Monte Carlo Geometry Processing | Hacker News Reader

Monte Carlo Geometry Processing

cs.cmu.edu

111 pointstosh16 comments

For what it's worth this was an interesting HN post on use of volumetric Monte Carlo ML to design build a better basketball backboard :

Show HN: A basketball hoop to maximize shots that go in [video] - https://news.ycombinator.com/item?id=22898653

This problem seems to be getting a lot of research lately. The problem being how to process PDEs on complex domains because meshing is a bitch. There seem to be a number of good approaches, NURBS Fea, this, Lawerence Livermore and Caltech just come out with a polygonal mesher.

I wonder how tractable and accurate they are from the solver side. Because generating high quality meshes really is a pain in the ass.

Why do we need to mesh them?

In order to apply finite element approximation approaches.

Can't you apply the model approaches directly?

The problem is gaining traction because it was used in the Covid 19 sim that was publicly released recently https://github.com/mrc-ide/covid-sim/blob/master/README.md

There was even a thread about it that made its way to the front page of HN yesterday but was flagged without an explanation https://news.ycombinator.com/item?id=23109211

I don't think they are very related.

What else would cause the sudden interest?

It was announced on Twitter just a couple of days ago: https://twitter.com/keenanisalive/status/1258152669727899650

This research would have been well underway before Covid-19 appeared.

It's a very interesting area, and has been for a while. Just look at the neat pictures in the article! I did my PhD on mesh-free approximations in the 2000s, and this kind of research has been picking up steam for decades as people exhaust possibilities for meshing and gain computing power.

As I started to read the paper, I got stuck at literally first 3 words beneath the title: "real-world geometry". Does that term mean anything? Geometry (as in meshes) is a map, not the territory. Does the expression real-world geometry actually imply the output of 3D scanners?

Anyone got any thoughts on how this might be interesting in terms of realtime rendering and visual effects?

It mentions the main cost being the construction of bounding volume hierarchies which I think is what RTX has got support for in silicon?

I think this is applicable to fluid simulation.

would this work on maxwell's equations?

The current approach outlines in the paper covers only linear elliptic PDEs as far as I can tell, so not without some more work.

Such beautiful diagrams.

Reply on news.ycombinator.com