Pacific Low-Cloud Feedback Under Warming and High CO2

The finding, kept at its actual scale

Low marine clouds cool the climate by reflecting sunlight. Their turbulent structures are much smaller than the grid cells used by conventional global climate models, so the direction and strength of their response to warming have remained an important source of climate-sensitivity uncertainty. The new study by Chammas and colleagues does not solve that global problem in one step. It does something narrower and useful: it resolves selected tropical Pacific low-cloud states at cloud scale, repeats the experiment thousands of times, and separates the response to warmer sea surfaces from the direct radiative effect of higher atmospheric CO2.

In the warming-only experiment, the median net low-cloud feedback was +0.14 W m^-2 K^-1, with a 95% bootstrap confidence interval of +0.02 to +0.25. A positive value means the simulated cloud changes leave the climate system absorbing more net energy as the surface warms. The interval excludes zero within this model design.

The distinction in the final column matters. The paper reserves feedback for the warming-only response. The combined experiments include both temperature-mediated cloud change and a rapid response to CO2, then express the total response per kelvin of imposed warming. The +0.43 value is therefore a diagnostic of the combined high-CO2 state, not a portable global-feedback coefficient.

How the experiment was built

The authors sampled 500 locations between 35 degrees south and 35 degrees north from an observed tropical Pacific low-cloud distribution. Four representative months captured seasonal variation. Every location-month pair was prepared in five states: present-day baseline, +4 K sea-surface warming, quadrupled CO2 with baseline sea-surface temperature, +4 K with doubled CO2, and +4 K with quadrupled CO2. That factorial design produced 10,000 candidate simulations.

Cases in which clouds rose above 4 km were excluded so the analysis remained about low clouds, and a perturbation comparison was retained only when both its baseline and altered case passed that filter. The analytical ensemble contained 7,083 simulations. The filter is defensible for the stated question, but it can select against cases most prone to deepening into convection, which the authors identify as a possible bias.

This is high resolution in a cloud-process sense. Each local domain resolves the dominant turbulent eddies rather than parameterising all cloud-scale motion. It is not a 49-metre simulation of the globe: the horizontal domain covers only 36 square kilometres, uses periodic lateral boundaries and cannot represent larger mesoscale cloud organisation. Large-scale forcing and sea-surface conditions came from one host global model, NOAA-GFDL CM4.

Google's role extended beyond hardware. Nine of the ten authors list Google Research as an affiliation, the simulation software was designed to scale across Google's tensor processing units, and Google-affiliated authors contributed methodology, software, resources and analysis. The paper discloses that nine authors are Google employees and hold Alphabet stock through standard compensation. It also identifies the code, processed statistics and 203.78 TB of raw simulation output as public research assets.

The mechanism is not simply thinner cloud

Every net response in the experiment is dominated by shortwave radiation. As low-cloud area falls, less incoming sunlight is reflected to space. The longwave response is smaller and negative, offsetting part of the warming effect without reversing it.

All values are ensemble medians in W m^-2 K^-1. Median components do not have to sum exactly to the median net response.

Warming alone reduced cloud fraction, but it also increased median in-cloud liquid water path by 2.4 g m^-2. The surviving clouds became optically thicker and brighter, partially compensating for the reflective area that was lost. Quadrupled CO2 at baseline sea-surface temperature moved both terms in the warming direction: cloud fraction fell and median in-cloud liquid water path fell by 4.7 g m^-2. The study links this rapid adjustment partly to weaker cloud-top longwave cooling, which reduces the turbulence that helps maintain low clouds.

In the combined +4 K and quadrupled-CO2 state, the warming-related environmental changes buffered some of the direct CO2 effect, so the response was less than a simple sum of the separate experiments. Even so, the compensating cloud thickening almost disappeared: median liquid-water-path change was -0.1 g m^-2 while cloud fraction remained lower. The high-CO2 result is stronger because reflective area is lost without the same brightening of the clouds that remain.

The quadrupled-CO2-only experiment is the cleanest warning against mixing unlike quantities. It held sea-surface temperature at baseline and produced a median +1.93 W m^-2 cloud-radiative adjustment. With no imposed temperature change, there is no kelvin denominator. It is a rapid adjustment in W m^-2, not a feedback in W m^-2 K^-1.

What this changes about climate sensitivity

The result strengthens an evidence direction that was already established. The IPCC Sixth Assessment concluded with high confidence that net cloud feedback is positive and assessed equilibrium climate sensitivity at a best estimate of 3 degrees C, with a likely range of 2.5 to 4 degrees C. A 2021 observation-based study estimated near-global marine low-cloud feedback at +0.19 +/- 0.12 W m^-2 K^-1 at 90% confidence. The new warming-only Pacific result is similar in sign and scale, but the domains, methods and confidence definitions differ and the values are not interchangeable.

The more consequential contribution is the nonlinear high-CO2 response. It supports the possibility that low-cloud behaviour becomes more amplifying as climate state changes, and that models which underrepresent rapid CO2-cloud adjustment may understate the upper tail in much warmer states. That is a reason to investigate state-dependent sensitivity. It is not a reason to replace the assessed global ECS range with +0.43, because the experiment did not calculate ECS.

The study used prescribed sea-surface temperatures rather than a coupled ocean-atmosphere system, sampled selected Pacific low-cloud regimes rather than every global cloud type, and did not include land, polar, high-cloud or carbon-cycle feedbacks. It cannot produce a global equilibrium temperature. The defensible conclusion is mechanism-level: positive Pacific low-cloud feedback is robust within this ensemble, and the tested high-CO2 state produces a substantially stronger total response.

Not a new cloud tipping point

The same physical research lineage includes the 2019 Caltech stratocumulus study, which simulated an abrupt breakup of one representative subtropical cloud deck above roughly 1,200 ppm CO2 in an idealised framework. The 2026 work is broader in sampling but does not repeat that concentration sweep. It tests baseline, doubled and quadrupled CO2 states, not the intermediate steps or a descending path needed to identify a threshold and hysteresis.

Both studies identify reduced cloud-top longwave cooling under high CO2 as an important mechanism. The new ensemble shows that direct CO2 adjustment matters across many Pacific cloud states and becomes much stronger by the quadrupled-CO2 experiment. It does not establish that every low-cloud regime breaks abruptly, recover the earlier 1,200 ppm threshold or simulate an extra global temperature jump.

Where the uncertainty still sits

The 95% intervals quantify variation across the retained ensemble and bootstrap resampling. They do not include every form of structural uncertainty. The large-scale forcing comes from one host global model; warming is a uniform +4 K rather than a realistic spatial pattern; the 6 km domain omits mesoscale organisation; simulations last five days; aerosol and microphysical assumptions are simplified; and the direct high-CO2 nonlinearity is sampled at only two elevated concentrations.

The next material tests are therefore clear: repeat the ensemble with multiple host models, extend it beyond the Pacific, use patterned warming, sample CO2 concentrations between doubled and quadrupled states, increase domain size and duration, and reproduce a meaningful subset with independent software and infrastructure. No substantive independent replication or technical response specific to the 24 July paper had emerged by this report's 26 July evidence cutoff.

How Solar Analytica carries the result

These values are published here as a literature-evidence layer. They do not enter the observational annual climate tables, the measured brightening and dimming record, or the location-level PV yield baselines. The paper models selected cloud states; it does not provide a new observed global time series or a site-specific solar-production adjustment.

The appropriate connection is contextual. The State of the Solar Climate tracks the measured and modelled conditions that affect solar assets. This report explains a process that may change future cloud reflectivity and climate sensitivity, while keeping that mechanism separate from the operating evidence already used in solar decisions. The distinction follows the Solar Analytica methodology: modelled evidence is labelled, scope travels with the number, and a regional process estimate is never promoted into a global observation.