Possible to summarize a simulation into a simpler simulation?

[ScrollPortals]

Is it possible to summarize a complex & heavy process simulation into a simpler simulation with rougher-estimations?
Is there a way to programmatically analyze the particle physics of a simulation? So that they could be converted into a simpler summarized simulation?

For instance. Look at our equations to describe the flow of water. How do we come up with those? What do we measure? Can we do the same for other liquid substances?

 

[anorlunda]

Is it possible to summarize a complex & heavy process simulation into a simpler simulation with rougher-estimations?

In circuits, we sometimes represent a complex circuit by a Thevanin equivalent. I suspect that is too simple for you.

Your question can't be answered unless you state the type and degree of simplification.

 

[newjerseyrunner]

When there are resiculous numbers of variables, I occasionally train a neural network to give me approximate answers, but they aren't exact and it's a black box.

People use maths approximations all the time. Quake had a brilliant but simple square root approximator that made the whole game significantly faster. I often create look up tables for trig functions.

 

[ScrollPortals]

You're right.
I guess things can be simplified to different degrees.

Unfortunately, the system I'm trying to simplify is always going to be different every time the application loads.

I'll worry about scaling up my simulation's universe & simplifying massive amounts of particles when I actually get to that point.
Right now I have a lot of other things to worry about and program.

 

[newjerseyrunner]

I actually did my graduating thesis on simulating fluids with particles, it might help you.

My biggest optimization was when two particles were moving in the same direction at the same speed and touched each other, they would merge. If a collision had enough energy and the particles involved were not the smallest particles, it would split.

I also simulated deep water with dynamically partitioned fields that the particles layer on top of.

The simulation did not look exactly like the same simulation run with real particles, but it was very close. My fields and dynamic particles simulation baked on my laptop while my real particle simulation required a rack of servers.

 

[ScrollPortals]

Those all seem like manual optimizations. It seems like you picked those things out yourself and optimized them.
I was hoping for an automated way to do that.

 

[anorlunda]

Those all seem like manual optimizations. It seems like you picked those things out yourself and optimized them.
I was hoping for an automated way to do that.

Ignore that the application is simulation. You have a piece of software. You are asking if there is automated software that can examine other software and write a better version of it. That is a level of AI which we have not achieved yet.

 

[newjerseyrunner]

Yeah, that would require a machine learning algorithm. The best one of those we have is a convolution neural network. It can come up with some amazing solutions to problems, but I'm not sure how helpful it would be for this.

My optimizations were coded by hand, but which particles merged with the field or popped out of it, and whether particles merged or broke was all algorithmic. I just launched the program and let it run for a few days, I never touched it.