FluidMLP: A general method for learning Lagrangian fluid simulation

Du, Fengshan; Ban, Xiaojuan; Zhang, Yalan; Guo, Yu; Obaidat, Mohammad S.

doi:10.1016/j.simpat.2022.102623

Cited by 1 publication

(1 citation statement)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In Lagrangian fluid simulations, FluidMLP [32] was proposed to generate a model by learning particles' physical properties. This model, which learned and compared particles' positions, velocities, and accelerations, improved accuracy notably when using acceleration data.…”

Section: Fluid Simulation With Deep Learningmentioning

confidence: 99%

Isoline Tracking in Particle-Based Fluids Using Level-Set Learning Representation

Kim,

Song,

Lee

et al. 2024

Applied Sciences

View full text Add to dashboard Cite

In this paper, we propose a learning model for tracking the isolines of fluid based on the physical properties of particles in particle-based fluid simulations. Our method involves analyzing which weights, closely related to surface tracking among the various physical properties of fluid particles, are significant. These weights are used as input values for the learning algorithm, enabling relatively accurate isoline tracking. In addition, compared to existing learning models such as linear regression, LSTM (long short-term memory), and learning representation (1-layer) models, our method obtained superior surface tracking results without accumulating errors. By using our proposed network structure to track the fluid surface, it learns and predicts values derived from existing fluid simulation algorithms, eliminating the need for computational processes for level-set values and enabling real-time surface tracking. As the scale of the simulation increases, our method significantly reduces the time and resources consumed compared to traditional methods and can track the fluid surface without additional resource consumption. Furthermore, due to our method’s simple network structure, the time consumed in the initial process of loading the model into memory is faster than models such as CNN and LSTM. Our proposed model occupies less than 30 kb of storage space, making it suitable for use in middleware. Lastly, to verify the generality of our method, we conducted tests in a total of five scenes, and in all test scenes, visually natural fluid isolines were represented.

show abstract

Section: Fluid Simulation With Deep Learningmentioning

confidence: 99%