Ryan Keisler open-source model GNN dự báo thời tiết toàn cầu - 10 ngày trong dưới 1 phút

TL;DR

Ryan Keisler vừa open-source model từ bài báo "Forecasting Global Weather with Graph Neural Networks" (arXiv 2202.07575) - một trong những nghiên cứu đặt nền móng cho làn sóng AI weather forecasting hiện nay. Model dự báo thời tiết 10 ngày toàn cầu chạy dưới 1 phút, hỗ trợ khởi tạo từ ERA5 hoặc ECMWF IFS, MIT license, kèm 3 scripts demo sẵn dùng.

Sensitivity map: du bao U500 tai t+24h nhan cam nhat voi initial conditions o dau (2026-01-03). Tinh toan bang JAX autodiff tu model.

Bai bao 2022 va vi tri trong lich su AI weather

Thang 2/2022, Ryan Keisler cong bo paper tren arXiv chung minh rang Graph Neural Network co the du bao thoi tiet toan cau o do chinh xac tuong duong GFS va ECMWF - nhung nhanh hon nhieu bac va khong can mo phong vat ly truyen thong. Day la mot trong nhung paper dau tien va co anh huong nhat trong linh vuc nay.

Sau do, Google DeepMind ra GraphCast (11/2023) - cite cung tu tuong, tang do phan giai len 0.25 degree, outperform ECMWF HRES tren 90% targets. Nam 2024, GenCast tiep tuc cap nhat probabilistic forecasting, dat 97.2% tren ENS targets. Keisler 2022 la "pioneer" - coarser resolution nhung da validate toan bo approach cho ca mot field.

Nay, sau 4 nam, code duoc public hoan toan. Khong phai port hay re-implementation - day la model chinh thuc tu tac gia.

Ky thuat dang sau: 3 GNN + 78 channels

Model gom ba thanh phan GNN ket noi noi tiep:

• Encoder GNN: Chuyen trang thai khi quyen tu lat/lon grid sang H3 hexagonal mesh de xu ly
• Processor GNN: Thuc hien 9 vong message passing tren hex mesh - day la noi hoc cac pattern vat ly
• Decoder GNN: Chuyen ket qua tro ve lat/lon grid chuan de output forecast

Moi grid point co 78 channels (6 bien vat ly x 13 muc ap suat). Moi buoc du bao la 6 gio, chain nhieu buoc de ra forecast 10 ngay. Training data la ERA5 reanalysis.

Diem thu vi: sensitivity analysis dung JAX autodiff - tinh ∂forecast/∂initial_conditions de biet du bao nhan cam nhat voi vung nao tren ban do (giong nhu backprop nhung cho thoi tiet).

10 ngay trong duoi 1 phut

Toc do la diem noi bat nhat:

Ve do chinh xac: performance tuong duong GFS va ECMWF operational model khi danh gia tren do phan giai 1 degree va dung reanalysis initial conditions. Metrics chuan: Z500 (geopotential height tai 500 hPa) va T850 (nhiet do tai 850 hPa).

Case study an tuong nhat: du bao Hurricane Sandy 8 ngay - model predict dung huong di va cuong do bao trong suot qua trinh tien vao bo dong nuoc My.

3 scripts, 2 data sources, MIT license

Repo tai github.com/rkeisler/keisler-2022 kem san 3 scripts:

1. 01_evaluation.py - Chay forecast tu ERA5, tinh area-weighted RMSE, visualize
2. 02_sensitivity.py - JAX autodiff sensitivity analysis
3. 03_hurricane.py - Demo Sandy 8-day track (predicted vs actual)

Hai nguon du lieu ho tro:

• ERA5 (Google ARCO) - du lieu lich su, tot cho evaluation va research
• ECMWF IFS analysis (AWS Open Data) - ngay gan nhat, tot cho near real-time forecasting

Yeu cau: Python 3.10+. Mit license - co the dung ca cho commercial project.

Ai nen thu ngay

• Researcher / sinh vien khi tuong: Codebase clean, well-structured - tot de hieu cach xay GNN weather model tu dau
• ML engineer: Muon co baseline weather model de compare hoac fine-tune
• Developer ung dung khi tuong: Can near real-time global forecast ma khong muon goi API thu phi
• Educator: Hurricane Sandy demo la tool giang day truc quan ve ung dung GNN trong khi tuong hoc

Gioi han can biet: do phan giai 1-degree (~111 km) - tho hon GraphCast (0.25-degree). Long-term rollout (1+ nam) bi smooth va xuat hien artifact. Chi deterministic, khong probabilistic.

Tiep theo

Day la release chinh thuc tu tac gia sau 4 nam - diem khoi dau de community experiment: fine-tune tren regional data, tang do phan giai, them variables moi, hoac dung lam backbone cho cac ung dung khi tuong chuyen biet.

Nguon: arXiv 2202.07575, GitHub rkeisler/keisler-2022, Ryan Keisler tren X.