Predicting the magnitude and timing of natural disasters is an important goal for scientists. But because of natural disasters’ rarity, there is no adequate data to allow them to make an accurate forecast. Massachusetts Institute of Technology and Brown University researchers have found a novel using artificial intelligence in predicting disasters. This new method could even surpass conventional methods.

Scientists employ machine learning and statistical algorithm to forecast disasters

In the latest study, researchers combined powerful machine learning and statistical algorithms that need fewer data to make accurate predictions to avoid using much data. 

Brown University’s applied mathematics and engineering professor George Karniadakis said that these stochastic events, such as (earthquakes, wildfires, hurricanes, etc.), are rare, and a lot of historical data is not available to make predictions. However, he said that there are enough past samples that can help scientists predict these future events. Karniadakis said the issue they address in their research is what would be the best data they can employ to minimize the number of data points needed. 

The researchers discovered that serial sampling with active learning was the best method. These statistical algorithms can study incoming information and gain insight from it to identify additional data sets that are equally important or more significant. In simpler terms, more may be accomplished with less knowledge. The machine learning model they employed is an artificial neural network called DeepOnet, which utilizes interconnected and stacked nodes to imitate the neuronal pathways of the brain.

How the system forecasts future disasters 

This technology combines the functionality of 2 neural networks into one, processing data across both networks. In the end, this enables enormous quantities of information to be examined in a very short time while also generating enormous quantities of data. By using DeepOnet and active learning approaches, the scientists could demonstrate that they can reliably identify warning signs of a catastrophic occurrence without a large amount of data.