Ah right, digits of pi isn't exactly like monkeys typing randomly for infinite time, it might be more similar to actual monkeys typing where there's no guarantee they will actually type random patterns.
Well, for some things it is. The main drive here is that for engineering applications and for predictions of physics, it may not be the right tool and it probably shouldn't be used alone
@tpg2114 Can fortran even do machine learning? All I know is that it's quite old. My numerical methods textbook always mentioned it, but we never actually used fortran.
Like, 90% of the time it can identify a dog. But we need airplanes to work 99.999999999% of the time
@JMac Well, Fortran is a (mostly) general purpose language, so it can do anything C or Python or whatever can do. And old Fortran, like F66 or F77, is horrendous and I can't read it.
@tpg2114 Ah okay. I knew it was like the standard for numerical at one point. I think my professor was biased against it because she wasn't great with it.
The supercomputer folks got confused when I asked for Boost to be installed on their systems 5 years ago -- they said they didn't know anybody using their machines that used anything other than Fortran
I can't read old Fortran. It's like trying to read Chaucer or Shakespeare in their contemporary English.
@tpg2114 I think a very different problem for engineering/science applications is that we usually want to understand how a result follows from the inputs, in the sense that a human can at least in principle explain the reasoning. That's not really possible for most types of machine learning.
@ACuriousMind Yeah, they are arguing about that now -- the counter point is that for a general, turbulent flow over an arbitrary body, we can't reason that very well either
@JMac Well, some are using it to generate reduced order models -- so get a similar answer, much faster. Some are using it to add resolution to low-order solutions. Others are using it to generate models for unused/unknown terms in things like RANS or LES
@alarge SPEC had it, but that doesn't mean the users were running codes in it. Our flow solver is part of SPEC, but it isn't the flow solver we actually use -- it's a toy version extracted to show great scaling
I'm just thinking of all the weird FEM mechanical designs they can come up with, like strangely organic shaped rims and stuff that look more like bone structures than traditional structures that they get by optimizing FEM for minimum mass or whatever. It seems like ML could help that process quite a bit.
@ACuriousMind I think part of the problem is that the gurus like to think they know what will happen based on X+10 years of experience, but that may not be true
@JMac Yeah, that's more ML for exploring design spaces -- you would use some ML algorithm to find some optimal solution, but you still have to use your FEM or whatever tool to evaluate the design generated by the ML algorithm
We're doing that now -- we had an engineer design a baseline combustor the "old" way to use as the starting point, then we found the Pareto front based on our inputs/outputs and are currently building the baseline, an optimal, and a non-optimal so we can test them
@tpg2114 Not that I really understand what's going on, but I guess that does kind of look amenable to machine learning to an extent. To sample the solution space somehow and thus make the optimizing faster. But I guess that's kind of what you were saying earlier.
Yeah -- that's how we're using it now. I want to replace and/or augment the RANS simulations with another ML algorithm to do faster estimates of the flow field
@tpg2114 you ever see this paper on Super Slo-Mo? It uses optic flow to create interpolation frames to convert regular video into super slow-motion video :D
@enumaris I haven't seen that specifically, but my summer intern last summer used some convolutional neural networks to in-fill frames in time from experimental data. I guess it's basically the same idea, we had 10 kHz movie and we upscaled it to 1 MHz
I guess that turns a super slo-mo to a hyper slo-mo :)
@ACuriousMind actually explainable AI is much harder for neural-types of ML. Traditional ML algorithms tend to be surprisingly explainable... e.g. linear regression is pretty easy to explain
Trees also tend to tell you exactly how their nodes are split
which can basically reduce them to a bunch of if-else statements
Yeah. Although, that's the key to turbulence if we can do it right. Large-eddy simulation (LES) coarse-grains the solution and then models the small scale turbulent effects on the coarse-grained solution. One of the assumptions that comes from Kolmogorov's hypothesis is that the small scales are universal
So if we could upscale our coarse-grained solution, we can get much better answers (in theory)
I would guess that if you just think of Neural Networks as powerful function approximators capable of modeling quite non-linear functions, you might be able to find some nice applications.
