This article presents a new data-driven spatial decomposition algorithm that allows to split a domain containing point particles into elementary cells, each cell containing a spatially-uniform distribution of particles. For that purpose, the algorithm relies on the use of statistical information for the spatial distribution of particles and then extracts an optimal spatial decomposition. After evaluating the convergence and accuracy of the algorithm on homogeneous and inhomogeneous cases, this optimal spatial decomposition is applied to study the case of particle agglomeration. Indeed in CFD context, recent developments on numerical simulations of particle agglomeration in complex and turbulent flows increasingly resort to Euler-Lagrange approaches. These methods are coupled with population balance equation (PBE)-like algorithms to compute agglomeration inside each cell of the Eulerian mesh. One of the key issues with such approaches is related to the respect of the spatially-uniform condition, i.e. agglomeration should be computed on a set of particles that are uniformly distributed locally in each cell. Yet, CFD simulations in realistic industrial/environmental cases often involve non-homogeneous concentrations of particles (due to local injection or accumulation in specific regions). We show that more accurate and mesh-independent predictions of particle agglomeration are made possible by the application of this new data-driven spatial decomposition algorithm.