Variable resolution modeling in the grey-zone with the Model for Prediction Across Scales (MPAS)

This blog is a follow-up on my previous post, which concluded with a statement that we intended to find the root-cause of the troubling results of various MPAS runs. Thanks to Dominikus Heinzeller and Michael Duda (NCAR) we finally discovered a bug in the pre-processing software of MPAS. This bug has been fixed, and the simulations were repeated afterwards. The results were that encouraging that I decided to write another blog on this topic.

To refresh our minds...why MPAS?

The traditional weather models of today can be separated into two groups: the global models and regional models. Global models (resolution > ~10km) provide global forecasts, but they are based on the hydrostatic assumption which means that they do not resolve meso-scale weather phenomena. In contrast, the non-hydrostatic regional models (resolution < ~10km) do resolve meso-scale weather phenomena, but they are in general more expensive. Both model groups have in common that they only allow for a fixed resolution.

The Model for Prediction Across Scales (MPAS), basically bridges the gap between the two model groups, allowing for weather predictions on a global scale and resolving for meso-scale weather phenomena in the region of interest (the USA in the picture below). This is made possible by 1) the global variable resolution (ir)regular Voronoi grid allowing for a smooth resolution transition, and 2) the scale-aware Grell-Freitas cumulus scheme that gradually fades out at high resolution (where clouds start being resolved).

As stated in the picture above, the variable resolution Voronoi grids utilized in MPAS allow for a smooth resolution transition from the coarse global domain towards high resolution in the domain of interest. This is a very interesting feature as this will allow us to define one model grid where we are able to refine the grid in the regions of interest, where our customers require high resolution weather forecast data.

Testing the meteorological and numerical performance of MPAS

The main goal was to demonstrate the potential of MPAS (a collaborative project between LANL and NCAR) in terms of the computational performance and the meteorological quality over the European continent. Matthijs compared the meteorological output with observations and the output of MeteoGroup's operational WRF model with a 3km regional mesh over Western Europe. We selected three case studies as shown in the following picture.

While MPAS has been tested thoroughly over the US continent, this study is unique in the sense that we have assessed MPAS with high resolution over Europe.

Model set-up and sensitivity tests

We have performed the WRF runs using our in-house configuration using version 3.6.1. The MPAS runs were executed at the SuperMUC Leibniz Supercomputing System. We have used the following domains:

  • Regional WRF 3km (grey-shaded area in the pictures below, 14M cells)
  • Global MPAS 3km (250 times WRF grid)
  • Global MPAS 15-3 km (left picture, 25 times WRF grid)
  • Global MPAS 30-3 km (similar to right picture, 5 times WRF grid)
  • Global MPAS 60-3 km (right pictures, 3 times WRF grid)

It is important to note that the global variable resolution mesh 60-3 km is only three times as large as the regional WRF grid!

The WRF runs have been performed with an adaptive time stepping scheme (~18s), and the MPAS runs used a fixed time stepping scheme. We investigated the influence of initialization time on the outcome as well by two runs with different initialization times with a 12 hour difference. All runs have been initialized with ECMWF data, and data assimilation was not used (not available in MPAS). The parameterization schemes are listed in the following table.

Image Description

Synoptic gale "St Jude"

Synoptic description

The storm traveled over the North Sea area on October 28, 2013. It was a strong depression with a minimum pressure of 975 hPa, and strong winds up to 12 Beaufort. The storm caused at least 17 casualties and ~100 million Euro's of damage in the Netherlands only.

Image Description

Pressure and wind barbs at October 28, 6UTC

The picture below shows the pressure fields and wind barbs for various WRF/MPAS runs. From this picture we find that:

  • The regional WRF and global MPAS 3km results (top-left, top-middle figures) show similar patterns, and they compare very well with the KNMI analysis shown in the figure above.
  • Coarsening the MPAS mesh outside Europe does not change the outcome in the high resolution domain itself (top-middle -> top-right -> bottom-right -> bottom middle figures)!
  • There are only minor differences between the two MPAS runs with different initialization times (bottom-left, bottom middle figures)

Wind time series

In the following picture we focus on the wind (speed/gust/direction) time series, and compare the simulated series with the observed series at the KNMI station Vlieland. In the observations (top-left figure) we see wind gusts up to 42 m/s (150 km/h). The simulation wind time series have been extracted (by means of a search algorithm) at the location of maximum wind near the depression core. From this figure we observe that:

  • In contrast with MPAS, the wind gusts in WRF have been slightly enhanced to account for unresolved turbulence. This becomes clear in observing the peak wind gust, which is in WRF higher compared with that in MPAS.
  • The wind time series are very similar in all MPAS runs. So, again we see that coarsening the mesh outside Europe does not affect the results much in the high resolution domain.
  • Later initialization shows higher wind speeds.

Föhn event

Synoptic description

We analyzed a föhn event that occurred in Switzerland in the period November 3-6, 2014. The KNMI analysis below shows a strong southerly flow over the Alps, causing long-lasting steady precipitation on the windward side of the Alps (induced by water supply from the Mediterranean Sea) and strong warm (föhn) winds on the leeward side.

Image Description

Cumulative rain during the event

The picture below shows the cumulative rain fall during the event. The black dashed line indicated the observed rain fall at the MeteoGroup station Intragna (close to the Italian boarder). We note that the simulated rain fall time series have been extracted (by means of a search algorithm) from the grid point with maximum rain fall. We see a delayed onset of rain fall in both the WRF and MPAS runs. In addition, the rain fell in a shorted period of time, compared with the observation. In general, the simulated final precipitation amount at later initialization is closer to the observed amount.

Image Description

Cumulative rain at November 6, 12UTC

In this section we compare the cumulative rainfall over Switzerland at November 6, 12UTC. As the MPAS results are fairly insensitive to the mesh configuration we only show here the 60-3km result. We see that the regional 3km WRF and the variable resolution (60-3 km) MPAS runs compare well, albeit that WRF produces more clustered precipitation patterns across Switzerland. The timing, location and magnitude of simulated rainfall compares well with the observations.

Temperature and wind speed at November 4, 6UTC (end of the night)

In the picture below we compare the temperature and wind speed patterns across Switzerland at November 4, 6UTC. The top figures show the observations, the middle figures the WRF results, and the bottom figures the MPAS 60-3km results. The observations in the black ellipses clearly indicate the föhn front (the inclined black line). On the east-side of the föhn front thee temperature exceeds 20 degrees and wind speeds above 15 m/s! This phenomenon is clearly present in both WRF and MPAS.

Convective hailstorm

Synoptic description

On August 30, 2015, a stationary warm front was positioned over the Netherlands. This front induced vertical instability, leading to severe hail with hail stone sizes of about 6cm in Oss (Netherlands).

Image Description

Cumulative rain during the event

The picture below shows the cumulative rain fall during the event. The black dashed line indicated the observed rain fall at the KNMI station Herwijnen. Similar to the previous case study, the simulated rain fall time series have been extracted (by means of a search algorithm) from the grid point with maximum rain fall. Both MPAS and WRF show a simulated precipitation trend comparable to that observed at Herwijnen. We see that the runs initialized at August 30, 12 UTC perform best (although with a delayed onset).

Image Description

Total rain fall at August 31, 2015

When we analyze the total rainfall on August 31 we see that both WRF and MPAS position the front more northward. While WRF produces more precipitation over the North Sea area in the early initialization at August 30, 0UTC, we have seen that later initialization at 12 UTC shows more enhanced precipitation (both in WRF and MPAS). We only show here the MPAS 60-3km result, as it is very similar in all the other MPAS mesh configurations. This is in line with what we already concluded from the previous 2 case studies.

Conclusions

The MPAS simulation results achieved after fixing the bug in the MPAS pre-processing software were much more consistent and in line! From this we may conclude the following:

  • The variable resolution MPAS performs well in all three extreme weather events investigated. The results were of similar quality compared with the uniform 3km regional WRF and the massive global MPAS runs.
  • Initialization closest to the time of the event gives the best result.
  • MPAS shows excellent scaling on modern HPC architectures (see the result of a scaling experiment in the figure below). This means that the MPAS 60-3km set-up has a global coverage which is only 3 times more expensive than a regional WRF!

Next steps

A paper of this study is intended to be published in the Climate Dynamics journal. The results have been presented at the European Geophysical Union (EGU) meeting in April 2017, the WRF annual workshop in Boulder in June 2017 and the European Meteorological Society (EMS) meeting in September 2017.

Acknowledgements

I would like to thank Matthijs Kramer for doing his internship at MeteoGroup. He was guided by Wim van den Berg and me from the MeteoGroup Research Department. I also acknowledge Dominikus Heinzeller from the Karlsruhe Institute of Technology, Gert-Jan Steeneveld from the Wageningen University, and Michael Duda from NCAR for providing the early-development versions of the MPAS model and the grid setups.

References

[1]: Bukovsky MS, Karoly DJ (2009). Precipitation simulations using WRF as a nested regional climate model.

[2]: Real-time MPAS convection experiments

[3] Ebert EE, Janowiak JE, Kidd C (2007). Comparison of near-real-time precipitation estimates from satellite observations and numerical models.