Collision detection (CD) has been the main part of several interactive applications such as computer vision, visual graphics, fluid mechanics, and solid dynamics, among others. In that context, one of the most preferred methods for collision detection is using particles to represent scenes and objects. Thus, one of the fundamental algorithms in particle simulations is the identification and iteration over nearest neighbors of every particle. During simulations, forces over the particles are computed and particle positions are updated in every simulation step. In order to find nearest neighbors for all particles, hash-based, grid-based, or tree-based approaches have been developed in the past. Prominent and fastest algorithms use virtual and explicitly allocated uniform grids to achieve high performance mainly when Graphics Processing Units (GPUs) are used. However, they also have disadvantages, mainly in terms of execution time and memory consumption. In this research work, we present a simple method, capable of improving performance in the collision detection stage, which can be applied to any particle-based simulation that allows us to represent much larger domains, since we consider particles of variable size to reduce the number of particles that allow us to represent objects. In addition, we also achieved an improvement in performance compared to previously existing methods that use particles of equal size.