Does anyone know why they don't use more modern directional representation systems, like curvelets, or shearlets?
My understanding is that they provide optimally sparse representation of this type of data - data that's smooth except on these wavefront sets, and that wavelets perform comparatively poorly.
Maybe a huge amount of legacy code using wavelets that's been debugged and experimentally validated. CFD code checking is laborious and errors often very subtle, and so tends to be pretty conservative.
In case anyone's interest was piqued by the complicated reaction mechanisms involved in even relatively simple combustion problems, you can automatically generate these mechanisms using software called "Reaction Mechanism Generator" http://rmg.mit.edu/ produced by a group in MIT's department of Chemical Engineering.
You wouldn't want to have to determine parameters and pathways for a 53 species 300+ reaction mechanism except in a highly automated fashion.
I don't know why the title was changed. The talk is by Adam Lichtl and Stephen Jones.
It features computer fluid dynamic (CFD) simulations of combustion in an accessible manner, though naturally the details are quite complex under the cover.
Keywords, for people wanting to dip their toes: discrete wavelet transform, space partition schemes (BSP, quadrees, octrees), space-filling curves.
The point lookup technique was a little fuzzy and I'd like some more details if possible.
The HN guidelines ask submitters not to change titles except when they are misleading or linkbait. Accordingly, we often change an edited title back to language used by the article itself—usually the original title, sometimes a subtitle or other heading. One reason for this is that we can't judge the accuracy of most edits.
HN titles also tend not to include the names of authors or speakers. We learned years ago that the site works better when the focus is on content, not personalities.
(Submitted title was "SpaceX's GPU-Powered Combustion Simulation – Adam Lichtl and Stephen Jones")
I am not sure there is a good "intro" for a lot of this per se since many of these ideas tie to research topics in applied math. I will give a few books that seem to touch on areas related:
* You first need a basic (or not so basic) numerical analysis course.
* For wavelets: Mallet's "A Wavelet Tour of Signal Processing"
* For spectral methods in general, a primer many people prefer is Trefethen's "Spectral Methods in MATLAB". It is introductory and you really should start here, but for its reliance on MATLAB and a matrix point of view it's not entirely practical.
* For GPU computing, a slightly advanced book is "The CUDA Handbook: A Comprehensive Guide to GPU Programming"
* For general-purpose parallel CPU computation: "Using MPI" by Gropp
A good general introduction to the topic is Strang's "Computational Science and Engineering" (http://bookstore.siam.org/WC07/) - basically numerical mathematics, but with applications in simulations for science and engineering.
My understanding is that they provide optimally sparse representation of this type of data - data that's smooth except on these wavefront sets, and that wavelets perform comparatively poorly.
http://en.wikipedia.org/wiki/Shearlet http://en.wikipedia.org/wiki/Curvelet