This study presents the development and evaluation of a deep neural network (DNN)-based surrogate model for simulating solitary wave propagation in coastal environments. High-fidelity training data were generated via three-dimensional CFD simulations using OpenFOAM, targeting both flat and sloped bottom conditions. The surrogate model predicts time-series distributions of water surface elevation, pressure, and velocity at any given location, enabling real-time visualization of wave deformation. Four model configurations were tested by varying hidden layer depth (4 or 5 layers) and activation functions (ReLU or Swish). Results show that the 4-layer model using the Swish activation function achieved the highest accuracy across all evaluation metrics, closely matching the original CFD results, especially under deep-water and sloped conditions. In contrast, deeper networks showed signs of overfitting and reduced generalization. The surrogate model drastically reduced computation time—yielding predictions in under one second compared to over 500 s with OpenFOAM—demonstrating its practical utility for rapid design evaluation and educational applications. These findings suggest that well-trained DNN-based surrogate models can serve as efficient and accurate alternatives to conventional numerical simulations in wave-structure interaction studies.