Inversion of geophysical data is often challenging and time-consuming, particularly for large scale surveys. The solution of the inverse problem requires numerous calculations of the forward problem, especially when calculating partial derivatives required for most linearized inversion schemes. The forward model is usually calculated numerically using accurate equations, but often less accurate and faster equations are used. In recent years, neural networks have become increasingly popular to replace the numerical forward modelling, as this may lead to a significant speed-up. Data normalization, prior to the training of neural networks, is crucial to obtain good results and faster convergence rate. This is especially true for geophysical data, as numerical data values may span over several orders of magnitude. In this abstract, we investigate several normalization approaches for TEM data, with a special focus on towed TEM data. Through extensive experimentations, we show that data normalization substantially affects the performance of neural networks when surrogating forward models. We also demonstrate the effect of normalized data variation on neural network’s performance and provide insights into which normalization approaches may be better than others. A significant improvement in performance accuracy is achieved when the appropriate data normalization technique is employed.