Skip to content
Resources > Latest News > How ai is changing climate modelling for water planning

How AI is changing climate modelling for water planning

As climate extremes accelerate and uncertainty becomes the norm, Australia’s water community is under growing pressure to anticipate what’s coming and make decisions that hold up under stress. For decades, climate and hydrological modelling has relied on physics-based numerical models to guide those choices – powerful and credible, but slow and limited in the number of futures they can explore.

Artificial intelligence is now reshaping that landscape. By rapidly processing vast datasets and learning complex patterns in climate and hydrological systems, AI is dramatically improving the speed, resolution and flexibility of forecasts. Its real value, however, is not perfect prediction, but helping water managers test a wider range of plausible futures, understand risk and plan with greater confidence.

Bureau of Meteorology Senior Research Scientist Catherine de Burgh-Day said the pace of change in AI-driven forecasting has surprised even experts in the field.

“We’ve gone from a place where most experts would have told you it simply wasn’t possible for an AI model to predict the weather like a traditional model,” she said.

“Now, AI weather models are being run in real time alongside traditional models in a growing number of weather agencies around the world.”

Faster modelling

Dr Sanaa Hobeichi from UNSW’s Climate Change Research Centre said modern AI models are trained on extensive archives of observations and simulations to learn how climate and hydrological systems behave. This allows them to act as extremely fast surrogates for the traditional numerical models.

AI can reproduce many of the essential physical behaviours of those models, but at a fraction of the computational cost. And that has lots of benefits, Hobeichi said

“For example, in terms of speed, what used to take hours on a supercomputer when using the hydrological and climate model can now be approximated with AI in seconds,” she said.

“AI models are very cheap to run, which means we can push towards much finer spatial and temporal resolution. We can do things at the scale of individual catchments or even individual assets, and that wasn’t possible with the traditional models.”

Hobeichi highlighted transfer learning as one of the more important developments, where AI learns from well-observed regions or catchments and applies this knowledge to a new catchment where data is limited.

“We also have multimodal AI, which can bring together lots of different data. For example, it can bring together observations, satellite data, forecasts and even text-based knowledge, which is especially useful for early warning systems because sometimes the emerging risks show up as weak signals scattered across multiple sources,” she said.

“Now AI is increasingly used to capture those and give us the early warning that we need, and that’s becoming particularly important now with more severe and unpredictable weather coming through. Or maybe not so unpredictable now, if AI is beginning to do it.”

De Burgh-Day said another benefit is in improving not just the resolution spatially, but also the resolution in terms of the distribution of possible outcomes.

“We can run an ensemble forecast to get a probabilistic outcome to help make risk-based decisions, then have 10 ensemble members, which is 10 times the cost of a model run,” she said.

“Alternatively, we could train an AI model to emulate another model, instead of trying to predict reality. The idea is to run the main model that’s well known just two or three times and fill the space in between with the AI emulator model.

“We can build these really welldefined probability distributions, which enable us to produce much better risk-based forecast products and risk-based decision-making tools. It’s another way in which we can use AI without replacing the traditional model.”

Learning from data

Group Leader of the Environmental Infomatics Group at CSIRO Dave Penton said some deep artificial neural networks are reportedly performing better than alternative approaches in predicting stream flow, estimating streamflow in ungauged catchments and predicting water quality parameters.

For example, CSIRO found improved performance for forecasting high-frequency dissolved oxygen in Australian rivers. The speed of these networks is an obvious benefit: detailed flood models take a long time to train and days to run, but once an emulator is built, it can produce scenarios in real-time.

Penton said the other benefit is the opportunity to explore why the data-driven model gives better results than the process and conceptual model.

“For some of those, there’s some catchment processes that might not be captured in your existing understanding. It helps you understand what will happen because ultimately, what you are trying to do is build models which are going to be robust for an unseen climate.

“Take the relationship between rainfall and run-off for example, which has decreased in some Australian regions over time.

Capturing those things, which are ultimately changes in the way that the catchment will respond to rainfall, is something that we need in our models, whether they are purely data driven or process-based models.”

No silver bullet

De Burgh-Day said there are some in the forecasting community who think AI will never be able to replace a climate model because, almost by definition, there’s no data to train it on that’s representative of what it’s trying to predict.

“But I’m always a little hesitant to say never, because if you asked almost any scientist in 2019 whether they thought AI models could be a replacement for a weather model, they would have said definitely not,” she said.

“There is research going on at the moment where groups are trying to teach the model about the relationship between CO2 and other variables so it can extrapolate into the future, but no one has done that in a way that comes close to being competitive with a traditional climate model yet.”

The debate is around the extent to which these models have learnt the physics, versus just learning the processes and the relationships, which is an important factor in how well they could be used to predict unseen events – scenarios outside their training set.

National Manager for Climate Services at the Bureau of Meteorology Karl Braganza said one of the hardest challenges for both traditional models and AI is forecasting at seasonal timescales – the space between short-term weather forecasts and long-term climate projections.

After about seven to 10 days, he said, forecasts are no longer strongly guided by current conditions and instead move into a different physical regime that is harder to predict.

“It’s actually more challenging than weather forecasting, which is very dependent on how things are right now, and more challenging than climate change projections,” Braganza said.

“That intermediate timescale is still very much an open question in terms of what machine learning can deliver.”

To effectively predict longer than about a month, other parts of the Earth system, in particular the ocean, need to be included.

“We’re getting there – we will have full earth system in our models in the future, and that will be a big boost to how well they go for sub seasonal and seasonal timescales. But we’re not there yet,” Braganza said.

Making it relevant

Useful climate science for the water sector is providing this risk-based information at useful spatial and temporal scales, Hobeichi said, and explaining uncertainties in a way that decision makers can work with and communicate.

“AI supports the provision of high-resolution data and it makes it possible to produce large ensembles and capture a wide range of uncertainty; this can help bridge the gap between climate and the impact. AI can really help make climate science more useful to the water sector,” she said.

Decision-making by human experts will be key to realising the potential of AI, especially framing the problem at a spatial and temporal scale.

“We have lots of data, but some data is good at some particular aspects, other data is better in other aspects. So really making decisions around the data, accounting for the biases before we give AI model the data, that’s also important,” said Hobeichi.

“AI is just a tool, and it’s a very powerful tool, but we can only get the best outcome if we bring together physics-based modelling, high quality observations, impact data, and advanced AI methods, plus definitely experienced human judgment.”

Penton agrees that people are at the centre of all decisions, from deciding whether to drive through flood waters, whether to invest in a career in agriculture, or to live on a flood plain.

“The whole process has to be wrapped up with engagement with people who are involved in the outcomes of the work. And that process takes quite a lot of engagement and investment.

We’re seeing the results on the Northern Rivers and in the Murray Darling Basin through the sustainable yields and a whole range of different basin planning.”

Climate science and hydrology experts agree: the water sector will benefit from the faster modelling, richer scenarios, better informed decisions, and increased accessibility of complex information that AI will bring. However, it will work best with physical models, not in isolation, and it does not remove the need for human expertise.

“Human judgement will be essential when using AI models”, de Burgh-Day said. “At the end of the day, a human needs to decide what to do with that information, and how to balance the risk and reward of any decisions made based on a forecast, whether it came from an AI model or not.”

This article was originally published in Current magazine. Take a look at the whole edition on Indigo.