Climate Science is just not that hard

From Andy Lacis, this is motivation to keep on looking at what Isaac Held calls the "fruit-fly" models of climate science.

Rabbet Run blog -- More from Andy Lacis

“It would seem more appropriate to assign "wickedness" to problems that are more specifically related to witches. The climate problem, while clearly complex and complicated, is not incomprehensible. Current climate models do a very credible job in simulating current climate variability and seasonal changes. Present-day weather models make credible weather forecasts – and there is a close relationship. Most of the cutting edge current climate modeling research is aimed at understanding the physics of ocean circulation and the natural variability of the climate system that this generates. While this may be the principal source of uncertainty in predicting regional climate change and weather extreme events, this uncertainty in modeling the climate system’s natural variability is clearly separate and unrelated to the radiative energy balance physics that characterize the global warming problem. The appropriate uncertainty that exists in one area of climate modeling doe not automatically translate to all other components of the climate system.”

I am continuing to work on the ENSO model, and as it gets simpler, the fit improves.  I started using an ENSO model that combines the SOI metric at 2/3 (Tahiti and Darwin) with the (negative to match SOI) NINO34 index at 1/3 (another interesting variation to consider is a median filter picking the middle value) This is a fit with a correlation coefficient of >0.80.

Fig 1: Recent SOIM fit. Yellow indicates regions of "poorer" fit

Another metric to consider is a variation of a binary classifier. The idea is simple. Since the model and data both show an oscillation about zero, then just by counting the number of times that both data and model agree with respect to positive or negative excursions, one can conveniently estimate fitness.  As it just so happens an 80% agreement in excursion classification also corresponds to around a 0.80 correlation coefficient.

Fig 2: Binary classifier for estimating correlation of model.

There is a limit to how high the CC can go, since the correlation between SOI and NINO34 tops off at about 0.86. This really has to do with nuisance noise in the system and the inability to identify the true oceanic dipole with confidence. So, for all practical purposes, a CC of 0.80 or identifying + or - excursions at 80% accuracy is quite good, with diminishing returns after that (since the fit is to the noise).

The other metric that I am exploring is an estimate of the agreement between two wavelet scalograms.   As one can see below, the fit in 2 dimensions appears quite good and the extra degrees of freedom provide better discrimination in identifying the better fit between two models.

Fig 3: Wavelet scalograms.

I am asking the participants at the Azimuth Forum as to how best to create an effective CC for comparing wavelet scalograms, but have had no response so far.

BTW, Archived discussions of the ENSO Revisited thread at the Azimuth Forum

Also a battle I had over at Real Climate  with a neo-denier. The moderators at RC dispatched the fella to the Bore Hole thread.

The Oil Shock Model Simplified

DC at the OilPeakClimate blog spent some time at re-analysis of the Oil Shock model and Dispersive Discovery model which were originally described in The Oil Conundrum Book (see menu above).

Whenever a model is re-analyzed, there is the possibility of improvement. In DC's case, he made a very smart move to try to separate the extra-heavy oil as a distinct process. The Shell Oil discovery data appears to combine the two sets of data, leading to a much larger URR than Laherrere gets. What he accomplished is to reconcile the lighter-crude Laherrere discovery data and the reality that there are likely ~500 GB of extra-heavy crude waiting to be exploited. Whether this effectively happens is the big question.

Read DC's whole post and potential discussion here, as he has made an excellent effort the last several years of trying to digest some heavy math and dry reading from the book.   He is also making sense of the Bakken oil production numbers in other posts.

As a PS, I have added an extra section in the book to describe the dispersive diffusion model describing the Bakken production numbers.

DC also posted this piece on the PeakOilBarrel blog. Almost 500 comments there.

CSALT re-analysis

I have previously described a basic thermodynamic model for global average temperature that I call CSALT. It consists of 5 primary factors that account for trend and natural variability. The acronym C S A L T spells out each of these factors.

C obviously stands for excess atmospheric CO2 and the trend is factored as logarithm(CO2).

S refers to the ENSO SOI index which is known to provide a significant fraction of temperature variability, without adding anything by way of a long-term trend. Both the SOI and closely correlated Atmospheric Angular Momentum (AAM) have the property that they revert to a long-term mean value.

A refers to aerosols and specifically the volcanic aerosols that cause sporadic cooling dips in the global temperature measure. Just a handful of volcanic eruptions of VEI scale of 5 and above are able to account for the major cooling noise spikes of the last century.

L refers to the length-of-day (LOD) variability that the earth experiences. Not well known, but this anomaly closely correlates to variability in temperature caused by geophysical processes. Because of conservation of energy, changes in kinetic rotational energy are balanced by changes in thermal energy, via waxing and waning of the long-term frictional processes. This is probably the weakest link in terms of fundamental understanding, with  the nascent research findings mainly coming out of NASA JPL (and check this out too).

T refers to variations in Total Solar Irradiance (TSI) due mainly to sunspot fluctuations. This is smaller in comparison to the others and around the level predicted from basic insolation physics.

Taken together, these factors account for a correlation coefficient of well over 90% for a global temperature series such as GISTEMP.

Over time I have experimented with other factors such as tidal periodicities and other oceanic dipoles such as North Atlantic Oscillation, but these are minor in comparison to the main CSALT factors. They also add more degrees of freedom, so they can also lead one astray in terms of spurious correlation.  A factor such as AMO contains parts of the temperature record itself so also needs to be treated with care.

My biggest surprise is that more scientists don’t do this kind of modeling. My guess is that it might be too basic and not sophisticated enough in its treatment of atmospheric and climate physics to attract the General Circulation Model (GCM) adherents.

So what follows is a step-by-step decomposition of the NASA GISS temperature record shown below, applying the symbolic regression tool Eureqa as an alternative to the CSALT multiple linear regression algorithm. This series has been filtered and a World War II correction has been applied.

The GISTEMP profile, interval 1940-1944 corrected as a 0.1 adjustment down.


Continue reading

Imaging via Particle Velocimetry of Sloshing

This is an interesting paper on capturing the volumetric effects of sloshing:

Simonini, A., Vetrano, M. R., Colinet, P., & Rambaud, P. (2014). Particle Image Velocimetry applied to water sloshing due to a harmonic external excitation. In Proc. of the 17th International Symposium on Applications of Laser Techniques to Fluid Mechanics.
The scale that they describe applies to containers filled with liquid subject to external forces.
Compare against the very large scale equatorial Pacific dynamics

linked from NOAA here

Absolute temperature, from which the anomaly is based


Anomaly of temperature. The emerging hotspots are what lead to El Nino conditions.

Forecasting versus Problem-Solving

The challenge of explaining climate phenomenon such as ENSO leads to an interesting conundrum.  Do we want to understand the physics behind the phenomenon, or do we want to optimize our ability to forecast?

Take an example of the output of a crude power supply. Consider that all one has is one cycle of output.

  1. The forecaster thinks that it is fair to use only one half of that cycle, because then he can use that to forecast the other half of the cycle.
  2. The problem solver wants the whole cycle.

Why is the problem solver in better shape?

  1. The forecaster looks at the half of a cycle and extrapolates it to a complete cycle. See the dotted line below.
  2. The problem solver looks at the continuation and discovers that it is a full-wave rectified signal. See the solid line below

In this case, the problem solver is right because the power supply happens to be a full-wave rectifier needed to create a DC supply voltage.  The forecaster happened to make a guess that would have been correct only if it was an AC power supply.

Lose your generality and that is what can happen. As Dara says, the key is to look for  structures or patterns in the data -- while reducing the noise -- and if that means to use as much of the data as possible, so be it.


Paper on Sloshing Model for ENSO

I recently archived the paper to ARXIV and submitted to PRL.

Get the paper here from ARXIV as a PDF.

The nearly year-long investigation is time-lined and outlined here.

The final model fit:

Fig 1: Figure from the paper. When the correlation reached 0.8, I thought it might be the right time to stop. The noise in the Darwin or Tahiti time-series was the limiting factor in how good the fit could eventually become.

Thanks for the good comments!

One application of the model described below

Continue reading

Demodulation and the SOIM

A few people have emailed me asking for a simplified explanation for the model of ENSO that I have been working on. I can appreciate this because presenting solutions to differential equations is not one of the most intuitive ways to convey the essential model.

I decided to go with a mathematical analogy that hopefully will appeal to a techie. Although analogies do not often work when presenting a scientific model, if it turns out that essentially the same mathematical formulation describes the foundation, then it hopefully will serve some use, as many people may already be familiar with the compared-to mathematical construct. So the idea is use the elements of the Southern Oscillation Index Model (SOIM), which includes (1) a basic wave equation, and (2) a forcing provided by stratospheric winds (i.e. the QBO) to convince you that you can understand a model that  effectively describes the seemingly erratic behavior of ENSO. Understanding this concept will be music to your ears if you are willing to listen and appreciate the concept.

Continue reading

Two modes to ENSO Variability

Since the last post in the SOIM series, where we identified a highly correlated behavior in tide gauge data to the ENSO SOI measure, it occurred to me that many of the earlier entries in this series seemed to go down a different path.  That path still involved Mathieu differential equations, but the parameters differed and the right-hand side (RHS) forcing function did not play as significant a role (described initially here with a couple of follow up posts).

Yet from the first of the tide gauge posts there was a clear indication that an additional distinct sloshing mode might apply, and that this could be related to the "phantom" Mathieu waveform we may have been chasing earlier.

Continue reading

An ENSO Predictor Based on a Tide Gauge Data Model

Earlier this year, I decided to see how far I could get in characterizing the El Nino / Southern Oscillation through a simple model, which I referred to as the Southern Oscillation Index Model, or SOIM for short (of course pronounced with a Brooklyn accent). At the time, I was coming off a research project where the task was to come up with simple environmental models, or what are coined as context models, and consequently simple patterns were on my mind.

So early on I began working from the premise that a simple nonlinear effect was responsible for the erratic oscillations of the ENSO. The main candidate, considering that the ENSO index of SOI was clearly an oscillating time-series, was the Mathieu equation formulation. This is well known as a generator of highly erratic yet oscillating waveforms.  Only later did I find out that the Mathieu equation was directly used in modeling sloshing volumes of liquids [1][2]  --  which makes eminent sense as the term "sloshing" is often used to describe the ENSO phenomena as it applies to the equatorial Pacific Ocean (see here for an example).

Over the course of the year I have had intermittent success in modeling ENSO with a Mathieu formulation for sloshing, but was not completely satisfied,  largely due to the overt complexity of the result.

However, in the last week I was motivated to look at a measure that was closer to the concept of sloshing, namely that of sea surface height. The SOI is an atmospheric pressure measure so has a more tenuous connection to the vertical movement of water that is involved in sloshing. Based on the fact that tidal gauge data was available for Sydney harbor (Fort Denison here)  and that this was a long unbroken record spanning the same interval as the SOI records, I did an initial analysis posted here.

The main result was that the tidal gauge data could be mapped to the SOI data through a simple transformation and so could be used as a proxy for the ENSO behavior. The excellent correlation after a delay differential of 24 months is applied  is shown in Figure 1 below.

Fig 1:  The first step is to map a proxy (tide gauge data) to the SOI data

That was the first part of the exercise, as we still need to be able to quantify the tidal sea surface height oscillations in terms of a Mathieu type of model. Only then can we make predictions on future ENSO behavior.

As it turns out the model appears to greatly simplify, as the forcing, F(t), for the right hand side (RHS) of the Mathieu formulation consists of annual, biannual (twice a year), and biennial (once every two years) factors.

 \frac{d^2x(t)}{dt^2}+[a-2q\cos(2\omega t)]x(t)=F(t)

The last biennial factor, though not well known outside of narrow climate science circles [3], is critical to the model's success.

Although the Mathieu differential equation is simple, the solution requires numerical computation. I (along with members of the Azimuth Project) like to use Mathematica as a solver.

The complete solution over a 85-year span is shown in Figure 2 below

Fig 2: The second step is to model the tidal data in terms of a sloshing formulation. The biennial factor shows a phase reversal around 1953, switching from an even to odd year periodicity. The yellow highlighted area is one of the few regions that a correlation is clearly negative. Otherwise the fit models the behavioral details quite effectively.

This required an optimization of essentially three Mathieu factors, the a and q amplitudes, and the ω modulation (along with its phase). These are all fixed and constitute the LHS of the differential equation.  The RHS of the differential equation essentially comprises the amplitudes of the annual, biannual, and biennial sinusoids, along with phase angles to synchronize to the time of the year. And as with any 2nd-order differential equation, the initial conditions for y(t) and y'(t) are provided.

As I began the computation with a training interval starting from 1953 (aligning with the advent of QBO records), I was able to use the years prior to that for a validation.  As it turns out, the year 1953 marked a change in the biennial phase, switching from odd-to-even years (or vice versa depending on how it is defined).  Thus the validation step only required a one-year delay in the biennial forcing (see the If [ ] condition in the equation of Figure 2).

The third step is to project the model formulation into the future. Or further back into the past using ENSO proxies. The Azimuth folks including Dara and company are helping with this, along with two go-to guys at the U of MN who shall remain nameless at the present time, but they know who they are.

Ultimately, since the model fitting of the tide data works as well as it does, with the peaks and values of the sloshing waters effectively identified at the correct dates in the time series, it should be straightforward to transform this to an ENSO index such as SOI and then extrapolate to the future. The only unknown is when the metastable biennial factor will switch odd/even year parity.  There is some indication that this happened shortly after the year 2000, as I stopped the time series at this point.  It is best to apply the initial conditions y and y' at this transition to avoid a discontinuity in slope, and since we already applied the initial conditions at the year 1953, this analysis will have to wait.

The previous entries in this series are best observed by walking backwards from this post, and by visiting the Azimuth Forum.   Science is messy and nonlinear as practiced, but the results are often amazing.  We will see how this turns out.


[1] Faltinsen, Odd Magnus, and Alexander N Timokha. Sloshing. Cambridge University Press, 2009.
[2] Frandsen, Jannette B. “Sloshing Motions in Excited Tanks.” Journal of Computational Physics 196, no. 1 (2004): 53–87.
[3] Kim, Kwang-Yul, James J O’Brien, and Albert I Barcilon. “The Principal Physical Modes of Variability over the Tropical Pacific.” Earth Interactions 7, no. 3 (2003): 1–32.


Keep it Lit

Good luck to the People's Climate Marchers.  I read Bill McKibben's book Long Distance several years ago, and realize that persistence and endurance pays off. I also realize that there are no leaders in the movement, and that we all have to pull together to get off of fossil fuel.  If we each do our share, the outcome will tend more toward the good than to the bad.