Members: Sinan Wang, Zebin Guo

Supervisor: Prof. Komura, Taku

---

Nowadays, with the advance of Metaverse, character animation in computer graphics has become a heating topic, with the splitting of directions of kinematic-based and physics-based animation. Given the complexity and real world meaning of fluid simulation, Significant work has been done in two respective fields. The work of Stream Function and DayDreamer has shown the potential of introducing the stream function into vorticity simulation, and the feasibility of building a fully differentiable model, respectively. As there are no notable combination between these two techniques, huge amount of potential influential work in the future is expected.

Therefore, we propose to have “physical-based character animation in fluids” as our final year project. We expect to solve the problem of the intractability of vorticity simulation in fluids with the help of differentiable world model trained by supervised learning model. As mentioned above, the combination of vorticity simulation and differentiable physical engine would be a significant step in fluid simulation, creating a new mode of building physical solver. By developing a discrete physical solver, we adopt supervised learning to make it differentiable. Then with the help of the differentiable world model, we can easily train the character for more verisimilitude animation purpose. It is expected that a well-developed differentiable solver capable of simulating numerous tasks, and a few well-trained characters can be provided in the final phase. We plan to finish the preliminary investigation before September and give out the first discrete fluid solver by the middle of November. The complete differentiable world model is expected to show up in the next February, and the final product will be delivered in next may.