A.I. can help humans.
In my search to see where AI can improve our lives, I came across the “A.I. for social good” movement. It seeks to apply A.I. techniques to the benefit of 17 of the UN’s sustainable development goals.
During my research, I came across some great resources:
- Nature paper on Social Good here.
- Google’s list of social good projects that involve medical research, building rooflines, flood risk mapping and forest fire prediction. In 2021 Google was supporting another list of 30 projects.
- Other organizations such as McKenzie have written extensively about using AI for Social Good.
googles weather model
Among the projects is a newly announced weather model. Looking deeper at Google’s newly announced weather model we can get a glimpse at how it works and how it was built.
The model is impressive as its accuracy is better than a gold standard physics model, and runs within less than a minute compared to hours.
The model was built on top of a massive weather data set “four decades of weather reanalysis data, from the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 dataset.”
To make predictions it works similarly to numerical models where you input mix of the previous and current state of the weather and run the model forward.
From the article:
For inputs, GraphCast requires just two sets of data: the state of the weather 6 hours ago, and the current state of the weather. The model then predicts the weather 6 hours in the future. This process can then be rolled forward in 6-hour increments to provide state-of-the-art forecasts up to 10 days in advance.
The cost/speed reductions and accuracy increase of this model are what makes this model interesting. The weather is known to be a stocastic system and hard to predict. The opportunity of increasing the speed of the model is for many people to create future predictions on demand without an expensive high-performance computing environment and the difficulty of operating the numerical model. This makes the model much more equitable to use by all kinds of people, and organizations.
Looking at the provided code it does seem quite simple to get started.
You also only need to download the model weights once, and only two data points are needed to start making the prediction. One may assume that you would download the model once, and grab the data points when you need them.
The model can predict 6 hours every hour. This can be very important when it comes to extreme weather events prediction. It might be a challenge to get the input at this 1 hour time resolution, however, if possible this could improve the early warning systems as they don’t require specialized equipment.
Research is ongoing, however, and the weather is changing a lot with climate change which may change the underlying data assumptions. That being said this is one of those applications where previous data will give better results than the numerical models can.