A new way to predict ship-killing rogue waves

Artificial-intelligence (AI) models are modern oracles. The neural networks that power them are flexible mathematical tools, capable of finding any pattern, fitting any shape and drawing any line. They are used to forecast the weather, anticipate road maintenance and diagnose diseases. The problem is, since they train themselves to accomplish those tasks, no one really knows exactly how they do it.

This “black-box problem” makes it hard to rely on such models, especially when, as in health care, they are making high-stakes decisions. It makes them less useful for scientists, too, who are interested not only in predicting an outcome but also in understanding why that outcome happened.