his paper presents a systematic methodology for construction of high-level performance models using least squares support vector machine. The transistor sizes of the circuit- level implementation of a component block along with a set of geometry constraints applied over them define the sample space. Optimal values of the model hyper parameters are computed using genetic algorithm. The novelty of the methodology is that the models constructed with this methodology are accurate, fast to evaluate with good generalization ability and low construction time. The present methodology has been compared with two other standard methodologies and the novelties are clearly demonstrated with experimental results.