I am wondering if we can gather enough data showing objects interact with each other is it possible to train AI to learn laws of physics. For example if we record behaviour of an siolated system involving few objects and use that data to build model such that it predicts motion for next few frames. Is it already in work somewhere? I am an amatuer so I would love to learn from you. 

This was actually done years ago, before we called it AI.

https://news.cornell.edu/stories/2009/04/computer-derives-natural-laws-observation

“The researchers have taught a computer to find regularities in the natural world that represent natural laws -- without any prior scientific knowledge on the part of the computer. They have tested their method, or algorithm, on simple mechanical systems”

I think there were other efforts, and there are more recent examples

Thanks for sharing. It sounds like what I was thinking about but given advances in computation power, image processing and general AI, would it be worth to try again?

1 hour ago, Mahi_sayli said:

Thanks for sharing. It sounds like what I was thinking about but given advances in computation power, image processing and general AI, would it be worth to try again?

If you did a search I’m sure you’ll find efforts from the last several years.

On 3/31/2025 at 8:51 AM, Mahi_sayli said:

I am wondering if we can gather enough data showing objects interact with each other is it possible to train AI to learn laws of physics. For example if we record behaviour of an siolated system involving few objects and use that data to build model such that it predicts motion for next few frames. Is it already in work somewhere? I am an amatuer so I would love to learn from you. 

Doesn't OpenFoam have a function for thermodynamic and hydrological physics? Or try simscale

On 4/29/2025 at 12:17 PM, Sohan Lalwani said:

Doesn't OpenFoam have a function for thermodynamic and hydrological physics? Or try simscale

Aren't those simulation software? That does not seem what OP is asking about. Most simulation software I know (and it may have changed) you define the models, add the desired input and boundary conditions and let it provide solutions (or approximations of it). From how I understand OP the idea is to take actual data, generate a model from it and then run it.

On 3/31/2025 at 5:51 PM, Mahi_sayli said:

if we can gather enough data showing objects interact with each other is it possible to train AI to learn laws of physics.

As far as I know the answer is yes. But the output from the AI* will not be a neat equation.

Some short, more technical, notes: Neural networks are powerful function approximators, yet their "black-box" nature often renders them opaque and difficult to interpret**. The model will be able to predict how a physical system evolves. The result of the machine learning is weights in a neural network, not compact equations.

some models struggle with discovering conservation laws. The following paper gives an overview of some challenges, including "black box" Lagrangians:

Accurate models of the world are built upon notions of its underlying symmetries. In physics, these symmetries correspond to conservation laws, such as for energy and momentum. Yet even though neural network models see increasing use in the physical sciences, they struggle to learn these symmetries. In this paper, we propose Lagrangian Neural Networks (LNNs), which can parameterize arbitrary Lagrangians using neural networks. In contrast to models that learn Hamiltonians, LNNs do not require canonical coordinates, and thus perform well in situations where canonical momenta are unknown or difficult to compute. Unlike previous approaches, our method does not restrict the functional form of learned energies and will produce energy-conserving models for a variety of tasks. We test our approach on a double pendulum and a relativistic particle, demonstrating energy conservation where a baseline approach incurs dissipation and modeling relativity without canonical coordinates where a Hamiltonian approach fails. Finally, we show how this model can be applied to graphs and continuous systems using a Lagrangian Graph Network, and demonstrate it on the 1D wave equation.
*) In this case "AI" means various types of neural networks related to physics.

Source: https://arxiv.org/abs/2408.14780

On 4/29/2025 at 8:17 PM, Sohan Lalwani said:

Doesn't OpenFoam have a function for thermodynamic and hydrological physics? Or try simscale

AI used in those softwares is for other purposes, not related to what OP is asking about (as @CharonY correctly noted).

*) "AI" in this case are machine learning and various neural network architectures such as Physics-Informed Neural Networks (PINNs).

**) paper that may be of interest.https://ar5iv.labs.arxiv.org/html/2003.04630

Edited May 4May 4 by Ghideon
reference to CharonY added

I've simulated the Second Law of Thermodynamics and the Quantum Zeno Effect on an Excel spreadsheet.

2 hours ago, KJW said:

I've simulated the Second Law of Thermodynamics and the Quantum Zeno Effect on an Excel spreadsheet.

What NASA level computer did you use

8 hours ago, CharonY said:

Aren't those simulation software? That does not seem what OP is asking about. Most simulation software I know (and it may have changed) you define the models, add the desired input and boundary conditions and let it provide solutions (or approximations of it). From how I understand OP the idea is to take actual data, generate a model from it and then run it.

If this is astrophysics I can recommend things like Universe Sandbox, otherwise no.

6 hours ago, Ghideon said:

As far as I know the answer is yes. But the output from the AI* will not be a neat equation.

Some short, more technical, notes: Neural networks are powerful function approximators, yet their "black-box" nature often renders them opaque and difficult to interpret**. The model will be able to predict how a physical system evolves. The result of the machine learning is weights in a neural network, not compact equations.

some models struggle with discovering conservation laws. The following paper gives an overview of some challenges, including "black box" Lagrangians:

Source: https://arxiv.org/abs/2408.14780

AI used in those softwares is for other purposes, not related to what OP is asking about (as @CharonY correctly noted).

*) "AI" in this case are machine learning and various neural network architectures such as Physics-Informed Neural Networks (PINNs).

**) paper that may be of interest.https://ar5iv.labs.arxiv.org/html/2003.04630

Thanks for the reference.

7 hours ago, Sohan Lalwani said:

9 hours ago, KJW said:

I've simulated the Second Law of Thermodynamics and the Quantum Zeno Effect on an Excel spreadsheet.

What NASA level computer did you use

Why do you think such simulations require a lot of computing power?

4 hours ago, KJW said:

Why do you think such simulations require a lot of computing power?

Your generating something that is by nature very complex, would that not require powerful computing?

Just now, Sohan Lalwani said:

Your generating something that is by nature very complex, would that not require powerful computing?

How far back does you memory of 'computing power' go ?

For instance the language Forth was developed for processors no more powerful than the humble 6502 (and if you know what that was you will know its address space and instruction list).

Forth was used to control alrge (at the time) astronomical telescopes for precision survey and study of the heavens.

I used to use a program called Design View, which was a luxury simulation program with full colour 3D visual capability that was originally designed for DOS - I still have the 3 floppies for loading it.
I say luxury because it was luxury for mechanical systems compared to a spreadsheet.

10 minutes ago, studiot said:

How far back does you memory of 'computing power' go ?

For instance the language Forth was developed for processors no more powerful than the humble 6502 (and if you know what that was you will know its address space and instruction list).

Forth was used to control alrge (at the time) astronomical telescopes for precision survey and study of the heavens.

I used to use a program called Design View, which was a luxury simulation program with full colour 3D visual capability that was originally designed for DOS - I still have the 3 floppies for loading it.
I say luxury because it was luxury for mechanical systems compared to a spreadsheet.

Far enough to know that just because someone duct-taped a FORTH interpreter to a 6502 and called it a telescope controller doesn’t mean we should all aspire to simulate modern physics like it’s 1983.

There’s a difference between running a stepper motor and modeling quantum decoherence. You might’ve gotten by with a floppy disk and 64KB of RAM back then, but let’s not pretend that Excel-on-DOS is the gold standard for simulating quantum systems or entropy dynamics.

We’re not debugging a stack of punch cards here, just acknowledging that simulating non-trivial systems with meaningful resolution tends to scale fast in complexity.

Just now, Sohan Lalwani said:

Far enough to know that just because someone duct-taped a FORTH interpreter to a 6502 and called it a telescope controller doesn’t mean we should all aspire to simulate modern physics like it’s 1983.

There’s a difference between running a stepper motor and modeling quantum decoherence. You might’ve gotten by with a floppy disk and 64KB of RAM back then, but let’s not pretend that Excel-on-DOS is the gold standard for simulating quantum systems or entropy dynamics.

We’re not debugging a stack of punch cards here, just acknowledging that simulating non-trivial systems with meaningful resolution tends to scale fast in complexity.

Another rude response and completely missing the point I (and KJW) was making.

1 minute ago, studiot said:

Another rude response and completely missing the point I (and KJW) was making.

Not blatantly rude, its as snarky as "How far back does you memory of 'computing power' go ?"

Just now, Sohan Lalwani said:

Not blatantly rude, its as snarky as "How far back does you memory of 'computing power' go ?"

You are the one that introduced 'snarky' incorrect comments about computing power.

I neither downvoted it nor reported you as author.

Instead I challenged the incorrectness with direct evidence to the contrary.

Instead of supporting your false claim you chose to be rudely dismissive.

1 minute ago, studiot said:

Another rude response and completely missing the point I (and KJW) was making.

Also, some of it is humorous

Context matters as well, I made a joke also prior. I don't think KJW took offense to this

1 minute ago, studiot said:

You are the one that introduced 'snarky' incorrect comments about computing power.

I neither downvoted it nor reported you as author.

Instead I challenged the incorrectness with direct evidence to the contrary.

Instead of supporting your false claim you chose to be rudely dismissive.

What an excellent challenge "Another rude response and completely missing the point I (and KJW) was making" is...

2 minutes ago, studiot said:

You are the one that introduced 'snarky' incorrect comments about computing power.

I neither downvoted it nor reported you as author.

Instead I challenged the incorrectness with direct evidence to the contrary.

Instead of supporting your false claim you chose to be rudely dismissive.

We seem to be going off topic, lets return to the original topic please.

10 minutes ago, studiot said:

Another rude response and completely missing the point I (and KJW) was making.

This is evidence?

Just now, Sohan Lalwani said:

What an excellent challenge "Another rude response and completely missing the point I (and KJW) was making" is...

Another snarky comment that is still completely missing my point.

Of course that was not my challenge, which was made in my previous post which was

That your original response to KJW was both rude and dismissive and incorrect.

10 hours ago, Sohan Lalwani said:

What NASA level computer did you use

I regard that as both rude and dismissive.

No one needs an NASA level computer to run an Excel spread sheet.

I then supported KJW's contention that analysis could be carried out on relatively simple equipment by supplying further examples, which received another rudely dismissive remark from your goodself.

Just now, Sohan Lalwani said:

someone duct-taped a FORTH interpreter to a 6502 and called it a telescope controller

Can you speak or parse or whatever Forth ?

I don't suggest there are not more but I only know on one other member here who has used it ( and I have forgotten most of what i knew about it) and appreciates its special characteristic in its ability to compact/compress code.

Edited May 5May 5 by studiot

4 hours ago, studiot said:

Another snarky comment that is still completely missing my point.

Of course that was not my challenge, which was made in my previous post which was

That your original response to KJW was both rude and dismissive and incorrect.

I regard that as both rude and dismissive.

No one needs an NASA level computer to run an Excel spread sheet.

I then supported KJW's contention that analysis could be carried out on relatively simple equipment by supplying further examples, which received another rudely dismissive remark from your goodself.

Can you speak or parse or whatever Forth ?

I don't suggest there are not more but I only know on one other member here who has used it ( and I have forgotten most of what i knew about it) and appreciates its special characteristic in its ability to compact/compress code.

Thats YOU, I also just clarified its a JOKE, learn to distinguish the two.

4 hours ago, studiot said:

I regard that as both rude and dismissive.

Did I say "WHAT NASA LEVEL COMPUTER DID YOU USE LIAR?!?!?!?!!?!?!." No, I simply made a joke within context. Read the context.

4 hours ago, studiot said:

Another snarky comment that is still completely missing my point.

Of course that was not my challenge, which was made in my previous post which was

That your original response to KJW was both rude and dismissive and incorrect.

I regard that as both rude and dismissive.

No one needs an NASA level computer to run an Excel spread sheet.

I then supported KJW's contention that analysis could be carried out on relatively simple equipment by supplying further examples, which received another rudely dismissive remark from your goodself.

Can you speak or parse or whatever Forth ?

I don't suggest there are not more but I only know on one other member here who has used it ( and I have forgotten most of what i knew about it) and appreciates its special characteristic in its ability to compact/compress code.

I made a joke, nothing here that I said is directly offensive, some comments are snarky like yours, others are humorous. If you have a problem with it simply notify me in a logical manner preferably and not go on a tangent.

7 hours ago, Sohan Lalwani said:

Far enough to know that just because someone duct-taped a FORTH interpreter to a 6502 and called it a telescope controller doesn’t mean we should all aspire to simulate modern physics like it’s 1983.

There’s a difference between running a stepper motor and modeling quantum decoherence. You might’ve gotten by with a floppy disk and 64KB of RAM back then, but let’s not pretend that Excel-on-DOS is the gold standard for simulating quantum systems or entropy dynamics.

We’re not debugging a stack of punch cards here, just acknowledging that simulating non-trivial systems with meaningful resolution tends to scale fast in complexity.

Am I saying "STUDIOT YOUR TOTALLY WRONG AND ARE BEING DUMB?" No, its a joke man.

7 hours ago, studiot said:

How far back does you memory of 'computing power' go?

Mods, this isnt snarky?

Anyway, I did request a ban

I just wanted to notify @KJW I made a joke, not an insult. :)

Note: In my opinion the title of the tread "Simulating Physics with AI" does not relay match the AI related questions in the opening post and this possibly causes some confusion. I try to address the content of the post rather than what the title could imply.

I found a video closely related to the question in the opening post where the presenter is using ideas from computer vision problems (convolutional  neural network) to work on temperature fields; closely related to the specific section:

On 3/31/2025 at 5:51 PM, Mahi_sayli said:

For example if we record behaviour of an siolated system involving few objects and use that data to build model such that it predicts motion for next few frames.

The video, titled "Can AI Uncover the Laws of Physics by Observing Apples Falling from Trees?" also briefly discusses the general question about AI capabilities to (re)-discover Lawes of physics (stated in the opening post). Here is the summary of the video, from YouTube (emphasis mine):

Moving beyond mere pattern recognition, machines are now capable of extracting new insights from hidden trends and patterns, generating lifelike images and coherent text, and making complex decisions in intricate environments. As these advancements progress at an astonishing pace, it begs the question of whether artificial intelligence (AI) will eventually attain the level of intelligence required to delve into highly intellectual pursuits, such as comprehending the fundamental laws of physics in nature. In this presentation, Dr. Baek will highlight some of the recent frontiers in physics-aware deep learning and demonstrate their application in solving complex mechanical engineering problems, ranging from designing composite materials to predicting quantum spin dynamics.

The video is a recorded presentation from University of Virginia, link https://www.youtube.com/watch?v=qhSkX7DjvSM

10 hours ago, Ghideon said:

Note: In my opinion the title of the tread "Simulating Physics with AI" does not relay match the AI related questions in the opening post and this possibly causes some confusion. I try to address the content of the post rather than what the title could imply.

I found a video closely related to the question in the opening post where the presenter is using ideas from computer vision problems (convolutional  neural network) to work on temperature fields; closely related to the specific section:

The video, titled "Can AI Uncover the Laws of Physics by Observing Apples Falling from Trees?" also briefly discusses the general question about AI capabilities to (re)-discover Lawes of physics (stated in the opening post). Here is the summary of the video, from YouTube (emphasis mine):

Moving beyond mere pattern recognition, machines are now capable of extracting new insights from hidden trends and patterns, generating lifelike images and coherent text, and making complex decisions in intricate environments. As these advancements progress at an astonishing pace, it begs the question of whether artificial intelligence (AI) will eventually attain the level of intelligence required to delve into highly intellectual pursuits, such as comprehending the fundamental laws of physics in nature. In this presentation, Dr. Baek will highlight some of the recent frontiers in physics-aware deep learning and demonstrate their application in solving complex mechanical engineering problems, ranging from designing composite materials to predicting quantum spin dynamics.

The video is a recorded presentation from University of Virginia, link https://www.youtube.com/watch?v=qhSkX7DjvSM

You dropped this

Thanks for getting it back on topic

I have an answer for OP: Several software options allow you to simulate physics. SimPHY, Physion, and myPhysicsLab are designed for educational purposes, providing interactive simulations for visualizing physics concepts. For more advanced simulations, COMSOL Multiphysics offers powerful multiphysics capabilities for engineering and scientific research.

On 5/4/2025 at 5:18 PM, Ghideon said:

As far as I know the answer is yes. But the output from the AI* will not be a neat equation.

Some short, more technical, notes: Neural networks are powerful function approximators, yet their "black-box" nature often renders them opaque and difficult to interpret**. The model will be able to predict how a physical system evolves. The result of the machine learning is weights in a neural network, not compact equations.

some models struggle with discovering conservation laws. The following paper gives an overview of some challenges, including "black box" Lagrangians:

Source: https://arxiv.org/abs/2408.14780

AI used in those softwares is for other purposes, not related to what OP is asking about (as @CharonY correctly noted).

*) "AI" in this case are machine learning and various neural network architectures such as Physics-Informed Neural Networks (PINNs).

**) paper that may be of interest.https://ar5iv.labs.arxiv.org/html/2003.04630

Thanks for sharing this. As you and @CharonY pointed, I am not talking about simulations. In typical simulations, physics is already known and we use known physics equations to solve the domain using boundary conditions. With AI, we won't be stipulating any physics equations, we will let the data figure out the relationship between various physical properties. Let's say we record a ball bouncing up and down using high speed cameras, and then let the AI models figure out the motion of the balls. We would record many other objects and add to the database, so AI just doesn't learn the behavior of one object. Eventually AI will come to model that has similar output as Newton's law.

28 minutes ago, Mahi_sayli said:

Thanks for sharing this. As you and @CharonY pointed, I am not talking about simulations. In typical simulations, physics is already known and we use known physics equations to solve the domain using boundary conditions. With AI, we won't be stipulating any physics equations, we will let the data figure out the relationship between various physical properties. Let's say we record a ball bouncing up and down using high speed cameras, and then let the AI models figure out the motion of the balls. We would record many other objects and add to the database, so AI just doesn't learn the behavior of one object. Eventually AI will come to model that has similar output as Newton's law.

What type of physics are you looking to stimulate if you don't mind me asking?

17 minutes ago, Mahi_sayli said:

With AI, we won't be stipulating any physics equations, we will let the data figure out the relationship between various physical properties. Let's say we record a ball bouncing up and down using high speed cameras, and then let the AI models figure out the motion of the balls. We would record many other objects and add to the database, so AI just doesn't learn the behavior of one object. Eventually AI will come to model that has similar output as Newton's law.

What you describe has been around not under the name of AI, but under the umbrella of machine learning. Neural networks as outlined by @Ghideon is an example of such an approach. The issues are also mentioned, including that most are black box approaches, which makes issues of overfitting problematic (i.e. the outcome could be heavily biased due to your training set, but you won't know why and how it affected it). In contrast to, say Newton's law which is a very reductive model (and hence is elegant, but also idealized) you'll get something that has many, many, unwieldy parameters. Trying to prune that down to something like a Newtonian equation will require significant follow-up work. Basically you will try to create a simplified system based on an overly complex ML-model. Whether that is easier than to derive them otherwise depends on the system, I guess (which is way above my knowledge level).

The first thing to note is that as with any computer system it can only output what it is programmed to output.

One of the features of much of human discovery has been the occurrence of the unexpected.

Columbus died thinking he had discovered Asia.

Fleming died knowing he had discoverd penecillin.

Rutherford's famous statement

Rutherford was astonished at the result: It was quite the most incredible event that ever happened to me in my life. It was as incredible as if you fired a 15-inch shell at a piece of tissue paper and it came back and hit you! (You may find other versions of this quote, because Rutherford described his experience on many different occasions.)

https://spark.iop.org/collections/rutherfords-experiment

The quote is from a famous experiment described here and I have referenced it as it describes the sort of process that leads to discovery.

That is to be alert for the unexpected.

So one possible approach is to program your AI to compare with the expected and output anything that is unexpected.

I remember being the only student in the class that got the 'right' wrong answer in a Chemistry experiment when the teachers gave use the wrong solution and everybody else falsified their results to fit expectations.