Rogue waves are rare surface waves in the ocean that are signicantly larger
than the general wave population. Although they pose a serious threat
to mariners, the causes of these waves in the real ocean are still poorly
understood, and they remain hard to forecast. This is due to the lack of a
high-quality observational dataset, the rarity of these waves and therefore
required amounts of data, the diculty of analyzing said data, and the lack
of a principled way to infer causation. This thesis consists of a collection
of 3 articles that address all of these issues through a combination of data
mining, interpretable machine learning, and causal analysis based on domain
knowledge. The rst article describes the assembly of a comprehensive wave
catalogue processing over 700 years of sea surface elevation time series from
158 buoy locations. The second article presents an analysis on the leadingorder
eects governing rogue wave formation based on interpretable machine
learning. The third article extends this to a fully nonlinear predictive model
by searching for a causally consistent neural network, and presents a path to
an improved rogue wave forecast. Finally, I discuss the implications of our
ndings for future rogue wave research, and outline how machine learning
can augment the scientic method and guide us towards scientic discovery.