Physics captured by data-based methods in El Ni\~no prediction

On average once every four years, the Tropical Pacific warms considerably during events called El Ni~no, leading to weather disruptions over many regions on Earth. Recent machine-learning approaches to El Ni~no prediction, in particular Convolutional Neural Networks (CNNs), have shown a surprisingly high skill at relatively long lead times. In an attempt to understand this high skill, we here use data from distorted physics simulations with an intermediate complexity El Ni~no model to determine what aspects of El Ni~no physics are represented in a specific CNN-based classification method. We find that the CNN can adequately correct for distortions in the ocean adjustment processes, but that the machine-learning method has far more trouble to deal with distortions in upwelling feedback strength.