A Need to Account for Chaotic Physics in Weather and Climate Models

It was recently demonstrated (Sardeshmukh et al 2023, J. Climate) that randomizing the diabatic tendency of every model variable at every grid point and at every time step not only did not degrade the NOAA/GFS model’s global 2-week forecasts but actually improved them. This was due to both an increase in the forecast ensemble spread and a reduction in the ensemble-mean forecast error due to a stochastic “noise-induced drift”. We have since established that such randomizations also do not degrade the long-term mean climate and variability (including skewness and kurtosis, with large implications for extremes) of the new NOAA/UFS atmospheric model in large 40-member ensemble simulations of the 20-year 1999-2018 period. The randomizations do have long-term impacts that are highly statistically significant, but it is difficult to establish to what extent they are beneficial given the biases and uncertainties in the long-term statistics of even state-of-the-art verifying reanalysis datasets (ERA5 and JRA3q). Our results are consistent with previous studies of this topic but are cleaner, more extensive, and physically comprehensible. They also point to a disturbing conclusion that the increasingly detailed representation of diabatic processes in weather and climate models pursued at most modeling centers (which also motivates many observational field campaigns) may be unwarranted. What may instead be warranted is a physically consistent and more accurate representation of the specified space-time covariance statistics of the random perturbations used in such “stochastic parameterizations”. Some possible ways to accomplish this through careful numerical experimentation and innovative use of high-resolution observations will also be discussed.