Reservoir Computing (RC), a simplified form of recurrent neural network , is composed with the Fourier Transform to produce data-driven prediction of multi-scale coupled quasi-geostrophic flows. Experiments are conducted using the Modular Arbitrary-Order Ocean-Atmosphere Model (MAOOAM) [10], a coupled quasi-geostrophic model that includes a 2-layer atmosphere (fast dynamics) and 1-layer ocean (slow dynamics). The Fourier Reservoir Computing (FRC) approach produces forecasts that extend the skillful forecast horizon beyond a comparable RC model trained on data in physical grid space. The FRC approach can be enhanced by applying localization in physical space, which extends the skillful forecast horizon further and facilitates practical application to high-dimensional geophysical problems. Plain Language Summary The science of predicting changes in weather and climate is typically enabled using a combination of idealized physical models and instrument measurements. Here, we offer a contribution as part of a growing community that is attempting to advance the use of machine learning for weather and climate prediction. We use a simplified coupled atmosphere-ocean model to generate synthetic ‘observa-tions’, and show that reliable forecast models can be generating using machine learning applied to these observation data alone. We also provide an approach for scaling this method for use in more realistic operational weather prediction scenarios.