Researcher and PhD candidate @ RWTH Aachen University. I work in the intersection between math, physics and software. Simulating the macroscopic world ๐๐ฆ
@chenna1985@NikhilYewale93 Depends how you define AD, but symbolically differentiated compile-time generated code for derivatives can be very close to manually optimized derivatives in our experience. To the point where bottlenecks are elsewhere...
@istvan_csanady Implicits many fundamental drawbacks though. For example, if you have to surfaces that are very close together, they will tend to merge with implicits unless you have some way of knowing that there's a gap. There are also fundamental issues with using implicits for simulation
@chenna1985 Aren't LLMs fundamentally unable to innovate? That's been my experience. It's useful for assisting with already solved problems but it's never ever been of any help at all for new problems that presumably hardly anyone else has ever looked at.
@chenna1985 ... will anyway be dominated by the areas where you have the highest resolution (in both space and time), whereas the coarser elements are likely to cost you much less. So it seems to me that you'd need a very special scenario for it to be worth the complexity...
@chenna1985 I agree. Are they actually ever worth it though? The primary use case, to my knowledge, is the ability to use local adaptivity in time as well as space. But, most often when you need temporal adaptivity you anyway need spatial adaptivity, and the overall computation cost ...
@heskelbalas@zhou_xian_@j_d_noone@jessfraz Looks like it's "visually accurate", i.e. it gives visually realistic simulations for the most part. But I wouldn't expect any decent force measurements (which you'd typically need for CAD-related stuff) since this requires entirely different simulation methods
@chenna1985 I'm also working on something CutFEM-like and I'm also writing my own code because it's nice to have the flexibility, and not have to work against the limitations imposed by FEM libraries that were not designed for it
MPM continues to be the GOAT multi-phase physics solver, and I always like to point out / brag that it was introduced into computer graphics by @DisneyAnimation.
@chenna1985 Huh, so there's really nonvalidation in any of the papers? That's unfortunate if it's th case...
Not super familiar with LBM, would you be willing to elaborate what makes it unsuitable?
@chenna1985 Yeah, I've always thought it makes more sense to use NN to somehow speed up conventional simulation rather than replacing it outright. i don't think NN alone could ever give enough accuracy and reliability
If there's one thing AI is *actually* good at (but not perfect), it's translation. This comes at a good time for the EU โ which traditionally has been limited by language barriers โ as European countries need to cooperate more than ever to secure their own interests
@chrisoffner3d Imo graphics is at a totally different level compared to engineering / applied math / physics. I just can't take more of those painful red/blue MATLAB plots ๐
@chrisoffner3d One of the reasons I'm pursuing a PhD in physics simulation in computer graphics instead of engineering or physics is because graphics people really care about presentation, including content presentation in papers