 45 minutes. That's how long it took, on average, for the 2011 Tohoku tsunami to go from earthquake to Japan's most destructive natural disaster. While that was enough time to signal a warning to those in most imminent danger, flood scientists say warning systems have to do better. Here's why. A standard computer takes about 30 minutes to run a conventional physics-based flood model. Depending on the dataset used, forecasters must also determine the source of a tsunami, which can take up to 35 minutes. Add that up, and it's clear that there is room for improvement. That's inspired researchers from Japan to look to AI for a solution which they say could cut modeling time by up to 99%. The team is led by researchers from three RIKEN centers in Japan, the Cluster for Pioneering Research, the Center for Advanced Intelligence Project, and the Center for Computational Science. Together, they've designed a machine-learning algorithm that takes in real-world data from Japan's Seafloor Observation Network for Earthquakes and Tsunamis, or SNET, and outputs realistic predictions of tsunami-caused flooding. The model was trained on more than 3,000 hypothetical scenarios, ranging from megathrust events, like the one that gave rise to the 2011 tsunami, to less intense outer-rise earthquakes. The team then tested the model using two sets of scenarios, a collection of 480 earthquakes of varying magnitude, and three, real historical tsunami events. The model results compared well with those produced by a conventional physics-based model. These included metrics such as inundation height, how high a tsunami stands above sea level, and flow depth, how high a tsunami rises above ground level. The difference, of course, was how long the models took to produce those results. The physics-based model delivered its predictions in approximately 30 minutes, the machine-learning model in a mere 5-hundredths of a second. This tremendous cut in time cost is due in part to the machine-learning model's intensive training. Instead of having to solve the fluid dynamics equations that predict flooding on the fly, the model relies on its massive knowledge of real and hypothetical events to venture a prediction. The caveat here is understanding sources of uncertainty. Any error in how data are measured by the S-Net observational network or in how their input into the model itself will undoubtedly affect the results. Dampening these sources of uncertainty is the goal of future and ongoing work. But the merits are undeniable. Being able to predict flooding fast could give citizens more time to make decisions about their safety, and it could give first responders the opportunity to mobilize faster and cover more ground. While natural disasters like tsunamis will likely never be preventable, there is reason to hope that there will always be ways to ensure that people are prepared for and informed of what's to come.