Tibetan Plateau Mountain Wave Simulation Using AI-Driven 3D Adaptive Mesh Refinement

April 12, 2026

A new paper titled “Tibetan Plateau Mountain Wave Simulation Using AI-Driven 3D Adaptive Mesh Refinement” has been published in Journal of Geophysical Research: Atmospheres. The study was led by Pu Gan (Beijing Normal University & CAS), with co-authors Jinxi Li (CAS), Xiaofei Wu (Chengdu University of Information Technology), Qizhong Wu (Beijing Normal University), Xun (Jerry) Zou (SIO), Zifa Wang (CAS), Jiang Zhu (CAS), Huifeng Yuan (CAS), Feng Xie (Anhui meteorological observatory), Xiao Tang (CAS), Leisheng Li (CAS), Fangxin Fang (ICL).

The study contributes to CW3E’s Strategic Plan (Atmospheric Rivers and Extreme Precipitation Research, Prediction, and Applications) and advances AI applications and model development in high-mountain regions.

Accurately simulating orography-induced mountain waves over steep terrain, such as the Tibetan Plateau (TP), remains a major challenge for numerical weather prediction (NWP) models due to grid distortions inherent in traditional terrain-following coordinates. To address this issue, this paper developed an AI-driven adaptive mesh refinement (AMR) framework within the Fluidity-Atmosphere model, which employs a 3D unstructured mesh to mitigate geometric distortions. A Long Short-Term Memory (LSTM) neural network is integrated into the AMR process, replacing traditional adaptation criteria with data-driven predictions. A series of idealized 2D and 3D experiments demonstrate that both the traditional AMR and LSTM-driven approaches reproduce mountain wave dynamics more efficiently than fixed meshes. Furthermore, the LSTM-based method suppresses numerical noise near complex terrain, preventing spurious over-refinement. In realistic simulations over the TP, the LSTM-enhanced model successfully captures the full life cycle of mountain waves, accurately reproducing key physical features such as vertical velocity structures, wave amplitude decay, and upstream phase tilt. Comparative tests reveal efficiency gains of up to 71.4% over fixed meshes and 23.8% over traditional AMR at high resolution, along with improved accuracy in vertical velocity, potential temperature, and wave propagation. These results validate the LSTM-based AMR framework as a robust and efficient approach for atmospheric simulations over complex terrain. By intelligently allocating computational resources while preserving physical fidelity, this method provides a scalable pathway toward next-generation atmospheric modeling, with future applications targeting realistic meteorological conditions over the TP (Figure 1).

Persistent precipitation biases in NWP, particularly in magnitude and location, highlight the need for improved terrain representation. Future work will focus on tighter integration between AI-driven adaptive meshes and precipitation-related parameterization schemes. Expanding training datasets to include complex topography and diverse high-impact events will support more stable, realistic, and computationally efficient simulations of terrain-forced atmospheric processes.

Figure 1. Comparison of vertical velocity differences and topographic features (low-resolution). (a) Vertical velocity difference: adaptive mesh refinement (AMR) mesh–Fixed mesh; (b) Vertical velocity difference: AI mesh–Fixed mesh; (c) Elevation at y = 80 km and (d) Slope difference (AMR–Fixed) at y = 80 km comparing results from the fixed mesh and adaptive mesh. From Figure 9 in Gan et al. (2026).

Citation:

Gan, P., Wu, X., Wu, Q., Zou, X., Wang, Z., Zhu, J., Yuan, H., Xie, F., Tang, X., Li, L., & Fang, F. (2026). Tibetan Plateau Mountain Wave Simulation Using AI-Driven 3D Adaptive Mesh Refinement. Journal of Geophysical Research: Atmospheres, 131(6), e2025JD045585. https://doi.org/10.1029/2025JD045585