polymathic-ai/UNetConvNext-rayleigh_taylor_instability

Benchmarking Models on the Well

The Well is a 15TB dataset collection of physics simulations. This model is part of the models that have been benchmarked on the Well.

The models have been trained for a fixed time of 12 hours or up to 500 epochs, whichever happens first. The training was performed on a NVIDIA H100 96GB GPU. In the time dimension, the context length was set to 4. The batch size was set to maximize the memory usage. We experiment with 5 different learning rates for each model on each dataset. We use the model performing best on the validation set to report test set results.

The reported results are here to provide a simple baseline. They should not be considered as state-of-the-art. We hope that the community will build upon these results to develop better architectures for PDE surrogate modeling.

CNextU-Net

Implementation of the U-Net model using ConvNext blocks.

Model Details

For benchmarking on the Well, we used the following parameters.

Parameters	Values
Spatial Filter Size	7
Initial Dimension	42
Block per Stage	2
Up/Down Blocks	4
Bottleneck Blocks	1

Trained Model Versions

Below is the list of checkpoints available for the training of CNextU-Net on different datasets of the Well.

Dataset	Learning Rate	Epoch	VRMSE
acoustic_scattering_maze	1E-3	10	0.0196
active_matter	5E-3	156	0.0953
convective_envelope_rsg	1E-4	5	0.0663
gray_scott_reaction_diffusion	1E-4	15	0.3596
helmholtz_staircase	5E-4	47	0.00146
MHD_64	5E-3	59	0.1487
planetswe	1E-2	18	0.3268
post_neutron_star_merger	-	-	-
rayleigh_benard	5E-4	12	0.4807
rayleigh_taylor_instability	5E-3	56	0.3771
shear_flow	5E-4	9	0.3972
supernova_explosion_64	5E-4	13	0.2801
turbulence_gravity_cooling	1E-3	3	0.2093
turbulent_radiative_layer_2D	5E-3	495	0.1247
viscoelastic_instability	5E-4	114	0.1966

Loading the model from Hugging Face

To load the UNetConvNext model trained on the rayleigh_taylor_instability of the Well, use the following commands.

from the_well.benchmark.models import UNetConvNext

model = UNetConvNext.from_pretrained("polymathic-ai/UNetConvNext-rayleigh_taylor_instability")