The agglomeration of colloidal particles is investigated numerically in the context of complex turbulent flows. For that purpose, Lagrangian one-point pdf methods are used to track the motion of parcels (which represent a group of real particles) in a turbulent flow simulated using RANS turbulence models while agglomeration is treated inside each cell of a mesh using PBE-like algorithms. One of the key issues with such approaches is related to the respect of the well-mixed condition, i.e. agglomeration should be computed on a set of parcels 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). To address this issue, a new data-driven spatial decomposition algorithm is proposed in this paper. Based on statistical information regarding the spatial distribution of particles (including the pdf), the algorithm extracts the optimal spatial decomposition that satisfies the well-mixed condition. After evaluating its convergence and accuracy, this optimal spatial decomposition is then used to obtain mesh-independent predictions of particle agglomeration.