real-time physics solver
Moderator: Moporators
real-time physics solver
i've been lately developing a new way to simulate physics (currently mainly fluids and sand, some other stuff under progress) using machine learning for real-time scenarios (games, ..)
here are some videos:
http://youtu.be/4ibq68ZspRI
various materials (I guess water looks the worst):
http://youtu.be/Ye9MhVwNxrc
comments?
here are some videos:
http://youtu.be/4ibq68ZspRI
various materials (I guess water looks the worst):
http://youtu.be/Ye9MhVwNxrc
comments?
[carebox]
Re: real-time physics solver
I have no idea about machine learning, but that looks pretty decent! I agree water looks the worst. I think this is mainly because splashes look unrealistic - they resemble a fluidic mist, rather than a bunch of droplets, which is what water splashes normally look like (or is it the lack of reflections?).
But some of the denser and more viscous fluids already look much more realistic, with droplets or patches forming and interacting (milk!). Sand also looks very decent, I wish games had this simulated at least to some extent!
Self-developed project or part of something bigger?
But some of the denser and more viscous fluids already look much more realistic, with droplets or patches forming and interacting (milk!). Sand also looks very decent, I wish games had this simulated at least to some extent!
Self-developed project or part of something bigger?
Team TR
Multi WR in Labyrinth with GRob
Best Internal Total Times, Pipe stats & Pipe archive
World kuski map, World Cup stats
Re: real-time physics solver
milk & water are exactly the same (both simulation and surface generation algorithm), except that water is semi-transparent and milk lacks foam
[carebox]
Re: real-time physics solver
Maybe it's that foam that makes it look unrealistic then? I think I can see why foamy splash may be valid sometimes, but seems like that's the case all the time in the pool video - the foam dominates droplets a lot. How does it look without teh foam?
Team TR
Multi WR in Labyrinth with GRob
Best Internal Total Times, Pipe stats & Pipe archive
World kuski map, World Cup stats
Re: real-time physics solver
it is a rendering issue, that means it doesn't matter - every system already has some rendering implemented, and typically much better than the one used here
[carebox]
- insane guy
- Kuski
- Posts: 1673
- Joined: 22 May 2002, 20:53
- Contact:
Re: real-time physics solver
awesome stuff mila!
water and one or two other fluids kinda look like are in a much smaller scale than the objects. Like you're watching the ocean from high up. This gives nice surreal touch (especially to first video) but prolly nat intended...
Anyway, can't wait to play games with real time physics like this!
water and one or two other fluids kinda look like are in a much smaller scale than the objects. Like you're watching the ocean from high up. This gives nice surreal touch (especially to first video) but prolly nat intended...
Anyway, can't wait to play games with real time physics like this!
"Every night, me go to sleep, me have wet dream..."
Re: real-time physics solver
the pool should be approx 6m diameter, so it so kinda biginsane guy wrote:water and one or two other fluids kinda look like are in a much smaller scale than the objects. Like you're watching the ocean from high up.
[carebox]
Re: real-time physics solver
If you don't import those physics in elma, I'll be disappointed.
Canada
Re: real-time physics solver
so the innovation here is real time? or has it been done before?
Re: real-time physics solver
It's cool and it is beyond the scope of anything I know how to do myself. But I can't help but wonder, how will you make money from this? Some part of me thinks you are the type of person who shares things because you have a future business plan in mind. I could be wrong of course, I only know you through occasional internet comments.
God Bless America
Re: real-time physics solver
they will probably sell/license this tech for use in games. ive seen similar stuff, for example this: https://www.youtube.com/watch?v=Jl54WZtm0QE was made in 2011, what is different with your thing? 6 years in this sort of tech is a long time and it doesnt look that different visual acuity-wise. much less processing power needed or is it just that using machine learning for this is a novel approach or what?
Re: real-time physics solver
yes, the only comparable real-time library is NVIDIA flex, and that one is 10-100x slower depending on the size of simulation (their dependency is not linear)Zweq wrote:so the innovation here is real time? or has it been done before?
[carebox]
Re: real-time physics solver
that's restricted to heightmap + particles generated for a short time on the surface, online in this method ( https://www.youtube.com/watch?v=1o0Nuq71gI4 ), which has already been used in some gamesVermin Supreme wrote:they will probably sell/license this tech for use in games. ive seen similar stuff, for example this: https://www.youtube.com/watch?v=Jl54WZtm0QE was made in 2011, what is different with your thing?
[carebox]
- nick-o-matic
- Kuski
- Posts: 872
- Joined: 27 Jan 2007, 12:53
- Location: Lappeenranta
Re: real-time physics solver
What is the machine learning part? Is it for finding a new state satisfying the Navier-Stokes equations (or probably some simpler equations since full scientific accuracy is not needed for games etc) after the timestep?
Linearity sounds impressive :)
Linearity sounds impressive :)
Re: real-time physics solver
Be careful milagros. milagrosnet. Skynet. Terminator who rides on a bike.
[i cruise, i lev, i live]
Re: real-time physics solver
yesnick-o-matic wrote:What is the machine learning part? Is it for finding a new state satisfying the Navier-Stokes equations after the timestep?
[carebox]
- nick-o-matic
- Kuski
- Posts: 872
- Joined: 27 Jan 2007, 12:53
- Location: Lappeenranta
Re: real-time physics solver
Awesome. Is your machine learning method applicable to solving any set of tightly coupled non-linear differential equations? And is this method something you have been working on in university and then began to apply it to fluid simulations?milagros wrote:yesnick-o-matic wrote:What is the machine learning part? Is it for finding a new state satisfying the Navier-Stokes equations after the timestep?
Re: real-time physics solver
yes (I wouldn't say any, but at least for a large class of problems) and yesnick-o-matic wrote:Is your machine learning method applicable to solving any set of tightly coupled non-linear differential equations? And is this method something you have been working on in university and then began to apply it to fluid simulations?
[carebox]
Re: real-time physics solver
Looks cool, I don't know a lot about physics, but I do know a lot about moving water. How did you think/work with the surface tension of the water in this? Is it calculated in at all? For instance, if you drop a ball in the water with speed, it would hit and land hard. Or if the water splashes then the drops forming would be separated from each other and quickly "close up", rather than sticking together... if that makes any sense.
Team HotHorses, and I'm converting to Icelandian now...
Re: real-time physics solver
not sure what kind of answer do you expect if you say you don't understand physicsniN wrote:Looks cool, I don't know a lot about physics, but I do know a lot about moving water. How did you think/work with the surface tension of the water in this?
it is used only for the "mercury" (parameters are fake) - at the scale used there surface tension is irrelevant and typically ignored (for example in all movies)
[carebox]
Re: real-time physics solver
looks super cool. would love to see it being used in some dark experimental sandbox game.
Re: real-time physics solver
tried to do a nice scene https://www.youtube.com/watch?v=nW_gd1Kity0
[carebox]
- Grace
- 38mins club
- Posts: 4856
- Joined: 19 Nov 2005, 10:45
- Location: Deep in your Imagination, Twirling your Dreams and Weaving your thoughts.
Re: real-time physics solver
I know you say surface tension is insignificant and therefore ignored, but realistically the cause of surface tension (hydrogen bonding between water molecules) is not insignificant on any scale and ignoring it is probably a significant part of the reason for the difference between real water physics and your water simulation. That said, it's significantly more difficult to emulate something such as hydrogen bonding in such a program, and I think you've done an absolutely phenomenal job simulating this stuff with your own program.
Cyberscore!
___________________________________________________
Targets: 6 Legendary, 23 WC, 20 Pro, 5 Good | AvgTT: 39:59:96
___________________________________________________
Targets: 6 Legendary, 23 WC, 20 Pro, 5 Good | AvgTT: 39:59:96
Re: real-time physics solver
the significance of each term of navier-stokes equations in any scale is quantifiable, so if it's only a fraction of a percent, it is insignificantHaruhi wrote:I know you say surface tension is insignificant and therefore ignored, but realistically the cause of surface tension (hydrogen bonding between water molecules) is not insignificant on any scale
also when you simulate galaxies, you don't need anything except gravity, even though stars would not hold together without strong nuclear force or electromagneticm
in any case, as I already mentioned the framework allows for surface tension and it is used for other materials, where it is significant (such as mercury)
[carebox]
Re: real-time physics solver
Is it deterministic? Let's say you want to make a physics puzzle game with premade objects that should work the same way everytime. For example object A: water container, object B: sandcastle, object C: tube that transfers water from A to B
Re: real-time physics solver
Milagros you are a figure head of the elma scene and probably a scholar, we are very interested in your project. I was wondering, how will this make money?
God Bless America
Re: real-time physics solver
well.. the algorithm is deterministic, CPU implementation gives always the same simulationZweq wrote:Is it deterministic? Let's say you want to make a physics puzzle game with premade objects that should work the same way everytime. For example object A: water container, object B: sandcastle, object C: tube that transfers water from A to B
however on GPU, there are some parallel summations done with atomic increments and the sum has a floating-point error depending on the order it was summed
then the epsilon error propagates and after a while it is a different simulation
I guess, it could be fixed by using fixed digit precision
Last edited by milagros on 5 Apr 2017, 12:46, edited 1 time in total.
[carebox]
Re: real-time physics solver
by selling the library/plugins to various game/simulation enginesgimp wrote:how will this make money?
[carebox]
- Grace
- 38mins club
- Posts: 4856
- Joined: 19 Nov 2005, 10:45
- Location: Deep in your Imagination, Twirling your Dreams and Weaving your thoughts.
Re: real-time physics solver
milagros wrote:the significance of each term of navier-stokes equations in any scale is quantifiable, so if it's only a fraction of a percent, it is insignificantHaruhi wrote:I know you say surface tension is insignificant and therefore ignored, but realistically the cause of surface tension (hydrogen bonding between water molecules) is not insignificant on any scale
also when you simulate galaxies, you don't need anything except gravity, even though stars would not hold together without strong nuclear force or electromagneticm
in any case, as I already mentioned the framework allows for surface tension and it is used for other materials, where it is significant (such as mercury)
Taking the most conservative values I could find for the enthalpy values of hydrogen bonding between OH --- O, the standard enthalpy total of water and the average instantaneous number of hydrogen bonds between molecules in a mass of pure water all at Standard temperature and pressure suggests that the electrostatic value of hydrogen bonding in pure water vapour should be about 12.07% of the total electrostatic attraction, although other research puts the instantaneous number of hydrogen bonds much higher, leading to a potential value as high as 18.38% of total electrostatic attraction.
This, as well, is just in water vapour (gaseous state). I'd anticipate at least somewhat of an increase in this electrostatic attraction in liquid state, hence surface tension. I could be doing my calculations incorrectly - and I'll openly admit to not having studied navier-stokes equations or fluid dynamics in any great detail as a biologist and chemist rather than physicist, but I think 12% probably counts as significant for water.
Aside from the fact that there is an entire biological family of arthropods (Gerridae, or Pond Skaters) with the defining characteristic being their use of water surface tension, you can also easily observe it even in a glass of water - which wouldn't be possible if it was only a fraction of a percent's worth of significance towards water dynamics.
Cyberscore!
___________________________________________________
Targets: 6 Legendary, 23 WC, 20 Pro, 5 Good | AvgTT: 39:59:96
___________________________________________________
Targets: 6 Legendary, 23 WC, 20 Pro, 5 Good | AvgTT: 39:59:96
Re: real-time physics solver
surface tension clearly scales with surface (n^2) and other forces with volume (n^3), thus as the simulation size increases, the relative importance of surface tension goes to 0
the representative scale, at which surface tension starts to lose its importance (a few times more and its unimportant), is related to the size of water drops
the representative scale, at which surface tension starts to lose its importance (a few times more and its unimportant), is related to the size of water drops
[carebox]
Re: real-time physics solver
it's fucking waterfall!
https://www.youtube.com/watch?v=jjYHqakTrWY
https://www.youtube.com/watch?v=jjYHqakTrWY
[carebox]
Re: real-time physics solver
https://vimeo.com/282230568
if you have at least GTX1070, you can download the demo
if you have at least GTX1070, you can download the demo
[carebox]