This work presents the development and application of a machine learning model to predict the unsteady performance of a turbocharger radial turbine subject to on-engine pulsating flow conditions. The model proposed, based on a fully connected neural network, predicts the instantaneous turbine torque and circumferentiallyaveraged relative inflow angle when given, as only inputs, the total pressure and temperature pulses and the time derivative of the total pressure pulse upstream of the turbine. The training data set for the model is obtained from an experimentally-validated Reynolds-averaged Navier-Stokes model and consists of various operating conditions characterized by different pulse amplitudes and frequencies. Heat transfer effects are neglected by the use of adiabatic boundary conditions while the rotational speed of the rotor is otherwise maintained fixed. Based on the results obtained, the model is shown capable of accurately predicting the turbine torque and local properties of the flow, such as the relative inflow angle, with a high degree of accuracy (coefficient of determination larger than 0.98). At first, the model is tested for both interpolation and extrapolation conditions. Given a training data set constituted by only three pulses characterized by different amplitudes, the model accurately predicts the turbine performance both for conditions inside and outside the range of amplitudes of the training data set. Lastly, the model is trained with a larger data set, including both variations of the pulse amplitude and frequency, in order to predict the performance of the turbine subject to a more general pulse shape. The error indicators show an improvement with respect to the extrapolation case due to the larger size of the training data set, showing also a great capability of the model to predict the unsteady performance of the turbine for a more general pulse shape. The model represents a fast and efficient approach for predicting the unsteady turbine performance as compared to more complex experimental set-ups and time-consuming 3D numerical simulations and a valid alternative to the more common 0D-1D models.