Physics-Informed Learning of Characteristic Trajectories for Smoke Reconstruction

Abstract

Figure 1: We present a physics-informed fluid reconstruction method using a novel Neural Characteristic Trajectory representation to preserve both short-term physics constraints and long-term conservation. In the challenging scene with smoke and obstacles, our method reconstructs decomposed radiance fields, obstacle geometry (serving as boundary constraints for smoke), smoke density, velocity, and trajectories from sparse-view RGB videos, and generates realistic renderings of novel views.