mdribeiro / deepcfd

DeepCFD: Efficient Steady-State Laminar Flow Approximation with Deep Convolutional Neural Networks

Mateus Dias Ribeiro 👋

Based in Braunschweig, Germany.

• 🔭 I’m a Staff Scientist at the German Aerospace Center (DLR)
• 🌱 Currently working on reduced order models for fluid dynamics via machine learning

deepcfd's Issues

Multi-region label

Dear, sir

Thank you so much for publishing this fantastic works. I went through the code, and have a small concerned:

How do you compute the multi-region label? Did you do it directly in OpenFOAM, or any other software?.

Once again, thank you sir for considering me! I'm really appreciated if you guide me through these question. Hope you all the best

result.json file contains weights?

Hi, mdribeiro:
Really appreciate your excellent work. May I ask that the result.json file contains the well-trained model's weights?
Thank you very much
Monica

Information leakage

First of all, thank you for making this repository available. I have learnt a lot through it so far. However, I found something troubling while looking at the code in this line

I would say that computing channel weights from all the available data cause information leakage from the test/validation set to creep into the model. Ideally, the channel weights should be computed from only the training set, which stems from the assumption that the model generalizes well on the test set just by looking at the train set.

Create MLP model

Create MLP model to allow the solution of other types of non-image data.

Final Trained Weights

Hi there,
Thanks for uploading this, I am exploring the code now and was wondering if you could provide the finalized weights of the model. I don't have the CPU to run the training unfortunately but would like to explore the result.
Josh

Class imbalance in this dataset

Hey guys,
I was exploring your code and found a huge class imbalance in the dataset. Your testing split contains 2samples which contribute almost everything to the MSE. Leaving those two out would reduce your achieved MSE of 2.04 to around 0.3. The CFD simulation of those samples looks totally different than the rest. In total there are about 6 or 7 of those samples in your dataset. Since I guess this is not intended I wanted to let you know :)
Best regards
Alex

How can I prepare own dataset?

Hello,

First of all thank you a lot for sharing your model.
I've read the paper and tested the model with the data you've provided but I wonder what would be the results if the model was trained with another data or bigger dataset?

In the paper you've stated that the dataset was obtained with simpleFOAM simulator but you've also used some random geometry generator and the shell script to save the data into the form which is then passed to the model.

Is there any chance that you could also share the data generation script or guidelines?

Best regards,
Adam

Create PINN Model

Create Physics-informed NN Model for problems that are governed by partial differential equations.

Create Tests for DeepCFD

Tests should be created to make sure the repository grows safely.

Layers per block in DeepCFD.py

Hi there,

Thank you for publishing this! I have been looking through the code and noticed that the number of layers per block was not specified in your DeepCFD.py script.

model = UNetEx(3, 3, filters=filters, kernel_size=kernel_size, batch_norm=bn, weight_norm=wn)

This results in the layers per block defaulting to 3 as defined in UNetEx class. I was just wondering if this was intended as this differs from your paper's implementation of 2 layers per block. Thank you again for this!

Kind regards,
Sean