In recent years, machine learning models have offered an efficient approach to studying geophysical fluid dynamics, particularly in scenarios where data availability is often limited. Here we present a study on the application of a Fourier neural operator (FNO) to the quasi-geostrophic (QG) system, an important system in geophysical fluid dynamics used to simulate large scale atmospheric flows. The primary objective of this research is to evaluate the performance of an FNO-based data-driven autoregressive model in predicting the evolution of the streamfunctions of the upper and lower layers of the QG system under various integration schemes, such as first-order Euler, fourth-order Runge-Kutta, as well as a simpler predictive approach where the FNO directly computes the next state in a sequence from the current state without intermediate calculations or corrections. The key question driving this study is the exclusion of the moisture channel from the training data, exploring whether or not we can effectively train the model on only partial states of data and still be able to get accurate assessments of large scale atmospheric flows. Our experiments demonstrate that while the FNO-based approach shows some promise in capturing the underlying dynamics of the QG system, excluding the moisture channel leads to challenges in achieving stable and accurate predictions. Our results demonstrate sensitivity of FNOs to missing state information, with evaluation metrics such as spectral analysis, Anomaly Correlation Coefficient (ACC), and Root Mean Square Error (RMSE) metrics showing us the impact of the moisture exclusion on the accuracy of the predictions.