ClimateAi Unveils Breakthrough 1km Resolution Climate Risk Forecasting using Physics-Informed Machine LearningCovered by AP News
Himanshu Gupta • June 27th, 2023.
The atmospheric pendulum has swung away from a three-year-long stubborn La Niña phenomenon, which brought its characteristic cooling to the Pacific Ocean. Now, its warm counterpart, El Niño, is here, which can bring even more dramatic shifts in weather patterns that range by location.
Governments from Peru to the Philippines are already preparing to spend billions of dollars to mitigate El Niño-related weather events. But companies also need plans in place to ensure their operations are resilient. Luckily, new technology can help.
Artificial intelligence can help mitigate the negative impacts of El Niño in three main ways.
For one, predictions of the timing of El Niño’s arrival with a solid lead time (over 9 months) are key for stakeholders. They need this time to get ready to take action to mitigate the adverse conditions.
Artificial intelligence can help us predict when El Niño will be here in a faster, cheaper way, with the same level of reliability (if not better) as physical models based on supercomputers. That offers companies and governments more resources and a wider window to prepare for shifts in global temperatures and other patterns.
The second way artificial intelligence can help mitigate the negative impacts of El Niño is to help companies understand its direct impacts on their supply chains. It creates actionable forecasts on business-specific metrics, such as crop yields for companies in the agri-food value chain. And because companies don’t need supercomputers to run these forecasts, they have quicker, less expensive access to this information.
For example, in Australia, the world’s second-largest wheat exporter, an El Niño can cause hot, dry conditions. These hurt the production of wheat. As a result, the country’s winter wheat crop, which is currently being grown, is expected to have yield declines (after a few years of a bumper crop thanks to La Niña-induced rainfall). A wheat yield forecast that accounts for El Niño’s impact is a key insight for agribusiness decision-makers.
The third way that artificial intelligence can help is by providing recommendations based on these forecasts.
Consider Australia again, both as a production site and as a node in supply chains. If Australian farmers had known to expect a dry year, they could have bought/planted drought-resistant wheat seeds to ensure profitability. Additionally, this lower wheat yield could have ripple effects across global supply chains, as it goes to buyers across Asia, including China, Indonesia, and Japan. If they had been aware ahead of time, they could have locked in sourcing contracts with wheat producers in other parts of the world with expected favorable wheat-growing conditions thanks to El Niño.
The technology involved in predicting El Niño has evolved and improved drastically over the past few decades. Artificial intelligence can help even more, and not only with ENSO. It has uses for other slow-moving climate events that we call “low frequency” events, such as long-term drought.
For these types of events, there is not as much historical data. Historical observational data doesn’t go back very far, as reliable satellite technology is only a few decades old. That makes low-frequency events harder to predict for any model. This is one of the big issues in forecasting slow-moving climate events such as El Niño.
Artificial intelligence has an advantage here. It can run simulated data — simulating centuries of ENSOs to get more data to learn about these types of events — then transfer these learnings onto its prediction capabilities. Similar to buzzy new-age Generative AI models like ChatGPT, AI-based weather and climate forecasting models train on simulated climate data (in the absence of historical observational data) and transfer these learnings onto observational data.
That can help capture some of the physical relationships and determine the more fundamental forces that contribute to the development of such low-frequency events with long tails and limited historical datasets.
Also, algorithms can be trained on many different kinds of models through deep learning. The algorithms can thus take into account more variables and don’t make as many assumptions about linear relationships between them. (The science for some climate events like ENSO is still unclear about some of these interactions and mechanisms.)
ClimateAi achieved this breakthrough at Stanford University four years ago, and a few research groups have also followed since then.