Ice was created in the context of my MSc at McGill University and Mila. The application is a standalone MPM solver written in C++/CUDA. It uses an embedded 3D convolutional neural network to asynchronously identify active regions of the simulation. The experimental results show performance improvements in practical scenarios with no deterioration in worst cases. Although it is not a plugin, a small set of HDAs were created to allow scene creation in Houdini directly. All simulations were rendered with Karma XPU. The following table shows, in video order, the different simulations along with their specifications and performance metrics.

Scene	Particle Count	Grid Resolution	Timestep (×10-4)	Sec/Frame	Adaptive Speedup
Ship Breach	54.0M	492 × 212 × 492	6.9	16.6	3.38×
Bullet Time	24.1M	688 × 364 × 492	4.2	17.9	1.05×
Avalanche	35.5M	736 × 148 × 760	21.0	7.5	1.53×
Train Push	24.9M	520 × 144 × 1748	2.6	28.8	1.55×
Rally Drift	14.9M	1400 × 196 × 536	1.1	40.7	2.00×
Glacier Collapse	39.5M	320 × 184 × 520	6.9	16.4	1.00×

Details regarding this project can be found in the thesis (coming soon...):