Adaptive optics (AO) has become an indispensable tool for ground-based telescopes to mitigate atmospheric seeing and obtain high angular resolution observations. Predictive control aims to overcome latency in AO systems: the inevitable time delay between wavefront measurement and correction. A current method of predictive control uses the empirical orthogonal functions (EOFs) framework borrowed from weather prediction, but the advent of modern machine learning and the rise of neural networks (NNs) offer scope for further improvement. Here, we evaluate the potential application of NNs to predictive control and highlight the advantages that they offer. We first show their superior regularization over the standard truncation regularization used by the linear EOF method with on-sky data before demonstrating the NNs’ capacity to model nonlinearities on simulated data. This is highly relevant to the operation of pyramid wavefront sensors (PyWFSs), as the handling of nonlinearities would enable a PyWFS to be used with low modulation and deliver extremely sensitive wavefront measurements.