Rare El Niño Events: How to Multiply a Storm ✛
Context
Extreme El Niño events, such as the 1997–1998 event, cause droughts, floods, and coral reef die-offs worldwide. Their frequency is estimated to double in the 21st century, but the precise distribution of the strongest events remains unclear because direct simulation requires hundreds of thousands of years of simulation. This work fits into the line of research on rare-event sampling in climate models: nearby studies have applied diffusion Monte Carlo methods to heatwaves and tropical cyclones, and adaptive multilevel splitting to simple models. This paper is the first to systematically test the TEAMS algorithm on an intermediate El Niño model where comparison with a long reference run is possible.
Methods
We used the Zebiak-Cane model, an intermediate ENSO model with an atmospheric component that responds to heating and a shallow-water ocean component. We ran an ensemble of 16 ancestors, each 25 years long, selected from a short direct simulation. At each step, we selected the 25% of members with the lowest maximum NINO3 index, and from the rest we generated offspring with small random perturbations to sea surface temperature (amplitude 0.1°C) applied 7.5 years before the threshold crossing. We repeated 50 independent runs with a budget of 2000 model years each. Return times were estimated using the block maxima method with a Poisson approximation, and uncertainty was assessed via bootstrap with 5000 samples.
Results
TEAMS reproduced the return-time estimates of the direct simulation up to 60,000 years at a cost of 100,000 model years versus 500,000 for the reference. The block maxima distribution, weighted by branching probabilities, matched the reference well, especially in the tail. The NINO3 threshold grew rapidly in the first generations but then plateaued around 4.2°C, suggesting a possible internal limit to extremity in the model. The success rate of generating offspring that exceed the threshold was about 77%. Trajectory analysis showed that a small perturbation can first trigger La Niña and then amplify the subsequent El Niño through non-normal growth of perturbations.
Implications
This work shows that rare-event sampling can provide statistically reliable estimates of the tail of the El Niño distribution at significantly lower cost than direct simulation. Unlike neighboring studies that applied methods to simpler or different systems, here we achieve for the first time quantitative agreement with the reference for an intermediate-complexity ENSO model. This strengthens confidence in the transferability of the approach to full climate models, where direct simulations of the required length are impossible.
Future development
Optimization of TEAMS hyperparameters (rejection fraction, number of ancestors, lead time) is possible using Bayesian search or grid analysis. Studying non-stationary climates, for example under anthropogenic warming, where extreme El Niño statistics may change, is promising. It is also worth exploring alternative strategies for managing splitting times to avoid bias toward earlier peaks and to improve boosting efficiency.
Impact
The method is applicable to expensive global climate models for assessing risks of extreme El Niño.
Next steps
Apply TEAMS to a full climate model, for example one of the CMIP models, and compare with available long simulations.
Key open problems
In climatology, the shape of the tail of the extreme El Niño distribution and its sensitivity to climate change remain open questions. This work provides a tool for obtaining reliable estimates under limited computational resources, directly addressing the problem of quantifying rare climate risks.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors applied the rare-event sampling method TEAMS to the Zebiak-Cane model to estimate the return times of extreme El Niño events. Comparison with a benchmark direct simulation of 500,000 years showed that TEAMS reproduces the estimates at a cost of about 20% of the direct simulation's expense.
Where its strength is
This work is the first to systematically test the TEAMS algorithm on an intermediate ENSO model, achieving quantitative agreement with the benchmark for the tail of the distribution. This bolsters confidence in the method's transferability to full climate models.
Where it could go next
- It might be interesting to see how the rare-event sampling method applies to other rare climate phenomena, such as extreme seasons or tropical cyclones, where diffusion Monte Carlo methods have already been used.based on: Quantum sieve catches one-in-a-million glitches arXiv:2606.06316
- A related avenue is using neural networks to accelerate rare-event simulation; perhaps combining them with TEAMS could further reduce computational costs.based on: Tropical Blueprint: A Neural Network Learns from an Algebraic Skeleton arXiv:2607.00237
- It would be interesting to compare TEAMS with adaptive multilevel splitting, which has been applied to simple models, to understand which method is more efficient for models of intermediate complexity.based on: Quantum sieve catches one-in-a-million glitches arXiv:2606.06316
Worth trying
- Test TEAMS on other ENSO indices, for example, eastern or western types of events.
- Check the method's sensitivity to perturbation amplitude and the interval before threshold crossing.
- Use TEAMS to investigate the joint distribution of several extreme characteristics, such as amplitude and duration.
Why this may matter
Advancing rare-event sampling methods for climate models could improve risk assessments of extreme phenomena such as droughts and floods associated with El Niño, and aid in planning climate change adaptation.
Nearby works in our archive
- Quantum sieve catches one-in-a-million glitchessimilarity 0.47
- The Climate Swings of Ancient Marssimilarity 0.47
- The Sun Conducts Ice Agessimilarity 0.46
- How Ancient Weather Chronicles Came to Lifesimilarity 0.46
- Tropical Blueprint: A Neural Network Learns from an Algebraic Skeletonsimilarity 0.46
- Gravitational Visitors: How Passing Planets Rewrote Earth's Historysimilarity 0.45
The archive is sparse around this work: the closest review is only 0.47, against a usual 0.64. It looks like few have drilled nearby.
TEAMS' cost of 100,000 model years is equivalent to simulating climate from 1997 to about the year 102,000 CE, but achieved by branching only 50 short runs.
The idea of branching trajectories to explore rare outcomes echoes the multiverse in Ted Chiang's 'Story of Your Life', where different branches of the future are viewed simultaneously.
Key numbers
- reduction in computational cost: 5-fold (20% of DNS cost)
- reference DNS length: 500,000 model years
- total TEAMS cost: 100,000 model years
- NINO3 plateau threshold: about 4.2°C
- offspring success rate: 77%