Machine-Learning Technique Speeds Extreme Event Analysis

Engineers at MIT in the U.S. have developed an algorithm based on machine-learning that quickly pinpoints the types of extreme events that are likely to occur in a complex system, such as when waves of varying magnitudes, lengths and heights can create stress on a ship or offshore platform. 

Compared with traditional methods, the team’s technique provides a much faster, more accurate risk assessment. Engineers typically gauge a structure’s endurance to extreme events by using computationally intensive simulations to model a structure’s response to, for instance, a wave coming from a particular direction, with a certain height, length and speed. These simulations are highly complex, as they model not just the wave of interest but also its interaction with the structure. By simulating the entire “wave field” as a particular wave rolls in, engineers can then estimate how a structure might be rocked and pushed by a particular wave and what resulting forces and stresses may cause damage.

These risk assessment simulations are incredibly precise and in an ideal situation might predict how a structure would react to every single possible wave type, whether extreme or not. But such precision would require engineers to simulate millions of waves, with different parameters such as height and length scale — a process that could take months to compute. 

As a more practical shortcut, engineers use these simulators to run just a few scenarios, choosing to simulate several random wave types that they think might cause maximum damage. If a structural design survives these extreme, randomly generated waves, engineers assume the design will stand up against similar extreme events in the ocean.

But in choosing random waves to simulate, engineers may miss other less obvious scenarios, such as combinations of medium-sized waves or a wave with a certain slope that could develop into a damaging extreme event.

“With our approach, you can assess, from the preliminary design phase, how a structure will behave not to one wave but to the overall collection or family of waves that can hit this structure,” says Themistoklis Sapsis, associate professor of mechanical and ocean engineering at MIT.

Instead of running millions of waves or even several randomly chosen waves through a computationally intensive simulation, Sapsis and former student Mustafa Mohamad developed a machine-learning algorithm so that they can quickly feed in various types of waves and their physical properties, along with their known effects on a theoretical offshore platform. From the known waves that the researchers plug into the algorithm, it then “learns” and make a rough estimate of how the platform will behave in response to any unknown wave. 

Through this machine-learning step, the algorithm learns how the offshore structure behaves over all possible waves. It then identifies a particular wave that maximally reduces the error of the probability for extreme events. This wave has a high probability of occurring and leads to an extreme event. In this way the algorithm goes beyond a purely statistical approach and takes into account the dynamical behavior of the system under consideration.

Sapsis says that the technique is not limited to ships and ocean platforms but can be applied to any complex system that is vulnerable to extreme events. For instance, the method may be used to identify the type of storms that can generate severe flooding in a city and where that flooding may occur. It could also be used to estimate the types of electrical overloads that could cause blackouts and where those blackouts would occur throughout a city’s power grid.