Deep learning has made strides in pattern recognition within the climate sciences within recent years, mainly for the recognition of weather processes such as atmospheric rivers and tropical cyclones within climate datasets. However, a critical question remains: can deep learning be used to drive pattern discovery, rather than recognition, within atmospheric and climate science? We show that deep learning can, indeed, be used to drive pattern discovery within atmospheric science by: 1) recognizing the state of an atmospheric pattern characterized by complex scale interactions, 2) expanding upon existing proxies for the pattern to create new metrics based on an expanded set of atmospheric variables, and 3) subsequently being used to test scientific hypotheses related to the physics of the atmospheric process.
We use the Madden-Julian Oscillation (MJO) as a case study. The MJO is a complex, non-linear atmospheric wave with active efforts to develop proxies for its state in both weather and climate applications. We train a state-of-the-art convolutional neural network (CNN) on MERRA-2 reanalysis that successfully identifies the state of the MJO. The developed deep learning proxy for the MJO can be successfully applied to other reanalysis datasets which supports its generalizability. Furthermore, the CNN understands the multi-variate structure of the MJO beyond the constraints provided by the truth classification. We use this fact to offer insights into how deep learning can be used as a tool to test scientific hypotheses related to the MJO, driving climate discovery. We end by summarizing the significant implications of these results for the usage of deep learning in the atmospheric and climate sciences beyond the MJO. Our findings serve as a turning point for using deep learning to enhance understanding of the physics of the atmosphere.