FLUID SIMULATION BASED ON MATERIAL POINT METHOD WITH NEURAL NETWORK

• Main Author: Akbar Dwikatama, Pandu
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/35912
• Institution: Institut Teknologi Bandung

To produce a real-time interactive fluid simulation, high framerate is needed. In previous research, deterministic fluid simulation algorithm based on Material Point Method or MPM had been optimized. It was optimized by parallelizing some of the computations by using GPU cores. Despite using parallelization, they’re eventually limited by the available hardware at the time. To allow more opportunities to optimize fluid simulation even more, without relying only on the hardware capability, a new probabilistic approach is introduced in this research. This research discusses the implementation of neural network on fluid simulation. The implication on its visual and performance will also be discussed, along with the potential of a further optimization in future research. Material Point Method or MPM was used as the ground-truth to provide training data. The data was used to train neural network model. Then the Neural network with calibrated weights was used to interpret some calculations in MPM algorithm. Those calculations are the calculations of particle acceleration. Particle acceleration was possible to be predicted by observing nine nearby grid nodes. By removing the actual calculations, the algorithm steps could be simplified. This method changes the simulation from deterministic simulation into predictive simulation. This research was conducted in two steps. First step was to implement the training of neural network. The second step was to work on the neural network implementation on the algorithm introduced in this research, which is called as NNMPM. Algorithm in NNMPM is a modified version of its state-of-the-art algorithm, MPM. Simulation was implemented using Cinder Library using C++. The result was tested both on CPU and GPU. The implementation on GPU was done using Compute Unified Device Architecture (CUDA). The neural network was trained using FANN Library, while the inference process was also implemented using FANN (on CPU). But because FANN library doesn’t support CUDA on GPU, inference process was also implemented as a matrix multiplication. In this research iv both calculation methods of inference process were tested to compare their performances. The result of this research showed that neural network implementation on a modified version of MPM algorithm could deliver a realistic visual of fluid simulation. It was concluded because NNMPM successfully met the realistic visual criteria used in this research. The criteria of realistic visual of fluid simulation are fluid movement to the lower plane, fluid movement to fill its container, ability to achieve stable fluid without any ripple, and ability to achieve stable fluid without any fluid flow. In the scenario which was used to get training data, NNMPM could meet all the criteria. In the other scenarios conducted in this research, NNMPM could meet at least two or three of the criteria. In term of performance, NNMPM showed its potential to increase fluid simulation framerate. Despite the lower framerate it achieved from matrix multiplication method compared to MPM, using FANN library to calculate the neural network inference actually showed higher framerate by 12.271%. The FANN library was still computed on CPU. For the future research, the neural network inference process could be optimized even more by efficiently calculate the matrix multiplication or by taking advantage of the GPU cores to parallelize the calculation.