Abstract by Samuel Giraud-Carrier

Personal Infomation

Presenter's Name

Samuel Giraud-Carrier

Degree Level


Abstract Infomation


Computer Science

Faculty Advisor

Seth Holladay


Automated Backward Smoke Simulation


Art directability is a crucial aspect of creating visually pleasing
visual effects. Effects artists confirm there is still a major lack
of art-directable simulation tools. Using only physically based 
simulation techniques can make it difficult to match a desired 
composition. Time-reversed simulation has shown 
to be a possible solution to this issue. However,
because the backwards mapping between simulation timesteps is ill-defined,
we propose the use of a deep learning architecture to solve this problem. 

With their capacity to approximate complex relationships between data
samples without knowing the underlying structure, deep learning methods
seem to be an intuitive approach to this problem. We propose a machine
learning algorithm which, when given a frame of 
simulation data, predicts the previous frame in a simulation. 
Used in sequence, we can effectively produce a frame range of 
reverse simulated data. When played forward, we predict our model 
will produce visually plausible results.