Using AI and machine learning to kickstart climate change fightback
Advances in artificial intelligence (AI) and machine learning may increase the chances of reducing carbon emissions through carbon capture or geo engineering projects.
Ranveer Chandra, managing director of Research for Industry and chief technology officer (CTO) of Agri-food at Microsoft, has spent his career researching climate and agritech solutions. Microsoft, in this regard, is pursuing its own commitment to carbon reduction. Chandra says AI can help “full solution” geoengineering projects become more affordable, targeted and transparent.
With AI, researchers can better estimate locations for, as well as impacts of, ocean or solar geoengineering to decide on the most appropriate, effective approach. AI may also replace expensive simulations or augment process-based models – especially with new and promising technologies such as the Python-based Causal ML.
Tackling an existential crisis
Chandra, however, warns that huge challenges remain: “The first is the cost of these solutions, to be implemented at scale,” he tells IT Pro. “Second, there’s a limited understanding of the full impact – primary and secondary – of any of these geo engineered solutions.”
Carbon capture and storage (CCS) is typically seen as more promising, but it’s “still a very expensive way to remove carbon”, he adds. “With AI, we are able to do seismic modelling at scale, at a speed-up of more than 1,500 times existing approaches that use partial differential equations and simulations.”
It means better modelling of the flow of carbon and planning of CCS operations; Microsoft is working with a range of partners, including Nvidia, on developing AI and machine learning approaches that will power CCS projects.
Ongoing and planned projects could sequester a combined amount of approximately 40 megatonnes of CO2 per year. But to keep temperature rises at 1.5ºC versus pre-industrial levels, a hundred times the storage capacity is required, he warns. “This is an existential issue; therefore we have to help invent techniques to mitigate climate change,” offers Chandra.
Professor Ted Shepherd, Grantham chair of climate science at the University of Reading, notes physics-based simulated models have been typically more reliable in terms of working out cause and effect than pure data-based approaches.
“Data science, as it is usually understood, is good at finding efficient solutions in situations with lots of data to explore, but climate intervention strategies are, by their very nature, out of sample — that is, not represented in existing data,” Shepherd says. “Data science methods tend to fall over when applied to out-of-sample problems.”
Innovations in causal AI – which identifies underlying causes of a behaviour or event that predictive modelling fails to – can help lower the risk of unwanted outcomes by getting a much better picture of cause and effect relationships, Shepherd says.
Microsoft research projects driving CCS advances
Northern Lights: Working with Norway’s government, Equinor, Shell and Total to standardise and scale CCS, for a North Sea storage reservoir from 2024.
KarbonVision: Using a computer vision approach to mapping geological faults from seismic data, reducing processing time it takes to detect potential leakage pathways of CO2.
Q-FNOs for 3D flow: Developing industry-relevant, scalable 3D simulations for CO2 flows and storage, involving typically complex, high-compute coupled PDEs.
Redwood: Working towards clusterless supercomputing on Azure, by building a more easily managed distributed programming framework on top of existing Azure HPC services.
Hyperwavve: Using a cloud-native fault-tolerant framework for hyperscale 3D seismic imaging, with Docker, Kubernetes and Dask parallel containerised seismic workloads at scale on Azure.
Innovating our way out of a climate crisis
Beyond CCS, more ambitious concepts include stratospheric aerosol injection (SAI) and marine cloud brightening (MCB). The idea is to reflect sunlight back into space to reduce global warming – either by spraying reflective particles into the stratosphere or ‘seeding’ clouds with salt crystallised out of the oceans, respectively.
Dr Vitali Avagyan, data scientist at TurinTech, notes that AI could help predict failures of CCS plants in real-time, as well as helping compare different, complex decarbonisation strategies.
“Rapid growth of environmental data from sensors, weather and climate models makes it difficult to interpret at speed,” Avagyan says. “AI can help measure collective impacts of CCS on entire energy systems.”
Shepherd notes that SAI seems “very feasible”, but no one knows exactly how it would play out. What if, for example, an SAI implementation in one country caused, say, problems in another place? A failure of the South Asian monsoon would be a disaster, for example – suggesting governance challenges.
Dr Timothy Farewell, head of science at Dye & Durham, highlights the need for solid assessing, filtering and cleaning alongside a strong understanding of interactions and processes involved.
“Some blind AI or machine learning models will look to extrapolate beyond the range of training data to more extreme conditions – leading to serious issues with accuracy,” Farewell confirms.
Jim Haywood, professor of atmospheric science at the University of Exeter, also tells IT Pro that greater knowledge of physical science is still needed to manage the risks and opportunities of SAI and MCB especially.
Shepherd agrees, adding: “We don’t really have that much confidence in the regional aspects of climate change. You have lots of factors; you need a very structured way of doing it, and that’s where the AI comes in.”
- SEE MORE Smashing green tech myths
Also, applying even well-established laws of physics to an atmospheric simulation means breaking up an earth-wide system into grid boxes – yet the equations needed should be continuous in space, rather than discrete. A typical grid box for a climate model might be 50 square-km in the horizontal and 1km in the vertical, Shepherd says. “It’s pretty coarse. You’re not representing a lot of things,” he confirms. “There are uncertainties. That makes the model much, much more computationally expensive.”
Ideally, an “ensemble” of runs, of lots of processes at very high spatial resolution, is needed to work out all the possible realisations as well, says Shepherd. “And different scientists will argue differently, for different trade-offs, as well as there being some processes that are not fundamentally understood – such as mixed-phase clouds of ice and liquid.”