Discover more from The Climate Brink
Global temperatures remain consistent with climate model projections
Here we update the comparison of climate models and observations through October 2023
With extreme global temperatures in recent months I see one question asked over and over again: does this mean that global warming is happening faster than we thought?
While there is growing evidence that the rate of warming has increased in recent decades compared to what we’ve experienced since the 1970s, this acceleration is largely included in our climate models, which show around 40% faster warming in the period between 2015 and 2030 compared to 1970-2014.1
In many ways, the more interesting question is how do temperatures in recent years (and months!) compare to what climate models have projected for this period.
The figure below shows the multimodel mean and the range (5th to 95th percentile) across model runs for the latest generation of climate models (CMIP6), along with the latest observations from Berkeley Earth. I’ve even included a preliminary estimate of October temperatures based on reanalysis data from JRA-55 for the month to-date.
In the CMIP6 ensemble the past few months (September and October) are above the multimodel mean, but remain well within the envelope of model projections.
However, the full CMIP6 ensemble might not represent the most accurate assessment of future temperatures; as we argued in Nature last year, CMIP6 contains a subset of models that are running quite hot (>5C climate sensitivity per doubling CO2) that generally do a poor job of representing historical global surface temperatures.
For this reason, the most recent IPCC report created Assessed Warming Projections that weight the CMIP6 ensemble by its performance in matching historical temperatures, and tend to show future warming more in-line with the prior generation of climate models (CMIP5).2
The figure below shows a comparison between CMIP5 models (which were developed in the lead-up to the 2014 IPCC 5th Assessment Report. Here we see that observations are on the high end of the model range in recent months, with September slightly above the 95th percentile of the model envelope.
However, we should not read too much into an individual month exceeding the model envelope. This has happened before – notably in large El Nino events like 1998 and 2016. In fact, we would expect roughly one in 20 months to be above the 95th percentile if models are accurately representing real-world variability.
While climate models include El Nino and La Nina behavior, these are emergent properties of the climate system and may occur in climate models spun-up in the mid-1800s at very different times than they occur in the real world. Because only a fraction of the 40-odd CMIP5 or CMIP6 models have an El Nino event occurring in 2023, its not surprising that they might not fully capture El Nino-related climate variability in any given year.
That being said, its worth noting that we are still pretty early in the current El Nino cycle. In terms of current conditions, 2023 is more similar to 1997 or 2015 than the (at the time) record-warm 1998 or 2016. It remains to be seen if we will see more exceptional warmth in the latter part of this year and early next as the El Nino event peaks or if this El Nino is behaving differently – potentially contributing more warming early on due to the rapid transition out of unusually persistent La Nina conditions – than we’ve seen in past events.
Finally, its useful to take a longer view of how climate models and observations have historically stacked up. The figure below shows CMIP5 model temperatures between 1900 and 2030 compared to Berkeley Earth observations over the same period:
Its also worth noting that the comparisons presented above use a 1900-2000 baseline period to align models and observations; the apparent visual agreement or disagreement between models and observations is quite sensitive to the choice of baseline period. Thankfully, there is a way to avoid the baseline sensitivity issue by looking at trends rather than just anomalies over time.
Taking the trendy view
The rate of warming over time (the trend) is insensitive to the choice of baseline period. That makes it a particularly useful way to compare climate models and observations.
The figure below shows a histogram of the trends in all the individual CMIP6 models compared to observations over the prior 15 year period (January 2008 through September 2023). Four different observational records are analyzed – Berkeley, NASA’s GISTEMP, NOAA’s GlobalTemp, and Copernicus/ECMWF’s ERA5 reanalysis product.
In general, observations are pretty close (albeit slightly below) the CMIP6 multimodel mean trend over the past 15 year period. While Berkeley, NASA, and NOAA all show similar trends of just above 0.3C per decade, Copernicus shows a bit more warming (perhaps reflecting its measurement of surface temperatures over the oceans rather than sea surface temperatures). However, there is a wide range of model warming over this period, with some models showing nearly twice as much warming over this short period and some showing no warming at all.
We can also look further back in time; the figure below compares the trend since 1970 (e.g. 53 years) rather than just examining the past 15 years.
Here we see a narrower range of model trends, and that observations are more notably on the low end of the CMIP6 ensemble for the full period. There is also a clear bifurcation between warming rates, with a subset of models showing warming rates of 0.3C to 0.35C per decade with the rest clustered around 0.15 to 0.25C per decade.
If we look at CMIP5 models instead, we see that observed trends over the past 15 years are in the upper half of model projections, but still well within the range. Here the range of trends over the past 15 years is notably narrower than we saw for the same period in CMIP6, representing the more realistic range of climate sensitivity (and transient climate response) across the CMIP5 ensemble.
If we look at CMIP5 trends since 1970, we see that observations are well within the range, though slightly below the mean CMIP5 model.
Sampling the whole cherry tree
While its useful to look at trends over the past 15 years (or 53 years), it is also a bit of an arbitrary period of time to select, out of all period where we could have looked at trends. The past 15 years in particular may be a bit unrepresentative, as it starts with a moderately strong La Nina event and ends in the current El Nino event.
However, there is a way to examine trends over time that does not depend on the choice of start date: simply look the trends for all possible start dates. The figure below shows the trend between today (September 2023) and each possible start date from January 1970 to October 2013 (e.g. 10 years ago).3
This figure is a tad complicated, so lets take some time to explain it. A point on the x-axis (say, 1980) shows the model mean trend and 5th-95th percentile of trends across all the CMIP6 models. Trends for each start date are also shown for the four different observational temperature records we are looking at in this analysis.
Here we see that observations are below the CMIP6 multimodel mean for virtually all possible trend periods. This is not a particularly surprising result, given that a subset of CMIP6 models is known to be running hot and the Assessed Warming Projections in the recent AR6 are notably lower than the multimodel mean (e.g. have a trend of 0.26C per decade from 2015-2030, vs 0.31C per decade for the CMIP6 multimodel mean).
That being said, CMIP6 contains a wide range of model projections, and observed trends remain well within this range for all trend periods examined.
If we turn to CMIP5 models, we find greater agreement between observed trends and model projections. Observed trends since 1970 are still slightly below the multimodel mean, but well within the envelope of model projections. In recent years (post-1998) there are periods where the observed trends are above the multimodel mean.
Overall, an examination of models and observations through present provides a number of useful takeaways:
Despite the extreme global temperatures experienced in the summer of 2023, observations remain broadly consistent with climate model projections this year, though this could potentially change as the current El Nino continues to develop.
The a subset of the latest generation of models (CMIP6) continues to run well above observations despite recent warmth. Researchers should avoid using an unweighted CMIP6 ensemble for global temperature analyses when possible as discussed here.
The prior generation of models (CMIP5) continues to perform quite well, and is consistent with the new Assessed Warming Projections used in the recent IPCC AR6.
2015 is used as a dividing point here as its when the latest generation of models switches between historical runs and future projections. Its also the date when the assessed warming projections in the recent IPCC report become available.
The IPCC AR6 Assessed Warming Projections begin in 2015 and only provide annual values, so they are less useful to use to compare to observations. There are ways of selecting a subset of CMIP6 models consistent with assessed warming (as we suggest in our paper), but I’m focusing on comparisons with CMIP5 for simplicity here.
Trends shorter than 10 years are not shown as the uncertainties balloon over short periods due to year-to-year variability. Also, while exploring all possible start dates is a useful sensitivity test, there can still be some unintentional cherry-picking when ending the timeseries during an El Nino event. The longer the trend period examined, the less sensitive the results will be to choice of end point (or start point).