Head upwind.

One of the areas where artificial intelligence really excels is in working out scenarios with a huge number of complex variables – like how radiation might spread after an accident at a nuclear power plant.

This is the focus of a new AI system developed in Japan, and it's showing us more accurately than ever before where the safest (and most dangerous) points could be following a meltdown. Spoiler: stay upwind.

While it's obviously better if nuclear plants don't fail in the first place, knowing which way the fallout will travel can be crucial in organising emergency responses and keeping people safe. It can quite literally save lives – and a lot of them.

The new AI, developed by a team from the Institute of Industrial Science at the University of Tokyo, is able to factor in accident variables and prevailing weather patterns to work out where the threat of radiation could be worst, up to 33 hours in advance.

"Our new tool was first trained using years of weather-related data to predict where radioactivity would be distributed if it were released from a particular point," says one of the team, Takao Yoshikane.

"In subsequent testing, it could predict the direction of dispersion with at least 85 percent accuracy, with this rising to 95 percent in winter when there are more predictable weather patterns."

You can see the model in action below:

That machine learning – using past data to calculate the most likely future outcomes – is crucial in maintaining a high level of accuracy. Existing fallout prediction systems are known to be limited in how much they can be relied upon.

Add in the unpredictability of the wind and it's not surprising that scientists have struggled to pin down how far and in what direction fallout will travel.

Armed with this new knowledge, evacuation procedures could be put in place more quickly and more efficiently. Getting masses of people out of an area is a big undertaking – and something authorities will only want to do when absolutely necessary.

"The fact that the accuracy of this approach did not decrease when predicting over 30 hours into the future is extremely important in disaster scenarios," says Yoshikane.

The new prediction model can provide useful information about which areas will be worst affected and need evacuating, and which areas have a lower risk – in these areas the residents might just get warnings about being careful what they eat and drink.

With the high temperatures associated with nuclear disaster, radioactive material can travel up to 2,000 metres (6,562 feet) into the air, the scientists report – reaching winds in the upper troposphere that can spread fallout all across the world.

At the lowest level, sea breezes and mountain valley winds can spread fallout locally. All these variables need to be accounted for to get a model that works.

The AI system was found to be least accurate at predicting fallout spread when applied to data collected in July, down to about 78 percent. The researchers put this down to the irregularity and unpredictability of typhoons during that month.

With a machine learning approach, though, the model should improve over time as it gets access to more training data.

If another Fukushima-type incident does happen, this technology will help us to be better prepared.

The research has been published in Scientific Reports.