<p>The dynamic manipulation of terahertz (THz) waves is important for sensing tasks where resonance features must be interpreted under varying measurement conditions. This work studies a graphene-assisted THz sensor based on a periodic array of gold nested split-ring resonators integrated with graphene patches, whose surface conductivity is adjusted through electrostatic biasing. Unit-cell behavior is evaluated through simulated scattering parameters (S₁₁ and S₂₁), and the resulting spectral responses are used to construct features for data-driven interpretation. The main contribution of this work is the integration of an additional classification stage into the numerical analysis of THz sensor data, providing an extended and more application-oriented sensing framework for sensor designers. To quantify the classification capability of sensing outputs, eleven learning approaches are benchmarked over ten independent runs and organized into four families: margin-based methods (SVM, polyhedral conic), instance-based learning (k-NN), probabilistic models (Gaussian NB, Bernoulli NB, LDA), and linear/neural discriminative models (Logistic Regression, Perceptron, Ridge, Deep Learning). In this evaluation, the margin-based group performs best overall, with SVM achieving the highest mean accuracy (96.42%) and low variability across runs (CV = 1.38%), followed by strong results from the linear/neural group (Perceptron at 94.17%, Deep Learning at 91.04%)incorporating a supervised classification step offers a practical way to interpret simulated THz response features in terms of application-relevant classes. Importantly, separability is not driven by any single resonant feature; rather, discrimination arises from the collective multi-frequency structure of the response. The reported performance and method rankings are therefore conditional on the present dataset, feature construction, and evaluation setup, and may differ for other sensor configurations, sensing conditions, or target analytes.</p>