Finite Element Method (FEM) simulations provide highfidelity solutions in solid mechanics but are often computationally expensive for large parametric studies and real-time applications. We present a neural-network-based surrogate framework that learns mappings from material and geometric parameters to nodal displacements using a compact Multi-Layer Perceptron (MLP). Datasets were generated with the ONSAS.m solver across parametric sweeps for two benchmark problems (uniaxial prism and a composite cantilever). The trained surrogate achieves low in-domain RMSE while offering orders-of-magnitude faster inference. We further analyze extrapolation performance, provide visual comparisons of predicted vs. FEM displacements, and discuss limitations when generalizing to heterogeneous materials. Finally, we outline directions to improve physics consistency and uncertainty quantification using operator learning and physics-informed models.