Using Graph Neural Networks to Learn Fire and Smoke Behaviour in a Physics- Based Simulator
Ladda ner
Typ
Examensarbete för masterexamen
Program
Publicerad
2021
Författare
Lam, Edvin
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Deep learning is prominent in computer vision as a means of learning spatial features
and processing them in useful ways, such as for image classification and image
segmentation. An application of this in physics is to use videos of fire to predict
how the flames and smoke evolve over time. This thesis explores a simplified version
of this in an attempt to lay the groundwork for future studies. In particular, simulations
of fire and smoke were made using the open source software Fire Dynamics
Simulator by NIST. From these simulations, frame by frame states of the physical
system could be extracted. A graph neural network using the MeshGraphNets
framework was trained to predict future states of the fire given two past states as
input. Some of the resulting neural networks were able to reasonably predict over
50 frames into the future and also showed some competence when faced with test
sets with a different physical geometry. The performances of the trained models varied
greatly depending on the loss functions used in training, suggesting that more
hyperparameter optimisation could be done in this regard.
Beskrivning
Ämne/nyckelord
graph neural network , fire dynamics simulator , deep learning , physics simulation