A new and original method for the regression of thermodynamic properties of pure fluids has been recently proposed by Scalabrin et al. [1]. This is an extended corresponding states (ECS) technique in which the shape functions are expressed in a multi-layer feedforward neural networks (NN) framework. NN assure a very high flexibility of the functional form for the regression, reaching an accuracy which is comparable to that afforded by multi-parameter equations of state (EoS) in the representation of all the thermodynamic properties. The aim of the present work is to apply such a technique to 1,1,1,2,3.3,3-heptafluoropropane (HFC-227ea), one of the new generation of refrigerant fluids. A limited amount of experimental data has been published on this fluid over recent years. Although reliable, these data are sparse and the present method works better when fitting is achieved on regularly distributed data. Thus, a new experimental campaign was conducted at the Laboratoire de Thermodynamique, ENSMR for the acquisition of a large number of Pp T data between 25 3 and 393 K and up to 20 MPa. The regression procedure was carried out on a small subset of well-distributed density data, the so-called 'training set', and the model was then validated along all the data sets, including the literature data which report isobaric heat capacity, speed of sound and Joule-Thomson coefficient values. The accuracy of the method is evaluated for all the available properties, both in and outside the fitting range. It can be stressed that an accurate prediction of all thermodynamic properties is reached by fitting only a limited number of PrhoT data. This important result is due to the combination of an accurate multi-parameter EoS for the reference fluid R134a, with an improved version of the CS principle, incorporating modem NN techniques for mapping the two shape functions phi = phi(rho, T) and theta = theta (rho, T), which are characteristic of the ECS method.