search

Ali-Stanford / PhysicsInformedPointNet

Implementation of physics-informed PointNet (PIPN) for weakly-supervised learning of incompressible flows and thermal fields on irregular geometries
8 stars 2 forks source link

Higher Dimensional PointClouds #5

Open armbrusl opened 10 months ago

armbrusl commented 10 months ago

Dear Ali,

I was wondering if you have some insights into a problem I have been facing when it comes to Physics-Informed pointNets. Do you know if the PointNet you used in your paper could also be adapted to take higher-dimensional (3–4) point clouds as an input? Specifically, I tried to encode the material variation, in addition to the geometries, using the point clouds. So the obstacle may have some material that differs from the surrounding media. The points within this box are also taken for the computation of the loss (sensor and pde). I have had some success with this, but generalization to unseen domains, that are to 'different' from the training domains, is a big issue at the moment.

All the best, and thank you for your time.

Ali-Stanford commented 10 months ago

Hi, I am happy that you used PIPN. I did not try PIPN beyond 2 dimensions; however, I imagine computationally it could be quite expensive. I am glad to hear that you obtained some success for a higher dimensional problem. Generalization is always a big challenge! -Best, Ali

On Wed, Jan 17, 2024 at 3:08 AM Leon @.***> wrote:

Dear Ali,

I was wondering if you have some insights into a problem I have been facing when it comes to Physics-Informed pointNets. Do you know if the PointNet you used in your paper could also be adapted to take higher-dimensional (3–4) point clouds as an input? Specifically, I tried to encode the material variation, in addition to the geometries, using the point clouds. So the obstacle may have some material that differs from the surrounding media. The points within this box are also taken for the computation of the loss (sensor and pde). I have had some success with this, but generalization to unseen domains, that are to 'different' from the training domains, is a big issue at the moment.

All the best, and thank you for your time.

— Reply to this email directly, view it on GitHub https://github.com/Ali-Stanford/PhysicsInformedPointNet/issues/5, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOBYOZOD2OUQCSDCNH2G7ETYO6WLRAVCNFSM6AAAAABB6LACTSVHI2DSMVQWIX3LMV43ASLTON2WKOZSGA4DKOJVGU2DGOA . You are receiving this because you are subscribed to this thread.Message ID: @.***>

armbrusl commented 10 months ago

Hello, I appreciate the quick response. Thanks, Leon