The sparse identification of nonlinear dynamics (SINDy) algorithm provides a data-driven method for discovering nonlinear dynamical systems [1], though its effectiveness diminishes in noisy environments, especially with limited data. Certain sampling methods combined with SINDy, such as Ensemble-SINDy [2], have shown promise in enhancing the robustness of results. In this work, we introduce GS-SINDy, an improved approach that integrates group sparsity thresholds with similarity measures based on the Earth Mover’s Distance (EMD), aimed at enhancing robustness in identifying nonlinear dynamics and learning functions within dynamical systems governed by parametric ordinary differential equations. Built on the sequentially thresholded least squares (STLSQ) optimizer, GS-SINDy serves as a novel sampling method, offering an alternative to Ensemble-SINDy in certain aspects. We also show the convergence of the GS-SINDy based on [3]. Another major advantage of GS-SINDy is its capability to handle problems across multiple datasets generated under varying parameters. By enhancing interpretability and accuracy across diverse parametric conditions, GS-SINDy isolates essential dynamical features across datasets, thus improving the model's adaptability and relevance. We evaluated GS-SINDy on complex systems such as the Lotka–Volterra, Brusselator, Van der Pol, Lorenz, Hopf, and FitzHugh-Nagumo systems. For each, six trajectories were generated with distinct parameter settings. GS-SINDy was tested on (1) concurrently solving all six trajectories and (2) individually solving each trajectory. It consistently showed superior accuracy and reliability compared to SINDy and Ensemble-SINDy under different noise levels, underscoring its effectiveness across diverse applications.