According to Brown and MIT researchers, scientists can use a combination of advanced machine learning and sequential sampling techniques to predict extreme events without the need for large datasets.

When it comes to predicting disasters caused by extraordinary events (earthquakes, epidemics, or “rogue waves that can destroy coastal structures”), computer modeling faces an almost insurmountable problem: statistically speaking, these events are so rare that they simply aren’t enough. . Data about them to use predictive models to accurately predict when they will happen next.

But a group of scientists from Brown University and the Massachusetts Institute of Technology suggest that this shouldn’t be the case.

in a published study Nature Computational ScienceIn , the researchers describe how they use statistical algorithms that require less data to make accurate predictions, combined with powerful machine learning techniques developed at Brown University. This combination allowed them to predict scenarios, probabilities and even timeframes for rare events despite the lack of historical data.

In doing so, the research team discovered that this new framework could provide a way to circumvent the need for the huge amounts of data traditionally required for this type of computation, and instead reduce the great difficulty of predicting rare events to a quality issue. excess amount

“You have to understand that these are stochastic events,” said George Karniadakis, professor of applied mathematics and engineering at Brown and author of the study. “An epidemic like COVID-19, an environmental disaster in the Gulf of Mexico, an earthquake, major wildfires in California, a 30-foot wave that capsizes a ship are all rare events, and we don’t do that because they’re rare.” Having lots of historical data We don’t have enough samples from the past to predict their future. The question we address in the article is: What is the best data we can use to minimize the number of data points needed?’

The researchers found the answer in a sequential sampling method called active learning. Such statistical algorithms not only analyze the data fed to them, but more importantly, they can learn from the information to mark new relevant data points that are equal or even more important than the calculated result. At the most basic level, they let you do more with less.

This is critical to the machine learning model the researchers used in the study. The model, called DeepOnet, is a type of artificial neural network that uses interconnected nodes in successive layers that roughly mimic the connections made by neurons in the human brain. DeepOnet is known as a deep neural operator. It is more advanced and powerful than typical neural networks because it is essentially two neural networks in one and processes data in two parallel networks. This allows it to analyze huge datasets and scenarios at breakneck speed to discard equally large probability sets once it knows what it’s looking for.

The bottleneck of this powerful tool is that, especially when it comes to rare events, deep neural operators need a lot of data to train to make efficient and accurate calculations.

In the paper, the research team demonstrates that the DeepOnet model, when combined with active learning techniques, can learn even if there aren’t many data points to look for, which parameters or antecedents to look for that lead to a devastating event that someone is analyzing.

“The point is not to take all possible data and put it into the system, but to actively search for events that would represent rare events,” Karniadakis said. said. “We may not have many examples from real events, but we may have these messengers. We use the math to determine them, and it will help you train this data-hungry operator along with real-world events.”

In the paper, the researchers apply the approach of determining the parameters and different probability ranges of dangerous spikes during a pandemic, finding and predicting rogue waves, and predicting when a ship will crack in half due to stress. For example, researchers have found that they can detect and measure when spurious waves occur by looking at possible wave conditions where waves twice as large as surrounding waves interact non-linearly over time, sometimes resulting in waves three times their original size.

The researchers have discovered that their new method outperforms more traditional modeling attempts, and they believe it provides a framework that can effectively detect and predict all kinds of rare events.

In the paper, the research team explains how scientists should design future experiments to minimize costs and improve prediction accuracy. Karniadakis, for example, is working with environmentalists to use a new method to predict climate events like hurricanes.