Physics engines are used in many fields in engineering such as computer vision and robotics to name a few. However, the dynamics of fluids or soft-bodies such as fabrics are difficult to simulate, computationally expensive and even inaccurate for complex physic engines. To cope with some of these limitations, learned simulators using machine learning methods are an interesting alternative.
The physical properties of a system can be represented by particles that interact with each other. In order to learn the interactions and propagate information throughout the particles we can represent the system as a Graph Neural Network (GNN).
The thesis goal is to develop a Graph Neural Network to learn to simulate multiple physical systems such as fluids, soft-bodies and rigid-body systems.
- Review relevant state-of-the-art literature.
- Design multiple simulations, such as fluids or soft-bodies, using physics engines (Nvidia Flex/Taichi language).
- Development of a Graph Neural Network.
- Evaluation of the physics simulated by the GNN.
Mrowca, D., Zhuang, C., Wang, E., Haber, N., Fei-Fei, L. F., Tenenbaum, J., & Yamins, D. L. (2018). Flexible neural representation for physics prediction. In Advances in neural information processing systems (pp. 8799-8810)
Sanchez-Gonzalez, A., Godwin, J., Pfaff, T., Ying, R., Leskovec, J., & Battaglia, P. W. (2020). Learning to simulate complex physics with graph networks. arXiv preprint arXiv:2002.09405.