Complex dynamical systems, such as active matter composed of large populations of self-propelled particles interacting with each other, present significant challenges for traditional numerical simulations. Recent advances in machine learning offer a promising alternative to address the most computationally demanding problems. In this work, we evaluate the prediction performance of several state of the art neural networks on an active matter dataset from The Well benchmark, a collection of diverse physical simulations. Specifically, we replicate experiments using three benchmark models: CNextU net, classical U-net, and Fourier Neural Operator, all trained under identical conditions and compared using the Variance scaled Root Mean Squared Error (VRMSE) metric. To further improve efficiency, we introduce a U-net variant that replaces the traditional hyperbolic tangent activation functions with Rectified Linear Units. Within a fixed ten hour training budget, the ReLU-based variant completes substantially more epochs than its Tanh-based counterpart and achieves competitive VRMSE values surpassing previously reported results on the active matter dataset. Our findings demonstrate that modest architectural changes, combined with careful hyperparameter tuning, can yield high-accuracy prediction.