The AMOC Collapse Debate: What CMIP6 Models *Actually* Tell Us About Resilience

The Atlantic Meridional Overturning Circulation (AMOC) is one of the most critical components of Earth's climate system, redistributing vast amounts of heat from the tropics to the Northern Hemisphere. Recently, a major scientific debate has erupted over its future:

• The "No Collapse" View: A recent Nature paper by Baker et al. (2025) argues that the AMOC "does not collapse" under extreme warming because a wind-driven remnant survives.
• The "Yes Collapse" View: Top climate scientists like Stefan Rahmstorf argue this is purely a semantic debate - the thermohaline (density-driven) component of the AMOC does collapse, bringing catastrophic cooling to the North Atlantic, regardless of what we call the remaining circulation.

To cut through the semantics, we need mechanistic evidence. Which models predict a "Baker-like" wind-dominated remnant, and which predict a "Rahmstorf-like" thermohaline collapse? And more importantly: can we predict a priori how a model will behave under extreme warming?

We analyzed an initial exploratory subset of 5 to 10 models from the CMIP6 ensemble to find out. Here is what we discovered.

---

Methodology: Parsing the Mathematical Components of the AMOC

The AMOC is not a single, uniform conveyor belt. In classical theory (e.g., Gnanadesikan 1999), it is driven by two overarching physical mechanisms:

1. The Thermohaline Component (T_n): Driven by cold, dense water sinking in the North Atlantic. This is the part vulnerable to freshwater influx (melting ice sheets) and has a known "tipping point" represented as a saddle-node bifurcation.
2. The Wind-Driven Component (T_ek): Driven by the powerful westerly winds over the Southern Ocean. The wind stress creates Ekman transport that pulls water northward across the equator to feed the Atlantic overturning. This component acts as a "floor" - it will persist for as long as the winds blow.

We ran a two-phase analysis on CMIP6 models using Google Cloud computing resources:

Phase 1: The Historical Baseline

For the historical period (2000-2010), we extracted:

• Total Atlantic overturning strength (msftmz at 26.5N) in Sverdrups (Sv).
• Southern Ocean zonal wind stress (tauuo), which we used to calculate the Ekman transport T_ek feeding the Atlantic.
• The Resilience Ratio (T_ek / T_n): A dimensionless measure of how "wind-dominated" a model's AMOC is prior to any extreme warming. This metric is the intellectual heart of our analysis. Why? Because if we can prove that a model's initial, pre-warming state (its Resilience Ratio) mathematically dictates how it will behave centuries later under extreme stress, we achieve something profound. We transform a messy, semantic debate ("does it collapse or not?") into a rigorous phase-space classification problem. Instead of arguing over definitions, we can simply categorize models by their starting conditions to understand their inevitable endpoint.

Phase 2: The Crash Test (abrupt-4xCO2)

We then subjected the models to the standard abrupt-4xCO2 experiment, where atmospheric CO2 is instantly quadrupled and run for 150 years. We measured the AMOC response in the "late" period (years 130-150) to observe the asymptotic behavior of the circulation.

---

The Results

1. The AMOC weakens, but doesn't "break" (in 150 years)

Across the 5 models that completed the full pipeline, total AMOC declined by 18% to 36%. Under 4xCO2, the AMOC bends but does not fully halt within 150 years. This aligns with the IPCC consensus that a full collapse is unlikely before the year 2100.

However, examining the total AMOC mathematically masks the critical bifurcating behavior taking place beneath the surface.

2. The Thermohaline Component Plummets

While the total AMOC dropped by ~30%, the isolated thermohaline component (T_n) plummeted by 28% to 64%. Meanwhile, the wind-driven Ekman transport stayed relatively stable, acting exactly as the "floor" theorized by Gnanadesikan.

3. The Resilience Ratio Predicts the Character of Collapse

The most striking result of our statistical analysis is that the Phase 1 resilience ratio predicts the Baker-vs-Rahmstorf character of the AMOC breakdown.

By knowing a model's starting Resilience Ratio, we could accurately predict how its circulation would look after 150 years of extreme warming. We classified the remnant AMOC under full warming into three distinct regime categories based on its wind-driven fraction:

• Wind-Dominated Remnant (Baker-like)
  Example: CanESM5
  Mechanics: This model had a high initial resilience ratio (0.87), meaning winds were already doing a lot of the heavy lifting. Under 4xCO2, its sensitive thermohaline circulation collapsed hard (-64%). But because it had such strong wind support, the total volume of water moving north hit a hard "floor". Today, 70% of its remnant AMOC is purely wind-driven. This physically illustrates Baker's core argument: the circulation appears stable solely because the winds refuse to let it die entirely.
• Sinking-Dominated Remnant (Rahmstorf-like)
  Example: MIROC6
  Mechanics: This model had a very low initial resilience ratio (0.23), meaning it was almost entirely dependent on the density-driven sinking of cold water in the North. Under 4xCO2, as the oceans warmed and stratified, the thermohaline circulation weakened. Crucially, because it lacked a strong wind-driven floor to catch it, the overall circulation continued to drop. Only 29% of its remnant AMOC is wind-driven. This aligns perfectly with Rahmstorf's warning: when the thermohaline component remains the primary driver, the entire system is structurally fragile and susceptible to a true dynamical collapse.
• Mixed Remnant
  Examples: CESM2, MPI-ESM1-2-LR
  Mechanics: Models with moderate resilience ratios (~0.5) showed intermediate non-linear behaviour. The initial shock caused thermohaline weakening, but they eventually settled into stable remnants that are roughly 40-50% wind-driven, displaying a blended sensitivity to both forcing mechanisms.

---

Methodological Vulnerabilities & Caveats

While this framing provides a structured view of the debate, there are important methodological vulnerabilities to consider in this initial analysis:

• abrupt-4xCO2 is not a freshwater hosing experiment. The classic AMOC tipping literature focuses on freshwater forcing and saddle-node bifurcation dynamics. abrupt-4xCO2 primarily warms the surface, changes stratification, and modifies winds, but it does not directly simulate Greenland meltwater pulses in a realistic way. Critics could rightly argue this tests weakening, not true collapse dynamics.
• 150 years may not reach the true attractor. AMOC adjustment timescales can exceed 150 years. If the thermohaline component is drifting toward a bifurcation but hasn't crossed it yet, the late-period metric may capture a transient plateau, not the true asymptote.
• The Ekman "Floor" Assumes Wind Stationarity. We assume that as long as the winds blow, the floor persists. But CMIP6 models disagree substantially on Southern Ocean wind stress response shifting poleward and intensifying. If T_ek is not stable, the predictive power weakens.
• Sample Size. Only 5 models completed the pipeline so far. IPCC assessments typically rely on 20-40 models; this is an exploratory analysis.

The Takeaway: Dynamical vs. Transport Collapse

Rahmstorf's position is not just semantic. A wind-only overturning might preserve some volume transport (Sv) but radically change heat transport to Europe, deep water formation, and carbon sequestration structure. We must explicitly distinguish between a dynamical collapse (loss of density-driven overturning) and a transport collapse (overall Sverdrups dropping to near zero).

Overall, the Baker vs. Rahmstorf debate should not be about semantics of "collapse" versus "remnant". Our quantitative analysis shows that both sides are describing valid, physically distinct mechanisms that exist within state-of-the-art climate models. If the real-world ocean resembles CanESM5 (highly wind-supported), the absolute AMOC decline will be buffered, leaving a wind-driven remnant. If the ocean resembles MIROC6 (sinking-dominated), the circulation lacks that floor and is fundamentally far more fragile.

What's Next? (Phase 3)

To resolve this model uncertainty, we need strict observational constraints. In our next phase, we will validate these model-derived resilience ratios against historical reanalysis wind stress data (ERA5) and direct AMOC volumetric transport observations from the RAPID array.

By estimating present-day real-world T_ek / T_n, comparing it to the model distribution, and weighting models by closeness to observations, we move to constrained ensemble prediction. This aims to definitively answer: Which of these simulated futures is most likely to be ours?