Turbulence modelling in turbomachinery flows remains a challenge, especiallywhen transition and separation phenomena occur. Recently, several researchefforts have been devoted to the improvement of closure models for ReynoldsAveraged Navier-Stokes (RANS) equations by means of machine learning ap-proaches which make it possible to extract the knowledge hidden inside theavailable high-fidelity data (from experiments or from scale-resolving simu-lations). In this work the use of the field inversion approach is investigatedfor the augmentation of the Spalart-Allmaras RANS model applied to theflow in low pressure gas turbine cascades. As a first step, the field inversionmethod is applied to the T106c cascade at two different values of Reynoldsnumber (80000-250000): an adjoint-based gradient method is employed inorder to minimise the prediction error on the wall isentropic Mach numberdistribution. The data obtained by the correction field are then analysedby means of an Artificial Neural Network (ANN) which makes it possible to generalise the correction by finding correlations which depend on physicalvariables. A study on the definition of the input variables and on the archi-tecture of the ANN is performed. Different kind of corrections are evaluatedand a particularly robust correction factor is obtained by limiting the rangeof the correction in the spirit of intermittency models. Finally, the ANN isintroduced in an augmented version of the Spalart-Allmaras model which istested on the T106c cascade (for values of the Reynolds number not consid-ered during the training) and for the T2 cascade. The prediction ability ofthe method is investigated by comparing the numerical predictions with theavailable experimental data not only in terms of wall isentropic Mach numberdistribution (which was used as goal function during the field inversion) butalso in terms of mass averaged exit angle and kinetic losses.< Réduire