« first day (3390 days earlier)      last day (1531 days later) » 

hpekristiansen
hpekristiansen
8:11 AM
I know additive RGB color mixing of light. -and substance CMY(K) color mixing of ink on white paper. -BUT what if there is no white background only paint(an infinite layer on e.g. black background). What would I need to read to understand color mixing like that?
 
Novalium Company
Novalium Company
8:22 AM
morgen haigen
 
Knight
Knight
8:45 AM
What is morgen haigen?
 
Ajay Mishra
Ajay Mishra
9:06 AM
I just cannot figure out one while looking upon the derivation of electromagnetic wave equation from the maxwell equation: You apply curl operator on one of the equation and outpops the wave equation. What does this mean?
Every field is wave?
 
John Rennie
John Rennie
9:17 AM
@AjayMishra if you have any field $F(x,y,z)$ you can Fourier transform it to represent it as an infinite sum $\int G(p_x, p_y, p_z) dp$.
And those function G are just infinite plane waves.
So basically, yes, every field is made up of waves.
 
Novalium Company
Novalium Company
9:33 AM
@Knight no idea
 
Slereah
Slereah
Well, not just waves solve the wave equation
ie the field $A = x$ solves the wave equation
although that one isn't a Schwartz function
 
Knight
Knight
10:11 AM
@NovaliumCompany Greame Johnson
 
Rudi_Birnbaum
Rudi_Birnbaum
Hi there! I fail to get a clear idea of canonical and mechanic momentum (esp. in a magnetic field). My confusion starts when I fail to understand how the canonical momentum ($\bf p$) behaves under a gauge transformation of the vector potential.
I do not see why $\bf p$ should change under $\bf A\to\bf A'$
where as ${\bf p}-q{\bf A}$ (mechanic momentum) changes because of the gauge transformation.
@JohnRennie even everything?
 
Slereah
Slereah
Classical EM canonical momentum is a bit tricky because the action isn't quite invariant under gauge transformation
But the change in the action doesn't change the equation of motion, is the important part
 
John Rennie
John Rennie
@Rudi_Birnbaum well you can Fourier transform anything. Whether the result is physically meaningful is another question :-)
 
Rudi_Birnbaum
Rudi_Birnbaum
@JohnRennie but I feel really a bit wave-like this morning ...
 
John Rennie
John Rennie
 
Rudi_Birnbaum
Rudi_Birnbaum
10:25 AM
lol!!
 
Rudi_Birnbaum
Rudi_Birnbaum
10:38 AM
Or to put it in another way: If $m\bf v = {\bf p}-e{\bf A}$, what is ${\bf p}$? (I suppose ${\bf p} = \partial L / \partial \dot{\bf r}$), but then how to arrive at $L$)....
 
Slereah
Slereah
@Rudi_Birnbaum Well, what do you start with, exactly?
 
Rudi_Birnbaum
Rudi_Birnbaum
so what is the explicit expression for ${\bf p}$ here..
 
Slereah
Slereah
If you want to derive something, first, what is your starting point
 
Rudi_Birnbaum
Rudi_Birnbaum
@Slereah Good question!
Its more I like to build a mental edifice of physics and connect the pieces with each other ...
 
Slereah
Slereah
Well I advise to start from having the Lagrangian, it's certainly easier :p
Although it's not necessarily the most intuitive
 
Rudi_Birnbaum
Rudi_Birnbaum
10:42 AM
Can you give me some hint ob the meanings of canonic and mechanic momentum?
 
Slereah
Slereah
I can tell you what the canonical momentum is
What is "mechanic momentum"?
 
Rudi_Birnbaum
Rudi_Birnbaum
mechanic is p - eA
 
Slereah
Slereah
Is that the definition, or is that what it is in this case?
 
Rudi_Birnbaum
Rudi_Birnbaum
according to what I have heared.
canonic is p
not sure
 
Slereah
Slereah
Canonical momentum is $\partial L / \partial \dot{x}$, which is $p - eA$
 
Rudi_Birnbaum
Rudi_Birnbaum
10:44 AM
There is a noe paper from Fritz London and he says its similar to Energy that partitions into kinetic and potential
 
Slereah
Slereah
I'm guessing the mechanical momentum is supposed to be the Noether invariant from translation?
 
Rudi_Birnbaum
Rudi_Birnbaum
yes, makes sense
OK and that cannot be conserved in the field, right?
 
Slereah
Slereah
Well, here's something to consider : if your mechanical momentum is just $p = m\dot{x}$, your particle is going to slow down in an EM field
or accelerate or whatever
 
Rudi_Birnbaum
Rudi_Birnbaum
Then is the question how does the mechanical momentum transform under a gauge transformtion
 
Slereah
Slereah
Therefore, it will not be invariant
 
Rudi_Birnbaum
Rudi_Birnbaum
10:47 AM
@Slereah yes
exactly
 
Slereah
Slereah
It is best to apply the usual Noether procedure to find it
Apply an infinitesimal translation
and see how the fields transform
 
Rudi_Birnbaum
Rudi_Birnbaum
You mean to show the fact that the mechanical momentum is not invariant?
 
Slereah
Slereah
I mean showing what the mechanical momentum actually is
 
Rudi_Birnbaum
Rudi_Birnbaum
Well I can accept this for the moment, I am much more interested in how the mechanical momentum transforms under a gauge transformtion
 
Johan Liebert
Johan Liebert
Why doesn't the edit appear for this question?( I have the edit privilege).
 
Slereah
Slereah
10:51 AM
It should transform as $m\dot{x} - e A \to m\dot{x} - e A - e \partial \alpha$
 
Rudi_Birnbaum
Rudi_Birnbaum
right because ${\bf p} = m{\bf v} -e{\bf A}$, which appears a bit circular at this stage. So right I might have to go through Noether to understand the definition ...
@Slereah thank you!
Ah!! Now I understand, in QM you set $m{\bf v} \to -i\hbar\nabla$!!
 
Slereah
Slereah
11:11 AM
The real definition of (mechanical) momentum is the Noether current under translation
The canonical momentum is the momentum you define via Legendre transform to do Hamiltonian mechanics
 
Rudi_Birnbaum
Rudi_Birnbaum
OK. I see! Now something else: I want to see how the mechanical momentum density $$(p - eA)\rho(r,r')$$ (QM) transforms under a gauge transfromation: $A\to A+\nabla \phi$.
with $\rho$ the 1-RDM
 
Slereah
Slereah
I'm afraid I don't know
 
Rudi_Birnbaum
Rudi_Birnbaum
I obtain $$(p-eA)\rho'$$ with $$\rho'=\rho \exp{(i/\hbar c)(\phi(r)-\phi(r'))}$$
but that leaves an awkward factor of $(i/\hbar c)(\phi'(r))$ in the $p$ term ... and that would mean its not gauge independent
I must do something wrong here ...
(sorry the mechanical momentum density is $$[(p - eA)\rho(r,r')]_{r'=r}$$ , I forgot to set $r'=r$ after the application of the operator)
 
Knight
Knight
How would you even state the notion of "Lorentz invariance" "without mathematics"? — ACuriousMind ♦ May 2 '16 at 18:06
Sometimes he makes me laugh so much
 
PM 2Ring
PM 2Ring
@JohanLiebert That question has a pending edit suggestion, so that suggestion has to be approved or rejected before anyone else can edit it. The mobile site hides the Edit button on such questions.
Feb 10 at 13:53, by PM 2Ring
On the mobile site, the edit button disappears if there's a pending suggested edit. But if you go to full site, a nightmare begins to unfold. ;) As Aaron said, you have to disable responsiveness, and select the desktop site in your browser. And even then it's a major pain.
 
Knight
Knight
11:27 AM
@PM2Ring Do you know driving?
 
 
1 hour later…
Novalium Company
Novalium Company
12:32 PM
When you google "morgen haigen" this is what comes up:
 
Knight
Knight
1:03 PM
Which one of them is you @NovaliumCompany?
Anyone interested in Set theory?
I got a question to ask
 
Johan Liebert
Johan Liebert
1:50 PM
@PM2Ring Yes I saw the suggested edit and approved it. Actually I was going to do the same.
 
 
2 hours later…
Novalium Company
Novalium Company
3:32 PM
@Knight No one lol
Guys, you know how in neural nets, gradient descent gets to tweak the weights and optimize them in the most efficient way possible. Ok but the designer of the network has to choose the number of nodes in each layer (and in most complicated networks, other stuff as well...). Why don't we make a program or something like that that will automatically try different amounts of nodes on different layers so to optimize performance?
Someone should have done something like that already, it's too obvious.
 
tpg2114
tpg2114
@NovaliumCompany Those are called hyper-parameters and people do work on how to tweak/optimize those also.
 
Novalium Company
Novalium Company
@tpg2114 Cool. Like, you can attempt to minimize some type of error there as well.
 
tpg2114
tpg2114
There's also something called an ODE-Net, which recasts a NN as a continuous ordinary differential equation and uses a stiff solver to move from input layer to output layer, getting rid of all hidden layers
 
Novalium Company
Novalium Company
@tpg2114 Sorry, what does it do in simple terms? I'm still a rookie.
 
tpg2114
tpg2114
@NovaliumCompany You certainly can -- but for any network of interest, there's just not enough data and not enough computer time to do a full study over the entire design space
ODENet? It basically changes the problem from taking inputs through a bunch of hidden layers (which the user has to select beforehand) and rewrite it in the form of a continuous equation
And then you can integrate the equation using standard computer integration
It removes some of the hyper-parameters
Anyway, that paper on arXiv is pretty good. And their code is on GitHub to play with
 
Novalium Company
Novalium Company
3:41 PM
What do you mean "rewrite it in the form of a continuous equation"?
 
tpg2114
tpg2114
You'll have to read the paper to see how they do it
Instead of this big network of nodes, you rewrite it as something like dH/dt = <stuff>
And then integrate dH/dt
 
Novalium Company
Novalium Company
What does the integration do?
 
tpg2114
tpg2114
Haven't taken calculus yet eh? That paper won't make a ton of sense then
 
Novalium Company
Novalium Company
I have to review calculus yep
 
tpg2114
tpg2114
Basically they try to find the function H which takes inputs and creates outputs. And they figured out how to write the derivative of the everything between the input and output layers. So they need to integrate the derivative to find the function H that gives them the "finished" mapping
Normally a NN uses weights and activation functions and summations across hidden layers and nodes -- that's the "function" H. But they want to get rid of specifying the hidden layers, so they re-write things to get rid of those
 
Novalium Company
Novalium Company
3:47 PM
@tpg2114 As far as I know, the derivative is basically "how fast something changes" or "the slope at a point on the graph". How in the world do you take the derivative of everything between input and output?
 
tpg2114
tpg2114
You'll have to read the paper to see how they do it -- it's quite clever.
 
Novalium Company
Novalium Company
@tpg2114 Yeah, and I'm quite stupid.
I looked at it and I have no idea what's going on. I'm still 17 dude.
 
tpg2114
tpg2114
@NovaliumCompany Not stupid -- but if you haven't taken calculus, it's probably going to be hard to follow. But so will any explanation I try to give
 
Novalium Company
Novalium Company
@tpg2114 Yeah, I will review calculus in depth. After all, I want to understand gradient descent equations first.
 
tpg2114
tpg2114
You're right, the derivative is the slope of a curve at a point. And you can approximate those as linear things -- f(x+dx) = f(x) + linear_estimate -- and if your dx is really really tiny, you end up with the derivative
So they look at a traditional neural network as H(layer + 1layer) = H(layer) + some_stuff and say "What if we made the +1layer really small in 'layer space'" so they can end up writing everything as a derivative
But it's of course much more complicated than that in detail
 
Novalium Company
Novalium Company
3:54 PM
Ok. This seems like a pretty complicated and inefficient way to get rid of hyper-params, any alternatives? Like, a program that randomly tries different combinations and sees which one produces the best model accuracy with the least amount of hyper-params.
I know it's computationally inefficient but I suppose it could be optimized somehow :D
 
tpg2114
tpg2114
That would be a genetic algorithm type approach -- those are certainly used to designing of all kinds of things.
I haven't looked for them used to design hyper-parameters in neural networks, but I'm sure somebody has done it
 
Novalium Company
Novalium Company
@tpg2114 What if nobody has done it? Dude, go build it and get your millions.
 
tpg2114
tpg2114
Haha, I just got back from a workshop on machine learning and I came away with 2 conclusions -- 1) the field is too mature to just go in and show something cool and not be able to explain the theory; and 2) the is way too immature to know which approaches will actually become useful long term
 
Semiclassical
Semiclassical
Given how universal the problem seems, there’s probably no one-and-done solution
 
tpg2114
tpg2114
So it seems to me a tough one to break into (from a research perspective) because it could be a whole lot of work to do to justify an approach that may not be useful/general
 
Novalium Company
Novalium Company
3:59 PM
Ok, I'mma get the mils then
 
tpg2114
tpg2114
From an industry or commercial perspective, it's a little easier -- explaining the theory isn't important, and it doesn't have to be a general solution for all things if it increases profit now
 
Semiclassical
Semiclassical
Right. One doesn’t need an “optimal” solution, just a satisfactory one
 
Novalium Company
Novalium Company
I don't care about fields and industries and so on. I wanna build something useful, and honestly when you choose hyper-params for your model, I believe your decision is based on the data and the network. If you can take it, why not let a machine analyze your stuff and do it better... donno... just thoughts.
 
tpg2114
tpg2114
@NovaliumCompany Look for a genetic algorithm package in python, or look for a multidisciplinary design and optimization package (OpenMDAO is one I use) and let it rip on your hyper-parameters. Could be a fun little study
 
Novalium Company
Novalium Company
@tpg2114 yeah, might do it.
But I started learning ML seriously literally 2 weeks ago.
 
tpg2114
tpg2114
4:02 PM
@NovaliumCompany That's a very true statement in all areas, not just ML. I run simulations of turbulent combustion, but I have to choose my domain, choose my grid, choose my models, choose my chemistry, choose my time steps, etc.. The final output is sensitive to literally every single choice I make.
And we've seen that two different people, with the same tools but making different choices, can get completely different answers
If we could get a computer to figure out how to choose the best parameters, we could actually make predictions instead of just matching (poorly, often) experiments
 
Novalium Company
Novalium Company
@tpg2114 You can choose your input data, but why bother with the between input-output stuff? A program can try random genetic combinations and sees which one performs the best.
 
tpg2114
tpg2114
@NovaliumCompany Because our simulations take 10,000 processors 6 months to run a single realization
 
Novalium Company
Novalium Company
fuk me dead that's true
 
tpg2114
tpg2114
There isn't enough computing power in the universe to optimize them
 
Semiclassical
Semiclassical
(Also, when I use the word “satisfactory”, I have in mind the following concept: en.m.wikipedia.org/wiki/Satisficing)
 
Novalium Company
Novalium Company
4:04 PM
Well, how do you choose your network structure?
 
tpg2114
tpg2114
@NovaliumCompany We're not using ML for these, we're using traditional flow solving approaches. Our "network" is basically the mesh we use to solve the equations -- and that's basically just from experience
 
Knight
Knight
@tpg2114 Hello! How are you? Seeing you after long time sir
 
tpg2114
tpg2114
We have to kinda know what the flow is going to do, at least roughly, to design a mesh. So it's a whole lot of engineer experience
 
Novalium Company
Novalium Company
Cool. How other people in the field choose their nn structure?
 
tpg2114
tpg2114
@Knight I'm good -- I was traveling for... 2 weeks for work
@NovaliumCompany Same way -- experience. They have solved related problems before and have a feel for what the minimum is needed
 
Knight
Knight
4:06 PM
@tpg2114 Travelling? Where?
 
tpg2114
tpg2114
Or, they go the deep learning approach and assume if you throw enough layers and nodes at it, it will work out in the end
@Knight Out the Mojave desert in California and then to Texas for a workshop on machine learning
 
Novalium Company
Novalium Company
@tpg2114 Based on what factors do people decide the structure of their network?
 
tpg2114
tpg2114
@NovaliumCompany It's going to depend on the problem you are trying to solve.
 
Semiclassical
Semiclassical
I wouldn’t be surprised if there’s a general principle at play: for any structure, there’s a set of problems they do badly at
 
tpg2114
tpg2114
If you are just trying to approximate a function, you only need one really wide layer (based on the Universal Approximation Theorem)
 
Knight
Knight
4:08 PM
@tpg2114 I thought you were a real theoretical physicist but you also have turned to Computers 💻
Real means doing only physics
 
Semiclassical
Semiclassical
Which would be fine, if it was obvious which structures do best for which problems
 
tpg2114
tpg2114
If you are trying to identify features in an image, layers can often represent certain things like detecting edges, finding separate regions, etc..
But that was discovered by a lot of trial and error
 
Novalium Company
Novalium Company
It'd be cool if we could make a global dataset of people's projects and their network structures so others can see how they do and choose a nn sturcture for themselves. Or you could even run that global dataset through a network so people just input their input data and it outputs a nn structure xDDD
 
tpg2114
tpg2114
@Knight I've always only used computers -- and a lot of pen and paper to derive things. My background is entirely numerical!
@NovaliumCompany image-net.org at least for image recognition
 
Semiclassical
Semiclassical
I imagine a big issue is just that it’s not so obvious how to classify “problems”
 
Novalium Company
Novalium Company
4:11 PM
A neural network structure clustering neural network xD
 
tpg2114
tpg2114
The contest section collects all of the references for the methods the team submits
 
Semiclassical
Semiclassical
If you narrow your range of problems, then I imagine it’s easier
But universal problem-solving is hard
 
Novalium Company
Novalium Company
But possible :D
 
tpg2114
tpg2114
@NovaliumCompany If you haven't started flipping through Deep Learning yet, it's a great book that's free online -- deeplearningbook.org
It's got suggestions for network design depending on problem type
 
Knight
Knight
@tpg2114 Means you are computational Physicist? And you use computers only for computations? Or are you working on topics of Computer Science?
 
Novalium Company
Novalium Company
4:12 PM
I just want to get rid of those hyper-parameters. A program can be made that will do it instead of me. It can analyze my dataset, compare it to other people's similar datasets and choose an nn structure for me, something like that.
 
tpg2114
tpg2114
@Knight I guess I'm a computational physicist -- really a computational fluids specifically. But that's not done in a bubble. I advise and analyze experimental work, I derive theoretical models to convert to computation models or verify computational results. And I have to write the code that does all of this work, so there's a lot of computer science involved also
 
Semiclassical
Semiclassical
The main issue Id see is that “figuring out what kind of problem you have” is a problem in itself
 
tpg2114
tpg2114
Codes to run on modern supercomputers aren't something you can buy somewhere -- they all have to be written
 
Semiclassical
Semiclassical
And that smacks of the halting problem, which truly is insoluble
 
tpg2114
tpg2114
@NovaliumCompany deeplearningbook.org/contents/guidelines.html page 427 has a section on automatic hyperparameter selection
 
Novalium Company
Novalium Company
4:16 PM
@tpg2114 coolio
 
Knight
Knight
@tpg2114 Oh my God! Means you do everything on your own. You develop a model, you write a code for computation and then run it and records the results for experimental purpose, WOW!
 
Novalium Company
Novalium Company
I believe that the most efficient neural network structure can be determined entirely from the input data. What do you think about that?
 
Knight
Knight
What about taking me as an assistant?
Paid assistant
 
tpg2114
tpg2114
@Knight Well, we build on the work of others -- I use a code that has about 30 years of development work into it, the last 10 of which I've been involved in
@NovaliumCompany Depends on what you include as "input data" -- you also need to know what output you want, and how reliable the output needs to be
If you include that as input then I guess? But that's just lumping everything about the problem into something called "input data" which isn't the usual nomenclature
 
Semiclassical
Semiclassical
Also, I imagine the phrase “most efficient” is itself a can of worms
Most efficient by time? Memory?
 
Novalium Company
Novalium Company
4:23 PM
Why not try to make a program that will take the project's input and desired output data and based on that alone (e.g if it's image dataset, number of images, image sizes...) will determine such a nn structure that produces the highest accuracy. (and you can maybe add an option for a balance between accuracy and memory efficiency)
 
tpg2114
tpg2114
@Knight We take paid interns or co-ops all the time. It's US citizen only, and we need people who are at a university and studying engineering or related programs
@NovaliumCompany Well, you just added a hyper parameter -- you are selecting for "highest accuracy." But what if I just want a ballpark answer and I want it really fast?
 
Novalium Company
Novalium Company
@tpg2114 What do you mean ballpark answer?
 
Semiclassical
Semiclassical
Make a computer that can predict what your goal is :P
 
Novalium Company
Novalium Company
@Semiclassical yeah cul
 
tpg2114
tpg2114
Like what if I don't want to know exactly where in the image something exists. I just want to know that the something is there. Or I want to know the position +/- 10 meters or something
The network structure for all of those might be different, even though it's all just "find something in an image" problem
 
Novalium Company
Novalium Company
4:26 PM
Be right back, gotta get food.
 
Semiclassical
Semiclassical
As a non-NN example: I had a piece of code that, as a certain number n increased from 1, it would take longer and require more memory
 
tpg2114
tpg2114
Or... maybe it's for a self-driving car, which means I need to be 99.99999999999999% certain the network is correct. But maybe it's for a shopping recommendation, and so I only need to be like 51% certain the network is correct.
 
Semiclassical
Semiclassical
To the point where I could run it on my laptop for n=5 in 2 hours and it’d top out at 10 gigs (or w/e the number was)
 
Knight
Knight
@tpg2114 Do you wear specs? (I mean those powered one not the goggles or glasses)
 
tpg2114
tpg2114
@Knight I'm not sure what you are referring to
 
Semiclassical
Semiclassical
4:29 PM
Now, suppose I wanted to know how long it’d take for n=6. Then I don’t really care if the answer is correct to the second. I care if the answer is one day vs one year
 
Knight
Knight
@tpg2114 Do you have myopia or hetero myopia and do you wear specs for correcting that?
 
Semiclassical
Semiclassical
(And even computing that, turns out to be hard. It’s easier than the main problem but it’s still a nightmare: the upper/lower bounds are so weak as to be useless)
 
tpg2114
tpg2114
@Knight So do I wear glasses? No
 
Loong
Loong
@Semiclassical That's like my currently running MCNP simulation. I hope it will finish after four days.
 
Knight
Knight
@tpg2114 I thought working on computers would have hurt your eyes like mine.
 
Semiclassical
Semiclassical
4:32 PM
“Hope” being a key word there
 
tpg2114
tpg2114
Well -- I probably need to get glasses. But I haven't. Last vision test I had was 20 years ago and I was 20/20 then, so until somebody tells me otherwise, I'm still 20/20 ;)
Since I can say with honesty "Last time I was tested!"
 
Loong
Loong
Yes, I made a quick test and extrapolated from there.
 
Semiclassical
Semiclassical
The saving grace for my case was that, after computing the hard part, I had a much easier computation to do after
 
tpg2114
tpg2114
@Loong There's a whole lot of value in extrapolation.
Which is one area that machine learning fails spectacularly
 
Semiclassical
Semiclassical
And when I did some research, I found an approach which let me skip the hard part entirely
Still a bit disappointing—the output of the hard part was not uninteresting—but the output of the easy part was sufficient for my purposes
 
Knight
Knight
4:36 PM
@tpg2114 This means that you're around 40 years or so now :) ?
I meant your age
 
Semiclassical
Semiclassical
(And the fact that it turned a 2 hour computation into a 2 minute one also made it easy to accept that bargain)
 
tpg2114
tpg2114
@Knight Not quite, unless you round really generously. I just haven't been tested since early high school
 
Knight
Knight
@tpg2114 Ah! you're in mid thirties :)
Most probably 36
I think your age is between 34 to 37 :)
 
Semiclassical
Semiclassical
Oook
 
Luyw
Luyw
4:50 PM
hello
If I have made a measurement with an uncertainty u, then if the value I am interested in is the inverse of that measurement, how can I find the uncertainty for it? it is not the inverse of u, right?
 
tpg2114
tpg2114
@Luyw Do you mean inverse like 1/M?
where M is your measurement
 
Luyw
Luyw
yes
 
Novalium Company
Novalium Company
I want a neural network for good grades.
 
tpg2114
tpg2114
So the simplest way to do it would be to write it as 1/(M +/- u), which will be equal to something else
 
Luyw
Luyw
I need to state my inverse adequately. Should I partial fraction that?
 
tpg2114
tpg2114
4:55 PM
Like 1/(M +/- u) = MI + u' where u' is something else
 
Loong
Loong
@Luyw If you only need a quick estimate, you can use the relative uncertainty u_rel = u/M
 
Luyw
Luyw
5:12 PM
so I have m with uncertainty u(m), I have just been reading that I can use propagation of errors as follows, u(1/m)=sqrt( [d/dm(1/m)*u(m)]^2 ). would this be the correct approach?
 
Semiclassical
Semiclassical
5:23 PM
ugh, why can't I get this experiment to work
 
Loong
Loong
@Luyw With "d/dm(1/m)" you mean "df/dm for f=1/m"?
 
Luyw
Luyw
5:45 PM
yes
 
Semiclassical
Semiclassical
ahah, I think i figured it out.
 
tpg2114
tpg2114
@Semiclassical Experimenting with anything fun?
 
Semiclassical
Semiclassical
rubber band thermodynamics :)
essentially, this demo: youtube.com/watch?v=ovVO8NDdon4
my problem was that I was seeing the opposite effect as expected
and what I think I'm seeing is that, in order to see the desired effect, you need a higher extension ratio not a small one
 
tpg2114
tpg2114
Huh, that's counter-intuitive
 
Semiclassical
Semiclassical
the other thing I'm seeing online is that it depends a lot on what kind of rubber you use :/
 
tpg2114
tpg2114
6:00 PM
Also don't jive with my expectations based on Young's Modulus behavior with temperature -- for rubber, it stays flat as T increases until you start to melt the thing
Although maybe it's only "approximately flat" since it's on a log scale and has values of 10**9
 
Semiclassical
Semiclassical
long quote incoming:
"In Gough–Joule effect, the temperature of rubber increases considerably when it is stretched to large extensions. But for small extensions (< 50%), a slight decrease in temperature is observed. This phenomenon is called thermoelastic inversion [5]. However, for small deformations, the temperature of rubber increases during compression and does not even change during shear or torsion.

Also, generally materials exhibit a positive coefficient of linear expansion. But for rubber, this coefficient changes from positive to negative depending upon its extension. Stretched rubber shrinks on heati…
So presumably the point is that, in the demo, the extension was high enough that it exhibited shrinking (and therefore increased tension)
But, when I was trying before, the extensions were small enough that I saw decreased tension and therefore expansion
so I think that explains it: with low (or no) extension, rubber really does expand like you'd expect for other materials
but at high extension, it instead contracts.
 
vzn
vzn
6:39 PM
lol nice work! phd physicist scratching his head over rubber band heating dynamics and we can also be sure QM is complete :)
 
Semiclassical
Semiclassical
oh hai non sequitor
also, note that I was scratching my head precisely because -I- didn't know what other people already knew
once I found what other people already did know, I stopped being confused
 
Physics Meta
Physics Meta
1
Q: flagging vandalism

Norbert SchuchIf I see a post which is vandalism (like this one, only visible to 10k users), what am I supposed to flag it as? (Concretely, I remember both "spam" and "rude or abusive" flags being rejected. Do I need to raise an "other" flag and spend time writing a reason?)

support flagging best-practices
 
 
2 hours later…
Novalium Company
Novalium Company
8:27 PM
DJ KHALEEEED...........................................................anadaone
 
 
2 hours later…
amanuel2
amanuel2
10:22 PM
WeirdChamp...
 

« first day (3390 days earlier)      last day (1531 days later) »