Track prediction of Tropical Storm Surigae (formerly 91W) by 60km MPAS

Life
Page content

Which matters more for 91W’s track, DA or the NWP model?

In my previous post, I mentioned that for forecasts initialized from 2026-09-22T12, all NWP models (GFS, ECMWF, CMC, ICON) did not predict that 91W (now Tropical Storm Surigae) will go northward in front of taiwan, while DLWP models, Google’s WN-cyclone and ECMWF’s AIFS predict that 91W will make the turn. (check those forecasts at https://www.tropicaltidbits.com/)

For the forecast initialized at 2026-09-23T18, ECMWF’s NWP model started to predict that 91W will turn around in front ot Taiwan, agreeing more with the DLWP model predictions.

By contrast, GFS did not. For the cycle initialized at 2026-09-23T18, GFS still predicted that 91W would approach northern Taiwan and then move southwards, due to the interaction with the other approaching storm on its east.

Since I’m running a low-resolution (60km) global MPAS routinely and working on data assimilation, I care about whether it is the initial condition or the NWP model that contributes more to this track prediction.

60km MPAS initialized from the GFS analysis at 2026-09-23T18

For the 60km MPAS initialized from the GFS analysis at 2026-09-23T18, 91W will approach northen Taiwan and continue going more northwards. While it has already moved relatively more west (124E) than EC’s prediction (which goes up around 128E), its track prediction is closer to ECMWF’s and those DLWP models. Due to its long distance from the approaching storm, 91W has not move southwards when it approched Taiwan.

mpas_d60km_0923T18

30km MPAS initialized from the GFS analysis at 2026-09-23T18

One question is: does the model resolution of 60km too coarse that it influences the move speed of those storms, so their interactions were too late, contributing to the more “correct” answer? To rule this out, another experiment was conducted using the 30km MPAS. Simliar results are shown here, that its track prediction is pretty consitent with the 60km MPAS prediction.

mpas_d30km_0923T18

60km MPAS initialized from the GFS analysis at 2026-09-24T00

For the 60km MPAS initialized at 2026-09-24T00, it even agreed more with the ECMWF and those DLWP models. It predicts that 91W will deviate in front of Taiwan and move northwards.

mpas_d60km_0924T00

By contrast, GFS’s 91W from this cycle is still struggling with its interaction with the other storm. This struggle continued, still being seen in the cycle 2026-09-24T12.

Due to the model instead of DA? Not sure.

Since GFS is based on FV3, I checked NASA GMAO’s GEOS-FP forecasts initialized from 2026-09-24T00, which shares similar dynamical core as the current GFS. As shown below, actually GEOS-FP’s prediction is more close to the ECMWF and my MPAS run. However, it is worth pointing out that GEOS-FP uses its own DA system and observation streams to generate the analysis thus its initial condition is different from the GFS analysis.

It will be interesting if GFS (GEOS-FP) uses the analysis from the GEOS-FP (GFS) as the IC to make a forecast, which can answer more questions.

geosfp_0924T00.gif

citation: figure downloaded from GMAO’s GEOS-FP website

Additional things to check for DA (Added on Sep 27, 2026)

Given that MPAS produces very different track predictions for 91W when initialized with the GFS analysis, it would be worthwhile to conduct another experiment to understand DA’s influence. Note that the GFS analysis does not initialize the hydrometeors in MPAS for my run. During my Ph.D. research assilimating surface precipitation for TC prediction, I noticed that updating hydrometeors during DA could increase TC intensity more easily, even in many cases it should not. We know that GFS-predicted track of 91W was strongly influenced by the storm following it, partly because the coming storm was so intense and large while moving fast. It would be helpful to do another experiment using GFS, initialized with the GFS analysis at 2026-09-23T18 but retaining the background hydrometeor fields instead of updating them.

(Added on Oct 21, 2026)

How are you, Discover?