Last year, we launched a $50,000 Kaggle competition with an ambitious goal: accelerate the development of machine learning emulators for high-resolution subgrid physics in climate models. The Kaggle community rose to the challenge, and we received 10,000+ submissions from hundreds of teams hailing from six continents.

But the real question remained: Do the ideas that won on an offline test set actually work when put to the test in a hybrid physics-ML climate simulation?

Today, we are excited to release our findings in our latest paper, "Crowdsourcing the Frontier: Advancing Hybrid Physics-ML Climate Simulation via a $50,000 Kaggle Competition." We systematically tested winning architectures and ideas from Kaggle in hybrid physics-ML simulations using the E3SM-MMF climate model.

So what did we find?

🚀 A Major Milestone: For the first time, we show that online stability with minimal drift is reproducibly achievable across multiple diverse architectures.

🌟 SOTA Performance: Building on the strong baseline set by Hu et al. (2025), these Kaggle-inspired architectures achieve state-of-the-art results across multiple metrics.

🏔️ The Next Frontier: Despite large diversity in architectures, all models share strikingly similar zonally averaged biases and systematically underpredict tropical precipitation when coupled.

Our study demonstrates that providing the machine learning community with the right interface can drive climate science research forward, but this story isn’t over. We have open-sourced our code, model checkpoints, and online-testing framework so that researchers around the world can continue to shrink the gap between research prototypes and operational relevance.

📄 Preprint: https://arxiv.org/abs/2511.20963

💻 GitHub repository: https://github.com/leap-stc/climsim-kaggle-edition

🤗 Model Checkpoints: https://huggingface.co/collections/jlin404/climsim-kaggle-models

🛠️ ClimSim-Online: https://github.com/leap-stc/climsim-online

BibTeX:
```
@article{Lin2025-ko,

title = {Crowdsourcing the Frontier: Advancing Hybrid Physics-ML Climate Simulation via a $50,000 Kaggle Competition},

author = {Lin, Jerry and Hu, Zeyuan and Beucler, Tom and Frields, Katherine and Christensen, Hannah and Hannah, Walter and Heuer, Helge and Ukkonnen, Peter and Mansfield, Laura A and Zheng, Tian and Peng, Liran and Gupta, Ritwik and Gentine, Pierre and Al-Naher, Yusef and Duan, Mingjiang and Hattori, Kyo and Ji, Weiliang and Li, Chunhan and Matsuda, Kippei and Murakami, Naoki and Ron, Shlomo and Serlin, Marec and Song, Hongjian and Tanabe, Yuma and Yamamoto, Daisuke and Zhou, Jianyao and Pritchard, Mike},

journal = {arXiv preprint arXiv:2511.20963},

year = {2025},

month = {11},

url = {https://arxiv.org/abs/2511.20963}

}
```