Hydrodynamic interactions in suspensions of small particles at low Reynolds number are classically described by Stokesian dynamics, where linearity of the Stokes equations leads to mobility or resistance formulations coupling all particles through long-ranged many-body interactions. While analytically well founded, these formulations rapidly become computationally prohibitive for large systems due to the dense structure of the mobility matrix and the need to resolve near-field effects. This work presents a surrogate modelling framework that combines physical insight from Stokesian dynamics with data-driven learning to efficiently approximate many-body hydrodynamic forces \cite{kimKarrila}. We introduce an Interaction-Decomposed Neural Network (IDNN) surrogate model \cite{IDNN}. The IDNN learns mapping from local neighbourhood geometry to the hydrodynamic force acting on a target particle. The implementation enforces permutation, rotation, and reflection invariance, meaning that the model is easily implementable in Lagrangian particle simulation solvers, such as OpenFoam. Trained on a large database of particle-resolved DNS data using the Boundary Element Method, the IDNN captures many-body corrections beyond merely pairwise superposition of neighbour contributions. The proposed approach provides a surrogate for the expensive matrix assembly, needed by the classical Stokesian dynamics and the costly evaluation of the velocity laplacian, needed by the Faxén model for performing the same task. We demonstrate how physics-informed surrogate modelling can preserve the structure of Stokesian dynamics while achieving gains in computational efficiency for particle-resolved flow simulations. We employ the IDNN model to simulate sedimentation of particle suspension of varying volume fractions and compare the results to classical Stokesian dynamics model.