Industry 4.0 has completely introduced a new ability to integrate physical and digital systems, therefore creating real-time monitoring, predictive insights, and better decision-making across industries. Digital twins lie at the core of this progress. A critical component in building a digital twin is the derivation of an accurate system model, a challenge addressed by numerous methodologies. Among these, the Koopman operator and Dynamic Mode Decomposition with Control have proven highly effective and have gained widespread acceptance within the scientific community. This study proposes a novel framework for the optimal selection of Koopman observables from a predefined dictionary of nonlinear functions. The objective is to determine the minimal set of observables that minimizes the prediction error between measured data and simulated samples from the obtained model. To address this, the study utilizes a genetic algorithm, showcasing its utility in optimizing the proposed approach for enhanced model fidelity.