Modulus CHT PINN MLP

Modulus CHT PINN MLP

Logo for Modulus CHT PINN MLP
Description
Parameterized PINN that uses MLP layers and incorporates PDEs in the loss function. You can incorporate Fourier transform based encoding for input layers for other variations of the network architectures.
Publisher
NVIDIA
Latest Version
latest
Modified
May 26, 2023
Size
325.58 MB

Architecture

PINN MLP is the base neural network architecture for learning directly from Partial differential equations.

Parameterized PINN that uses MLP layers and incorporates PDEs in the loss function. You can incorporate fourier transform based encoding for input layers for other variations of the network architectures.

The pretrained model checkpoint comes from the 3D Three Fin Heatsink example as described in the Modulus documentation.

Training

The model was trained on a parameterized 3D Fin geometry using the Navier-Stokes and heat flow governing equations without the need for training dataset.

How to use?

You can use the inference script to run the model and vary the parameters of the design. You can use the training script as a starting point to train for a different geometry. Refer Modulus Documentation for more details.

Input

The model accepts spatial co-ordinates and value of the design parameters such as the fin height and thickness. For a complete list, Modulus Documentation.

Output

The model outputs the flow quantities, i.e., u, v, and p and temperatures in solid and fluid at the given input co-ordinate and design point.