Although there are numerous variants of NeRF, most are primarily focused on the task of novel-view synthesis.
Directly integrating existing NeRF models for graphics simulation yields unsatisfactory results.
Specifically, we found the geometry of current NeRF models is not sufficiently accurate to enable physics-inspired interactions, and they cannot provide a realistic and complete illumination field to relight the inserted physics entities, such as water and snow.
To address this challenge, we have systematically investigated various techniques to improve NeRF models, focusing on answering the critical research question of "what constitutes a good simulatable NeRF representation?".
ClimateNeRF&nbs