I wrote the overall process through the Hackathon Here in Korean.
In this post, I will share the presentation content in English.

Presentation

Hello, everyone! Welcome to our presentation. I’m Yelim Kim from NVIDIA Team, ”Typhoon Sensei". Today, we are excited to share our project.

Before we begin, we would like to introduce our team members who worked together on this project.

And also our mentors who supported us throughout this project.

The expected path of typhoon 'Kong-rey' was a major societal issue indeed.
This was due to the fact that the ECMWF, which was considered the best-performing model in the field, had produced inaccurate results for the prediction of Kong-rey’s path.
Our team was eager to solve this problem.

The blue points on the figure represent the expected path produced by the ECMWF model, while the black points indicate the actual path taken by the Kong-rey. As one can notice, significant discrepancies existed between the results, and this caused unprecedented damages taken by both Korea and Japan.
This caught our attention and led us to choose our subject for the hackathon to be "Predict Typhoon Kong-rey and potentially other typhoons using FCN".

To start with, we'd like to briefly introduce the FCN model. This model uses AFNO for predictions, and this type of architecture is an enhanced version of the vision Transformer model. The mixed operations are composed of global convolutions and are effectively implemented through the Fourier transform.

The FCN model begins with projecting the input variables from ERA5 dataset on a 720*1440 latitude longitude grid onto a 2D patch grid. Each patch is represented by a d-dimensional token, and the patch sequence and location encoding are passed together onto the AFNO layer. The training process of FCN is carried out gradually, and the configuration of the FCN model we received will be explained in detail in a future slide.

Our goal for the project was to make the most out of our knowledge. Many students in our group had conducted meteorological research and had background knowledge of front-end programming as well. Most importantly, we all wanted to improve the model through the two weeks of training. Therefore, we tried to materialize it through two key ideas. They were,
1. The model is already developed but training it within the limited time is unfeasible. How can we improve this situation?
2. Is there any possibility of applying the model for typhoon forecasts?
In the following slides, we explained the problem we encountered on each topic and suggested a solution if there was one.

the Initial problem is Unsuitable data representation for maximum wind velocity. ERA5 data used is grid data, leading to average value drops. In extreme situations, the difference from the average is significant. Due to the lack of awareness of peak wind speed, danger radius comprehension is challenging, leading to underestimating wind speed values.
Applying this data to the model introduces additional prediction difficulties due to double smoothening.

The second problem was of the variables that were used during the training of the FCN model. To quote from Nvidia's official website, "FCN is a model that provides accurate short to medium range global predictions.”
In other words, the FCN model is not a model designed to predict typhoons. We found a problem with this.
As can be seen from the data on the left, variables below at least 200 hPa are required.
A way to improve this is to delete the variables that are not necessary for typhoon prediction and insert the variables at 250hPa that we need as a minimum. However, to train the model with these variables, we judged that it was impossible to finish during the hackathon period and decided to leave it for future research.

The third problem is that of the ERA5 dataset. This sentence is also quoted from the ERA5 official website “These variables, are sampled at a temporal resolution of 6 hours.”
The ERA5 dataset, which FCN used for training, has data at 6-hour intervals. We judged that this interval was far too short for forecasting and proposed retraining with a dataset with a shorter time interval as a solution.

The FCN model we got at the beginning of the hakathon has a structure on the left of the slide.
However, we found it challenging to improve the already completed model, so we tried to find another ways to approach it.
As a result, inspired by the feature of FCN - auto-regressive prediction, just here the orange box, we devised a new structure, as shown in the image on the right. We used GT data as input for each time step in the auto-regressive predictions.
Also, we combined this approach with the ensemble to present comprehensive results.

The red dots from the figure is our prediction without using an ensemble.
Just as we initially aimed for, we have successfully achieved improved results compared to the existing models.

Furthermore, we set another ultimate goal was to achieve results similar to KMA(those of the meteorological agency).
Therefore, we focused on improving the model and visualizing the results using ansemble, and also we consulted Jeff about the number of ensembles, and he suggested starting with 10, but due to the OOM issue, we had to limit it to 3 for now.
We have generated results and visualizations using these 3 ensembles, which led to positive outcomes.

Now I’ll introduce problems related to the topic of creating visualization and UI for forecasting. At first, we thought about how it would affect people based on the predicted path of the typhoon.
To do so, accurate data on the radius of influence was crucial. We selected the radius based on the Typhoon white book.
The figure on the right is the result we made from ERA5 dataset.



The second problem is processing multiple typhoons.
As one can see on the left, when the eye of a typhoon is detected, a typhoon appears on the right side, tracks the eye, and detects another typhoon, leaving the predicted path. So, we specified and classified the location value of the typhoon, and obtained the result shown on the right.

Thirdly, we thought about how to make people use this forecast on typhoons.
We let the users select the typhoon, date, location, and time. As output values, we gave a warning message on the top to inform people of the size and strength of the incoming typhoon, and what kind of precautions should be taken based on the predicted results. We also added the location of the typhoon in the corresponding time zone and the expected path predicted by the FCN model.

Now we’ll be discussing insights we gained from the Hackathon.
We gained ideas for potential research topics that we may embark on in the future.
First, The model we currently utilize is based on the FCN model published in 2022, using AFNO architecture. However, in a research paper published in 2023, SFNO-based FCN was used.
We can't use this code because it hasn't been opened to the public yet, but once it’s released, we’ll be able to get better results using it.

The next subject is the problem on variables which we mentioned earlier. If we can improve this area and train the model,
we will be able to get better results for typhoon prediction.

The following are possible improvements that can be made to the UI.
Currently, the UI is provided only for Typhoon Kong-rey, but if you want to receive real-time data, you can create a script using ChatGPT by loading the ERA5 dataset and checking U, V component wind and preciptation.
Following this process, one can receive results for cases using real-time datasets.

Lastly, each team member wrote what kind of research they could do by grafting it to FCN in their own laboratory.

Thank you for reading.
And finally, let me introduce the website our team has created.

Website

We have built the Typhoon Track Forecast Service.
Please go to the website and check the descriptions for each tab at the bottom of the main page.

Furthermore, on the last tab, you will find introductions, contributions, and contact information for our team members. If you have any points of curiosity or areas for improvement regarding our project, please feel free to let us know.

I would like to take this opportunity to express my gratitude to all the mentors, participants, and our incredible team members. Thank you all for your support and contributions.

Insight from Mentor

This is a message left by Jeff, my mentor from NVIDIA, after my presentation. If we refer to the indicators Jeff provided and use them to train the model, we may obtain better prediction results.