Machine Learning and Climate Tipping Points: Predicting the Unpredictable

Several large parts of the climate system, from ice sheets to ocean currents, can flip into a different state once warming pushes them past a threshold. A 2022 assessment in Science concluded that five of these tipping elements are already "possible" at roughly 1.1°C of warming above pre-industrial levels. Machine learning is now used in two ways to see such shifts coming: finding faint statistical warning signs in noisy records, and running climate models fast enough to explore the risks. Both approaches have real strengths, and both run into hard limits.

Key takeaways

• Near a tipping point, a system recovers more slowly from small disturbances. This "critical slowing down" can show up in data as rising autocorrelation and variance.
• A 2021 deep learning method was more sensitive and specific than those classical indicators across 268 empirical and model time series, but it has never been validated on a tipping point that has not happened yet.
• For the Atlantic Meridional Overturning Circulation (AMOC), the same kind of analysis has produced alarming results and sharp rebuttals, and a 2024 paper argues the uncertainties are too large to predict when it might tip.
• Machine learning emulators can run weather and climate simulations at a small fraction of the usual computing cost, but a leading example does not extrapolate to substantially different future climates, which is exactly where tipping points live.

What is a climate tipping point?

A tipping point is a threshold beyond which a change in part of the climate system becomes self-sustaining, with large consequences for the wider Earth system. Armstrong McKay and colleagues, writing in Science in 2022, describe it in those terms and list nine global "core" tipping elements and seven regional "impact" elements. The phrase "tipping element" itself was introduced to climate science in a 2008 PNAS paper by Timothy Lenton and colleagues, which also compiled the first short list of candidates and asked experts to rank them.

The underlying idea comes from dynamical systems theory. Picture a marble resting in a bowl. Nudge it and it rolls back. Now slowly reshape the bowl so one side flattens and then disappears. For a long time nothing seems to change, and then the marble drops out of the bowl and rolls to a new resting place. Mathematicians call the moment the bowl loses its edge a bifurcation. The marble's drop looks sudden, even though the reshaping was gradual.

Examples of candidate tipping elements include the Greenland and West Antarctic ice sheets, tropical coral reefs, regions of abrupt permafrost thaw, the Amazon rainforest, and ocean circulation patterns such as the AMOC and convection in the Labrador-Irminger Sea. The 2022 assessment sorted them by the warming at which they become possible or likely.

Tipping element	Status per the 2022 assessment
Greenland ice sheet, West Antarctic ice sheet	Possible at about 1.1°C
Tropical coral reefs	Possible at about 1.1°C
Abrupt permafrost thaw	Possible at about 1.1°C
Labrador-Irminger Sea convection	Possible at about 1.1°C; likely within 1.5 to under 2°C
AMOC, Barents Sea ice, boreal forest	Possible by 1.5°C
Mountain glaciers	Possible by 1.5°C; likely at 2 to 3°C under current policies
Amazon rainforest	Not likely before about 3.5°C from climate drivers alone; deforestation could bring it sooner

Source: Climate Tipping Points explainer of Armstrong McKay et al. (2022). "Possible" means the warming level is above the minimum estimate of the threshold; "likely" means it is above the central estimate.

Three details are worth holding onto. First, thresholds are ranges, not lines on a map, and the ranges are wide. Second, the paper's authors note that today's warming already sits at the lower end of some of those ranges. Third, a system can tip in more than one way. Peter Ashwin and colleagues distinguish tipping caused by a bifurcation, tipping caused by random noise kicking the system out of its current state, and rate-dependent tipping, in which a fast enough change pushes a system off course without any bifurcation at all (Philosophical Transactions of the Royal Society A, 2012). That distinction matters later, because the best-known warning signal is tied to only the first kind.

Critical slowing down: the warning signal everyone looks for

As a system nears a bifurcation, its restoring force weakens. Go back to the marble: as the bowl flattens, a nudge takes longer to die away. Mathematically, the rate at which the system returns to equilibrium approaches zero. This is called critical slowing down.

It leaves fingerprints in time series. Because the system is sluggish, each measurement resembles the previous one more strongly, so lag-1 autocorrelation (the correlation between consecutive observations) rises. Because disturbances linger and pile up, variance rises too. Marten Scheffer and colleagues made the case in a widely cited 2009 Nature review that generic early-warning signals of this kind may exist across ecosystems, financial markets, and the climate, even though forecasting tipping points is very difficult (Nature 461: 53-59).

The idea is attractive for a practical reason: you do not need a full physical model. You only need a long, consistent record of one variable. The trouble is that real climate records are anything but clean.

• Noise. Year-to-year variability, measurement error, and long-term trends can all mimic or mask a rising autocorrelation.
• Preprocessing choices. Detrending a series, choosing a window length, and filling gaps in the data can each change the answer.
• Changing observation networks. If more instruments came online over time, variance can rise for reasons that have nothing to do with the climate.
• Only one kind of tipping. Noise-induced and rate-dependent tipping may not announce themselves with critical slowing down at all.

These are not minor caveats. They are the main reason classical early-warning tests so often return results that are only marginally significant, and why researchers turned to machine learning.

How deep learning reads the warning

The most influential machine learning approach came from Thomas Bury, Chris Bauch, and colleagues, published in PNAS in 2021. Their paper, "Deep learning for early warning signals of tipping points," trains a neural network (a model that learns patterns from examples, see how neural networks work) not on climate data but on simulations. According to the University of Waterloo, the researchers built a "universe of possible tipping points" of some 500,000 models, so the network would learn the shared features of tipping in general rather than the quirks of any one system.

Think of a medical student who studies thousands of textbook cases of a disease before ever meeting a real patient. The student learns the pattern from the textbook cases and then recognizes it in a patient who looks a bit different. The network does the same with time series. It looks at a window of observations and judges whether the system is drifting toward a bifurcation.

The reported results are specific. Tested on 268 empirical and model time series from ecology, thermoacoustics, climatology, and epidemiology, the algorithm was more sensitive and more specific than generic indicators such as lag-1 autocorrelation and variance (Europe PMC record of Bury et al., 2021). A second ability is unusual: it can estimate the normal form of the approaching bifurcation, which offers clues about what kind of state lies on the other side of the tipping point. Classical indicators say that something may be approaching, but nothing about what comes next.

Two cautions follow. The first is the sim-to-real gap. A network trained on idealized models can be fooled by features of real data that the models never produced, which is why the validation on real transitions matters. The second is more basic. Every successful test so far is retrospective, on transitions that already happened. University of Waterloo coverage described feeding the algorithm current climate data as the team's next step. Nobody can run a controlled test on a tipping point that has not happened, so predictions about the present are always a matter of degree.

For readers who want the general background on this kind of forecasting, how time series forecasting works covers the basics, and overfitting and underfitting explains why a model that fits its training examples can still fail on new ones.

The AMOC: a case study in contested signals

The Atlantic Meridional Overturning Circulation is a system of ocean currents that carries warm surface water northward in the Atlantic, where it cools, sinks, and returns south at depth. It shapes the climate of Western Europe and the tropical rain belts. What would happen if it stalled is explored in What If the North Atlantic Current Stopped?.

The AMOC has been the proving ground for early-warning analysis, and the story shows how far a signal is from a verdict. The table lays out four key papers and a rebuttal.

Study	Approach	Finding	Caveat or response
Boers, 2021, Nature Climate Change	Early-warning indicators on eight AMOC indices from sea-surface temperature and salinity data	Reported significant early-warning signals across the indices; the system may have moved from stable conditions to near a critical transition	Chen and Tung argue the variance rise is an artifact of expanding observational coverage and that two proxies have reversed signs
Ditlevsen and Ditlevsen, 2023, Nature Communications	Statistical estimators of the timing of tipping	Collapse around mid-century, or potentially any time from 2025 onward, under continued emissions	Contrasts with IPCC assessments based on climate models; timing is contested
van Westen, Kliphuis, and Dijkstra, 2024, Science Advances	Simulated AMOC tipping in a full climate model, then looked for an observable signal	Proposed a physics-based indicator, the minimum of AMOC-induced freshwater transport at the Atlantic's southern boundary; reanalysis suggests the AMOC is heading toward tipping	Found that classical statistical indicators are sensitive to the time interval analyzed before tipping
Ben-Yami, Morr, Bathiany, and Boers, 2024, Science Advances	Critique of extrapolating from historical data	Uncertainties are too large to reliably estimate when a component will tip, even if it is approaching a tipping point	Illustrated with the AMOC; raised concerns about data coverage, gap filling, and modeling assumptions

The last row deserves emphasis because one of the authors, Niklas Boers, also wrote the 2021 paper in the first row. A scientist revisiting his own method and concluding that its timing claims are too uncertain is the process working as intended.

Behind the argument is a short instrument record. Continuous direct monitoring of the AMOC only began in 2004, according to the World Climate Research Programme's summary of the Ditlevsen study, so most of the long history has to be reconstructed from proxies such as sea-surface temperature patterns. Different reconstructions can tell different stories.

The IPCC's Sixth Assessment Report took a cautious position: decline of the AMOC over the 21st century is very likely, but there is only medium confidence that it will not collapse abruptly before 2100. Medium confidence is not reassurance. It is an honest statement that models and observations disagree enough to leave the question open.

Can machine learning tell us when the AMOC will tip?

Not with the evidence available. Machine learning can flag a pattern consistent with an approaching transition, and a trained network may be better at it than a single autocorrelation statistic. The 2024 critique in Science Advances argues, though, that errors in the underlying data and assumptions propagate into any timing estimate. A better classifier does not fix a record that is too short and too uneven.

The Amazon: when loss of resilience becomes a measurement

The Amazon rainforest has been another test. Chris Boulton, Timothy Lenton, and Niklas Boers applied established early-warning indicators to satellite vegetation data from 1991 to 2016 and reported that the forest's resilience has declined since the early 2000s, a pattern consistent with an approach to a critical transition. The decline was faster in drier regions and in areas nearer to human activity (Boulton, Lenton, and Boers, 2022, Nature Climate Change; the preprint reports the decline beginning around 2003).

This example matters for how the field is evolving. The signal is a statement about resilience, which is measurable, rather than a prediction of a date for dieback. It also shows why the Amazon is a different kind of risk from the ice sheets: the 2022 assessment judged that climate drivers alone are unlikely to push the forest over before about 3.5°C, but that deforestation could bring the threshold closer. Human land use changes the answer.

Machine learning emulators: running climate models faster

A second use of machine learning has nothing to do with early-warning statistics. It is about speed. Full Earth system models are physically detailed and expensive; running them at the resolution needed for regional questions ties up supercomputers for weeks. Emulators trained on past model output or observations try to reproduce the key behavior at a much smaller cost. Three systems illustrate the range:

System	Developers and year	What it does	Verified claims and limits
GraphCast	Lam et al., Science, 2023	Predicts hundreds of weather variables for 10 days at 0.25° resolution	Runs in under one minute; outperformed the most accurate operational deterministic system on 90% of 1,380 verification targets
NeuralGCM	Kochkov et al., Nature, 2024	Combines a differentiable solver of atmospheric dynamics with machine learning components	Competitive with machine learning models for 1 to 10 day forecasts and with the ECMWF ensemble for 1 to 15 days; with prescribed sea-surface temperature, tracks climate metrics over multiple decades at 140 km resolution; authors report "orders of magnitude" computational savings; does not extrapolate to substantially different future climates
ClimaX	Nguyen, Brandstetter, Kapoor, Gupta, and Grover, 2023	A general-purpose transformer pretrained on CMIP6-derived climate data and fine-tuned for forecasting and projection tasks	Authors report better results than existing data-driven baselines on weather and climate benchmarks

The practical value is the number of runs you can afford. Climate scientists characterize uncertainty by running many simulations with slightly different starting conditions or settings. A cheaper model allows more of them, which gives a better picture of rare extremes such as intense heat waves or heavy rainfall.

The limit is just as important, and the NeuralGCM authors state it plainly: the model does not extrapolate to substantially different future climates. That is the central tension for tipping-point work. An emulator learns from the past. A tipping point, by definition, takes the system somewhere it has not been. Hybrid designs that keep the physical laws of fluid motion in the loop, and approaches such as physics-informed neural networks, are attempts to narrow this gap by tying a flexible network to equations that remain valid outside the training data.

The chaotic nature of weather adds its own ceiling. Tiny differences in starting conditions grow over days, which is the logic of the butterfly effect, and turbulent flows resist exact prediction (why turbulence is hard to predict). Machine learning speeds up the calculation without removing the underlying unpredictability.

Why predicting a tipping point stays hard

Pulling the threads together, a handful of obstacles recur.

1. No ground truth for the future. Methods can be tested on past transitions, such as those recorded in paleoclimate proxies. They cannot be validated on a transition that is yet to occur, so any forecast is a probability statement.
2. Short and uneven records. Direct observation of key systems is often a few decades long, and proxy reconstructions carry their own uncertainties.
3. Different ways to tip. A warning signal tuned to bifurcations may miss noise-induced or rate-dependent transitions.
4. Moving targets. Thresholds depend on what else is happening, from land use in the Amazon to freshwater input in the North Atlantic.
5. Interactions. Tipping elements can influence each other, and no single study captures all the connections.

A sensible reading of the evidence is that early-warning methods, including machine learning ones, are useful for prioritizing where to look, where to put instruments, and which questions to ask of climate models. They are poor tools for putting a date on an event.

For related examples of how machine learning meets the limits of forecasting hazards, see AI and disaster risk assessment and the case of AI in seismology, where prediction remains out of reach. For the basic physics behind the warming that pushes systems toward these thresholds, see climate change science: how it actually works.

What the 2025 tipping points report added

The Global Tipping Points Report 2025, released by the University of Exeter, the Potsdam Institute for Climate Impact Research, and WWF with international partners, shifted the conversation from risk to a first reported crossing. According to the National Oceanography Centre's summary, it concluded that warm-water coral reefs have passed their thermal tipping point, with a central threshold estimate of 1.2°C (range 1 to 1.5°C) against current warming of about 1.4°C, and that widespread mortality is underway. It also flagged ice sheets, ocean currents, and the Amazon as approaching thresholds.

The same report counted the positive side. It judged that solar power and wind power have already crossed positive tipping points globally, as have electric vehicles, battery storage, and heat pumps in leading markets. These are tipping points in the economic sense: falling costs drive adoption, which drives scale, which drives further cost reductions. Whether such social and economic dynamics can be modeled with the same machine learning tools is an open research question, and no model discussed in this article claims to do it.

For what a large ice loss would mean at the extreme, see what if Antarctica completely melted and how glaciers form and move.

FAQ

Q: How does the Paris Agreement relate to tipping points? A: The Paris Agreement aims to limit warming to well below 2°C and to pursue 1.5°C. The 2022 assessment by Armstrong McKay and colleagues found that several tipping points may be triggered at 1.5 to under 2°C and many more become likely at 2 to 3°C, the range expected under current policies. Threshold estimates are wide, so the staying-below-1.5°C case rests on risk rather than on a precise line.

Q: How reliable are neural network early-warning systems? A: They have been tested only on past transitions. On 268 empirical and model time series, the 2021 PNAS method beat lag-1 autocorrelation and variance for sensitivity and specificity. For present-day systems, no validation is possible until after the fact, so outputs should be read as probabilities and compared against physical understanding.

Q: What is the difference between a weather model and a climate model? A: A weather model predicts the state of the atmosphere at a specific time, for days to a couple of weeks ahead. A climate model estimates the statistics of weather, such as average temperatures and the frequency of extremes, over decades under a given emissions scenario. Machine learning has had its clearest success so far in weather forecasting, as with GraphCast, and is moving into climate emulation.

Q: Can climate tipping points be reversed once triggered? A: Tipping is defined as self-sustaining change, so recovery usually requires much more than removing the original push. Reversibility differs by element and is hard to assess in advance. Reducing warming is still the main way to lower the chances of crossing thresholds in the first place.

Q: Are there positive tipping points too? A: Yes. The Global Tipping Points Report 2025 says solar and wind power have crossed positive tipping points globally, along with electric vehicles, battery storage, and heat pumps in leading markets. The mechanism is a feedback loop in which falling costs boost adoption, which in turn lowers costs.

Q: Is the AMOC about to collapse? A: Scientists disagree. One 2023 analysis estimated collapse around mid-century or potentially at any time from 2025 under continued emissions, while the IPCC expressed medium confidence that an abrupt collapse will not happen before 2100. A 2024 paper argued that the uncertainties are too large to estimate when it could tip.

The takeaway

Machine learning has given tipping-point science sharper instruments, not a clairvoyant. Neural networks trained on simulated bifurcations can spot the faint slowing-down signature that older statistics barely catch, and fast emulators let researchers explore many more futures than supercomputers once allowed. But the central problem remains: a tipping point has never been tested in advance, and the records we have are short, patchy, and open to dispute, as the AMOC debate shows. The most defensible use of these tools is to find out where the risks are rising and where better observation is needed, while treating any specific date as an open question.

Sources

• Armstrong McKay, D.I., et al. (2022), "Exceeding 1.5°C global warming could trigger multiple climate tipping points," Science 377(6611) (DOI; explainer).
• Bury, T.M., et al. (2021), "Deep learning for early warning signals of tipping points," PNAS 118(39) (DOI; University of Waterloo news).
• Boers, N. (2021), Nature Climate Change 11: 680-688 (DOI); Chen, C. and Tung, K.-K., Matters Arising (arXiv:2312.13568).
• Ditlevsen, P. and Ditlevsen, S. (2023), Nature Communications 14: 4254 (DOI; WCRP summary).
• van Westen, R.M., Kliphuis, M., and Dijkstra, H.A. (2024), Science Advances 10(6) (DOI); Ben-Yami, M., et al. (2024), Science Advances 10(31) (DOI).
• Kochkov, D., et al. (2024), "Neural general circulation models for weather and climate," Nature 632 (arXiv:2311.07222); Lam, R., et al. (2023), Science 382 (DOI).
• Lenton, T.M., et al. (2008), PNAS 105(6): 1786-1793; Scheffer, M., et al. (2009), Nature 461: 53-59; Ashwin, P., et al. (2012), Philosophical Transactions of the Royal Society A 370: 1166-1184 (arXiv:1103.0169).
• National Oceanography Centre, Global Tipping Points Report 2025 summary.

---

Read next:

• How Machine Learning Is Accelerating Materials Discovery
• Deep Learning and Microscopy: Seeing Materials at the Atomic Scale
• AI at CERN: How Machine Learning Found the Higgs Boson and What Comes Next