Ozone levels in the earth’s troposphere (the lowest level of our atmosphere) can now be forecasted with accuracy up to two weeks in advance, a significant improvement over current systems, which can only forecast ozone levels three days in advance. The new artificial intelligence system could lead to better ways to control high ozone levels and even contribute to climate change solutions.
“This was extremely difficult. Nobody had ever done anything like this before. We are, I believe, the first to attempt to forecast surface ozone levels two weeks in advance” Yunsoo Choi, professor of atmospheric chemistry and AI deep learning at the University of Hawaii’s College of Natural Sciences and Mathematics, agreed. The findings have been published in the scientific journal Scientific Reports-Nature.
Ozone, a colorless gas, can be beneficial when used in the right place and amount. It protects the earth’s stratosphere (“the ozone layer”) by filtering out UV radiation from the sun. However, high concentrations of ozone near the earth’s surface are toxic to the lungs and hearts.
Ozone is a secondary pollutant that can harm humans. Exposure can cause throat irritation, breathing difficulties, asthma, and even respiratory damage. Some people are more vulnerable than others, including the very young, the elderly, and the chronically ill. Applying deep learning to air quality and weather forecasting is like looking for the holy grail, just like in the movies.
Alqamah Sayeed
“Ozone is a secondary pollutant that can harm humans,” explained doctoral student Alqamah Sayeed, a researcher in Choi’s lab and the study’s first author. Exposure can cause throat irritation, breathing difficulties, asthma, and even respiratory damage. Some people are more vulnerable than others, including the very young, the elderly, and the chronically ill.
Ozone levels are now frequently mentioned in daily weather reports. However, unlike weather forecasts, which can be reasonably accurate up to 14 days in advance, ozone levels have only been predicted two or three days in advance – until now.
This new study’s vast improvement in forecasting is only one part of the story. The other is how the team pulled it off. A numerical model is used in conventional forecasting, which means that the research is based on equations for the movement of gases and fluids in the atmosphere.
Choi and his team were well aware of their limitations. Because the numerical process is slow, obtaining results is costly, and accuracy is limited. “After the first three days, the accuracy of the numerical model begins to decline,” Choi explained.
The machine learning algorithm was created by the research team using a one-of-a-kind loss function. A loss function aids in AI model optimization by mapping decisions to their associated costs. Over conventional loss functions, researchers in this project used index of agreement, or IOA, as the loss function for the AI model. IOA is a mathematical comparison of the differences between what is expected and what actually occurs.
In other words, team members gradually refined the program’s reactions by incorporating historical ozone data into the trials. The combination of the numerical model and the IOA as the loss function eventually allowed the AI algorithm to accurately predict the outcomes of real-life ozone conditions by recognizing what had happened in similar situations in the past. It is similar to how human memory is constructed.
“Consider a young boy who notices a cup of hot tea on a table and reaches out to touch it out of curiosity. The child realizes the cup is hot and should not be touched directly the moment he touches it. The child’s mind has been trained as a result of that experience “Sayeed stated. “In a very basic sense, AI is the same. You provide input, and the computer outputs. The process is refined over time through many repetitions and corrections, and the AI program comes to ‘know’ how to react to previously presented conditions. On a fundamental level, artificial intelligence develops in the same way that a child learns not to rush to get the next cup of hot tea.”
The team used four to five years of ozone data in the lab in what Sayeed described as a “evolving process” of teaching the AI system to recognize ozone conditions and estimate forecasts, improving over time.
“Applying deep learning to air quality and weather forecasting is like looking for the holy grail, just like in the movies,” said Choi, who enjoys action movies. “For a few years, we had some difficult times in the lab. There is a procedure in place. We’ve finally gotten our hands on the Holy Grail. This system is functional. The AI model ‘understands’ forecasting. Despite the years of effort, it still comes as a surprise to me, even today.”
Before laboratory success can lead to real-world service, many commercial steps must be completed before the rest of the world can benefit from the discovery. “If you can predict the future – in this case, air quality – you can do a lot for the community. This could be very important for the planet. Who can say? Maybe we can figure out how to solve the climate change problem. Weather and ozone forecasting may be obsolete in the future. This could contribute to the planet’s security “Choi stated.
