emerged as a promising alternative to classical approaches. The current generation of deep mobility solvers are essentially graph neural network that operate on small clusters of particles at a time. While these networks are orders of magnitude faster than the classical methods, they exhibit two major flaw. First, they omit long range multi body interactions, and secondly, they scale quadratically with the particle count, and therefore are only more efficient up to a certain limit. In this project, we aim to develop a new generation of neural architectures for solving the Stokes mobility problem, that exhibit the same linear scaling as classical methods, and account for long range multi body interactions. The main purpose of this project is to generate a considerable amount of training data. This will involve solving the Stokes mobility problem thousands of times for particle clouds containing on the order of 10 000 particles. Such a task is trivially parallelizable and therefore well suited for HPC systems.