Prediction is Very Difficult, Especially if it is About the Future (Thank You, Niels Bohr)
If you like getting wrapped up and tumbled about in complex systems, the ocean is happy to oblige. Consider any artificial structure subject to the whims of the sea. For example, the judges’ tower being erected (despite controversy) at Teahupo’o for the surfing portion of the Paris 2024 Olympics. When construction is finished, it will stand three stories high; an aluminum perch stapled to a reef in a Tahitian lagoon, seemingly safe from the roaring liquid cylinders, exploding several hundred yards away.
But the ocean offers no guarantees. A shift in tide, swell, wind direction, or any number of variables combined could see a freak wave march past its typical detonation point and batter the spindly pylon legs of the tower. What happens next is hard to predict. Themis Sapsis speaks of “motions becoming highly nonlinear,” which, in layman’s terms, means chaos and unpredictability. Now apply the same principles to offshore oil and gas rigs, wind turbines, or containerships.
Themis Sapsis speaks of “motions becoming highly nonlinear,” which, in layman’s terms, means chaos and unpredictability. Now apply the same principles to offshore oil and gas rigs, wind turbines, or containerships.
In addition to directing the Center for Ocean Engineering at MIT, he is also the principal investigator at the Stochastic Analysis and Nonlinear Dynamics (SAND) Lab. He is inspired by challenges in ocean engineering, though his work addresses problems that touch mechanical engineering, aerospace engineering, environmental engineering, and climate dynamics.
Part of their work involves addressing key challenges related to the prediction, quantification, and mitigation of extreme events. Challenge number one is the fact that extreme events are rare, which means a lack of data—an essential component for training a machine learning model to aid in the prediction process. The second key issue according to Sapsis, is tied to the complex dynamics that govern extreme events. Chaos compounds the problem.
Imagine you have ground truth data on the daily weather conditions for the last 40 years. Can this information be used to improve the modeling of extreme weather events? According to Sapsis, “Application of standard machine learning methods is limited due to the chaotic character of the underlying dynamics, and chaos is associated with high sensitivity, which makes it very hard to create meaningful associations between data and model behavior.”
“At SAND Lab, we spend a lot of time and effort trying to formulate algorithms at the intersection of dynamical systems, probability, and machine learning, to address exactly these types of challenges.”
To address the rarity aspect of extreme events, Sapsis, and his researchers start by designing multiple optimal experiments to consider—which allows them to identify the experiments that are the most informative for predicting extreme events.
Think of a ship forced to contend with random waves. You could run an experiment in which you generate random waves for an extended period and then analyze the statistics from the time series. However, it would take a significant amount of time to yield actionable data because extreme events (the appearance of random waves) are so rare. Sapsis calls this the “brute force approach.”
Not one for brute force, Sapsis explains the technique developed at Sand Lab: “We try to parameterize all the possible random waves through a mathematical construction, and then choose the parameters of the waves so that we can explore events that are not only extreme but also feasible.”
This is not a thought experiment; there are real world implications. You don't want to spend all of your computational or experimental resources attempting to understand waves that don’t show up. But it’s a fine line, because, as Sapsis explains, there is little to be gained from focusing on a boring scenario (i.e., “non-extreme events; the situation that happens every day,” as he puts it).
This type of careful exploration that is both interesting but also feasible, is one of Sapsis and SAND Lab’s key innovations, and it allows them to capture statistically extreme events while keeping computational costs at an acceptable level.
Interestingly, the algorithms that Sapsis and his team have been developing to predict extreme events with limited data work just as well when it comes to optimizing for extreme performance while using the minimum amount of computational resources. Conceptually, they are similar problems, after all. The common factor among scenarios is the presence of a high-dimensional parameter space to explore.
The applications here range from how to perform the exploration of an unknown area for an event that is of high interest (Blanchard and Sapsis, 2022, Ocean Engineering), to transforming a cell through cellular reprogramming to achieve desirable properties (Zhang et al., 2023, Nature Machine Intelligence).
These days, Sapsis and his team are also exploring questions around the intensity and frequency of future extreme weather events caused by climate change. The greatest stumbling block: modeling events that have never happened before. In the face of uncertainty, scarcity of data, and finite computational resources, Sapsis says SAND Lab is developing computational algorithms that combine aspects from multiscale analysis, machine learning, and extreme event theory to analyze the characteristics of future extreme events due to climate change.
“The ultimate goal of our climate-focused analysis is to inform stakeholders, so they can better prepare and mitigate the effect of climate events,” he says. It is valuable work with far-reaching implications across disciplines and sectors. Policy makers, government, and insurance agencies all stand to benefit from the work at Sand Lab. Urban planners designing the cities of the future benefit from understanding the probability of future extreme weather events occurring and their frequency of occurrence. As do the designers of off-shore wind turbines, among other things.
“It's not a matter of over-designing or under-designing a structure,” according to Sapsis. “The probability of extreme events defines the financial feasibility of these huge projects. It's critical to be able to provide high accuracy so stakeholders can design structures more accurately and more effectively.”
As for the current generative AI boom, Themis sees hope rather than hype: “Extreme events could be a beautiful area [for generative AI] because we have a dire need to generate new possibilities of extreme events, or extreme events that are previously unseen. At Sand Lab, we’re trying to push the limits and take advantage of what generative AI has to offer.”
Admittedly, generative AI and the prediction of extreme events might seem like strange bedfellows—generative AI craves an abundance of training data; there is a lack of data around extreme events. But Sapsis is working diligently on a solution: “One of the of the frontiers we're trying to push is exactly how to generate events that have never been seen before. These events basically live, as we say, in the tail of the distribution—the unknown that can be explored only through the combination of physical laws and previous experience (i.e. data.).”