Machine-guided discovery of a real-world rogue wave model

Area of Science:

• Oceanography
• Physics
• Data Science

Background:

• Machine learning excels at pattern matching but struggles with scientific discovery due to misaligned goals.
• Scientific theories require causal consistency, interpretability, and reasoning capabilities beyond mere accuracy.

Purpose of the Study:

• To develop a method for using machine learning to discover interpretable, causal scientific models from data.
• To create a symbolic model for predicting oceanic rogue waves.

Main Methods:

• Utilized causal analysis, deep learning (artificial neural networks), and symbolic regression.
• Trained neural networks on observational data from wave buoys, focusing on causal features.
• Employed parsimony-guided model selection to distill complex models into interpretable equations.

Main Results:

• Developed a symbolic model for rogue waves that retains predictive power and allows for scientific interpretation.
• The new model reproduces known wave behaviors and generates well-calibrated probabilities.
• Achieved superior predictive performance on unseen data compared to existing theories.

Conclusions:

• Demonstrates machine learning's potential for inductive scientific discovery.
• Highlights the synergy between deep learning and symbolic regression for creating interpretable scientific models.
• Paves the way for enhanced forecasting of extreme oceanic events like rogue waves.