Project

Abstract: This master thesis explores ways to apply geometric deep learning to the field of numerical simulations with an emphasis on the Navier-Stokes equations. With the recent success of Deep Learning, there should be room for experimentation also in the field of fluid simulations. Here we lead such an experiment. We propose an end-to-end differentiable architecture that allow object-to-mesh predictions of fluid simulations. We provide a comparison with a baseline and visual results on three different datasets: airfoils, backward facing steps and winged drones.

Week 18 (17/06 - 21/06)

21/06 Is the deadline for the written report

Week 17 (10/06 - 14/06)

Week 16 (03/06 - 07/06)

Week 15 (18/05 - 31/05)

Week 14 (18/05 - 24/05)

Week 13 (13/05 - 17/05)

Week 12 (06/05 - 10/05)

Week 11 (29/04 - 03/05)

Week 10 (22/04 - 26/04)

Holydays

Week 9 (15/04 - 19/04)

  • WE –> marriage
  • Friday: Holyday

Week 8 (08/04 - 12/04)


Almost half!!

Week 7 (01/04 - 05/04)

  • WE –> Greyssonay

Actually I am here below


Week 6 (25/03 - 29/03)

  • WE –> stay in Lausanne

Week 5 (18/03 - 22/03)

  • WE –> pdg with Auriane

  • Changed to Theo’s code and re-read surface data.

Week 4 (11/03 - 15/03)

  • WE –> Crans with Auriane

  • Friday
    • tried to install fineopen with no success but with a good understanding of what was what. Openned a request of assistance
  • Thursday
    • Morning –> IP
    • Created the data in an efficient way
  • Wednesday
    • Try to create adjacency
  • Tuesday
    • played with paraview and tried to understand what was going on with Pierre’s model
  • Monday
    • Check results of the net

Week 3 (04/03 - 08/02)

  • WE –> StartHack

  • Thursday
    • Morning –> IP (Image Processing)
    • numpy2vtk function and visualize firsts results
  • Wednesday
    • catch up IP
    • NDA agreement
  • Tuesday
    • Made the net work
  • Monday
    • Made net work on toy data and trying to make it work on real data

Week 2 (25/02 - 01/03)

  • WE –> France
  • Friday:
    • Have a non-working neural-net on the graph
    • not working but will repair it Monday
  • Thursday:
    • Went to Reinforcement Learning
    • Start writing the net & prepare the data to be fed to the net –> did SVM to do a baseline
    • ~See Andrea for numerical problems~ worked late + Auriane instead

How to coarsen the graph?

  1. Michael’s methods changes the nodes :’(
  2. Sequentially with numpy? (it is a bit aleatoric…)
  3. Random (–> bad idea?)
  4. Do we really care? at the end its still a graph…
  5. How to coarsen the surface itself?

In the following A is the weighted Adjacency matrix (i.e. W in some notations)

  • Wednesday:
    • Received email from Numeca & tried to get started installing Fine/OpenLabs
  • Tuesday:
    • Read the data as .npy
    • moved the code on the server
    • Finished fine-tuning of vim
  • Monday:
    • Extracted VTK data
    • got started with vim

      Week 1 (18/02 - 22/02)

  • Checked the code pygsp and cnn_graps
  • read some papers
    • Name the papers
  • Before I did a presentation on numerical methods for PDE’s with a focus on CFD