Nick Schutgens, University of Oxford

DA in support of model evaluation and improvement

E. Oikawa, T. Nakajima


The evaluation of models with observations seems a routine procedure but the following three issues ensure that it is far from routine. First, observations have errors and this needs to be accounted for. Second, observations have a very different spatio-temporal sampling than models and, again, this needs to be accounted for. Third, models may have parametric errors that hide structural errors.

The above situation is even more relevant to aerosol models. Aerosol remote sensing observations are notably sparse, and have large errors. Also, aerosol models rely on emission inventories for anthropogenic aerosol and parametrisations for natural aerosol that specify the initial injection of matter into the atmosphere. Errors in these emissions easily lead to big deviations from observations without the need for any structural errors.

The above three issues make DA a tempting methodology to assess structural model errors. As a case study we performed DA experiments for three generations of a global aerosol model. Our DA method is an ensemble Kalman filter where the ensemble is due to perturbed emissions and meteorology. Errors in the emissions result in model biases that quickly dissappear due to DA. The annual difference between analysis and forecast is then interpreted as due to structural errors. The evaluation of the standard model is compared against evaluation of the DA, using observations independent and complimentary to the observations that were assimilated. Similarly, short-term forecasts starting from analyses for these three model versions are compared.

Combined, these studies reveal potential structural errors in sea spray and dust emission parametrisations as well as aerosol wet growth. Limitations in the methodology due to observational errors are also discussed, suggesting that this kind of study may work better with long-term averages than the hourly data currently used.