A team of researchers from Brown University and the Massachusetts Institute of Technology (MIT) have combined statistical algorithms, which need less data to make accurate and efficient predictions, with a powerful machine learning technique and trained it to predict scenarios, probabilities, and sometimes even the timing of rare events, such as earthquakes, pandemics, or tsunamis, despite the lack of historical records about them, to try to predict them.
And it is that when it comes to predicting disasters caused by extreme phenomena, computational modeling faces the almost insurmountable challenge that, statistically speaking, these phenomena are so rare that there is not enough data about them to use predictive models that allow us to accurately predict when will happen.
Thus, the research team found that this new framework may provide a way to circumvent the need for the massive amounts of data traditionally required for such calculations, essentially reducing the huge challenge of predicting rare events to a matter of quality. over quantity, as published in the journal ‘Nature Computational Science’.
“You have to keep in mind that these are stochastic phenomena,” says George Karniadakis, professor of applied mathematics and engineering at Brown and author of the study. An outbreak of a pandemic like COVID-19, an environmental catastrophe in the Gulf of Mexico, an earthquake, huge forest fires in California, a 100-foot wave capsizing a ship… these are rare events, and because they are rare We don’t have a lot of historical data. We don’t have enough samples from the past to predict them in the future. The question we address in the study is: What is the best possible data we can use to minimize the number of data points we need?”
The researchers found the answer in a sequential sampling technique called active learning. These types of statistical algorithms are not only capable of analyzing the data that is fed into them, but more importantly, they can learn from the information to label new relevant data points that are equally or even more important to the outcome that they are. is being calculated. At the most basic level, they allow you to do more with less.
That’s central to the machine learning model the researchers used in the study. Called ‘DeepOnet’, the model is a type of artificial neural network that uses nodes interconnected in successive layers that roughly mimic the connections made by neurons in the human brain.
‘DeepOnet’ is known as deep neural operator. It is more advanced and powerful than typical artificial neural networks because it is actually two neural networks in one, processing data in two parallel networks. This allows you to analyze gigantic data sets and scenarios at breakneck speed to spit out equally massive sets of probabilities once you learn what you’re looking for.
The bottleneck of this powerful tool, especially when it comes to infrequent events, is that deep neural operators need tons of data to train themselves and make effective and accurate calculations.
In the study, the research team demonstrates that, combined with active learning techniques, the DeepOnet model can be trained to know what parameters or precursors to look for that lead to the disastrous event being analyzed, even when there are not many data points. .
“It’s not about taking all possible data and feeding it into the system, but about proactively looking for events that signal rare events,” explains Karniadakis. We may not have many examples of the actual event, but we do have those precursors. Through mathematics, we identify them, which along with actual events will help us train this data-hungry operator.”
In the study, the researchers apply the method to the determination of parameters and different probability ranges of dangerous peaks during a pandemic, to the detection and prediction of tsunamis, and to the estimation of when a ship will break in half due to stress.
For example, in the case of tsunamis—those that are twice the size of surrounding waves—the researchers found that they could discover and quantify when they would form by looking at likely wave conditions that interact nonlinearly with time. , giving rise to waves that are sometimes triple their original size.