Recovering Irreducible Three-Body Interactions with Lagrangian Neural Networks

summary

Video file (mp4)

The gist

The gist: Lagrangian Neural Networks can recover an irreducible three-body interaction from motion data alone by parameterizing a scalar Lagrangian and learning only the potential, which has been

In short

Lagrangian Neural Networks (LNNs) learn a physical system by parameterizing its potential and deriving dynamics from motion data. This study successfully used LNNs to recover the irreducible three-body Axilrod-Teller interaction from particle trajectories. The learned potential accurately reproduced the known radial scaling, angular dependence, and sign reversal of this complex three-body force.

Key concepts

Lagrangian Neural Networks (LNNs)
LNNs learn mechanical systems by defining a scalar Lagrangian function that describes the system's energy and motion. The dynamics are then derived by applying the Euler-Lagrange equations to this function. This method biases the learning toward physically realistic motions that respect the underlying structure of mechanics.
Axilrod-Teller Triple-Dipole Interaction
This is a specific, irreducible three-body force term in many-body physics that cannot be broken down into simple pairwise interactions. It depends on the collective geometry of three particles, specifically involving an angular factor that changes sign based on their arrangement. Recovering this interaction proves the model can capture complex many-body effects.
Symbolic Regression
This is a machine learning technique used to automatically discover mathematical formulas from data without assuming a specific functional form beforehand. In this paper, it was used on the learned potential to reconstruct the exact Axilrod-Teller expression, including its radial law and coupling constant, confirming the model's success.

Terminology used across episodes

This episode discusses

The paper

Recovering Irreducible Three-Body Interactions with Lagrangian Neural Networks · Read on arXiv

Sena Kalabalık, Şukrཡ C⌇ağlar, Abdullah Umut Hamzao⦚gullari, Arkadaıs Ozak཯n

Department of Physics, Boıgazi University · Department of Computer Engineering, Boıgazi University

Lagrangian Neural Networks (LNNs) learn mechanical systems from trajectory data by parameterizing a scalar Lagrangian and deriving the dynamics through the Euler-Lagrange equations, biasing the model toward motion that respects the variational structure of mechanics. The interactions studied so far have been pairwise-additive, leaving open whether LNNs can recover an irreducible many-body interaction from motion alone. We address this using the Axilrod-Teller triple-dipole interaction, the leading three-body dispersion term, whose dependence on the collective geometry of three particles cannot be represented by any sum of pairwise terms. We prescribe the kinetic energy exactly and learn only the potential, as a function of interparticle distances and angles. Combined with a signal-aware sampling strategy that concentrates training data where the three-body signal is strongest while preserving broad coverage of configuration space, this recovers the interaction from accelerations alone. The learned potential reproduces the radial scaling, angular dependence, and characteristic sign reversal of the Axilrod-Teller interaction, and generates stable trajectories. Embedding it within a pairwise Lennard-Jones potential lets us examine its observability: the three-body term is recovered once its dynamical signature exceeds the residual modeling error, which it does cleanly at amplified coupling but not at the coupling of a real noble gas, bracketing the transition between the two. Finally, symbolic regression on the trained potential, with no assumed functional form, returns the Axilrod-Teller expression with angular numerator, radial law, and coupling constant intact. An LNN with domain-specific structure can thus recover an irreducible three-body interaction from trajectories as a closed-form law, pointing toward learning interaction laws whose structure is not known a priori.

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "Recovering Irreducible Three-Body Interactions with Lagrangian Neural Networks".

Jane: The gist: Lagrangian Neural Networks can recover an irreducible three-body interaction from motion data alone by parameterizing a scalar Lagrangian and learning only the potential,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: We’re talking about this paper, "Recovering Irreducible Three-Body Interactions with Lagrangian Neural Networks." The core claim is that these neural networks can learn an irreducible three-body interaction directly from motion data alone.

Jane: That sounds like a big deal because most models we use are good at learning pairwise interactions, but they struggle when the physics requires three particles to interact in a specific way.

Tom: Exactly. The authors address whether LNNs can recover an irreducible many-body interaction from motion data, specifically targeting the Axilrod-Teller triple-dipole interaction.

Jane: This is interesting because the paper suggests that this three-body term’s dependence on the collective geometry of three particles cannot be represented by any sum of pairwise terms.

Tom: Right. So they are trying to show that motion data alone can reveal physics that standard pairwise models miss, and they do this by learning only the potential.

Lu: The paper sets up a system where they prescribe the kinetic energy exactly and learn only the potential as a function of interparticle distances and angles. This setup is key because it forces the model to learn the interaction itself from those dynamical constraints.

Meng: Why is that setup important for recovering something irreducible? Does fixing the kinetic energy help isolate that three-body effect?

Tom: It helps by isolating what the potential must be based on how particles move under those specific dynamical rules. Combined with their signal-aware sampling strategy, they can recover the interaction from accelerations alone.

Jane: So, if you only have trajectories and no explicit knowledge of a three-body term, this framework says that the data is enough to find it if you use a clever way to sample the training data.

Lu: The paper states that their learned potential reproduces the radial scaling, angular dependence, and characteristic sign reversal of the Axilrod-Teller interaction. This shows they aren't just finding some random fit; they are capturing the actual physical behavior of this specific three-body force.

Meng: So, what’s a practical implication for us? If we can do this, what does that change for how we model complex systems in AI?

Tom: It means we have a route toward learning interaction laws whose structure is not known a priori. It points toward using learned dynamical models as tools for scientific discovery, where numerical agreement with trajectories establishes that the model works.

Jane: So, it's about moving from just modeling known interactions to discovering new ones based on observed dynamics and then extracting an analytic formula from the AI.

Lu: The paper does this by pairing physics-informed neural learning with symbolic methods to move from dynamical data toward interpretable analytic forms. That’s a powerful combination, I think.

Meng: And we need to be careful about the limits of that interpretation, though. Symbolic regression is only as good as the model it's applied to; if the underlying physics isn't captured perfectly, you might just get a messy expression back instead of a clean formula.

Tom: Exactly. The paper shows they can do this cleanly at amplified coupling, but not at the coupling of a real noble gas, which brackets that transition point between finding the interaction and seeing the noise.

Conclusion: Tom: So to wrap up, we looked at this work on "Recovering Irreducible Three-Body Interactions with Lagrangian Neural Networks" by Sena Kalabalık, Abdullah Umut Hamzao˘gulları, and Arkada¸s Ozakın ¨. They showed that LNNs can recover a specific three-body potential from motion data.

Jane: In simple terms, it means we can use trajectory data to reconstruct the actual mathematical law describing a complex interaction that isn't just about pairs of particles.

Tom: It’s not about building a better force field; it’s about discovering what the force is based on how things move, which is a new way to approach scientific discovery in AI.

Jane: It suggests that we can learn the structure of physics by learning the dynamics, and then use symbolic regression to extract that structure back into an explicit mathematical expression.

Tom: It opens up a path for studying other complex systems where higher-order interactions are difficult to formulate or their structure is not known beforehand.

Jane: That’s a very cool direction for how we think about AI—not just solving problems, but discovering the underlying laws of nature from behavior, and then translating that into math.

Tom: We've got a framework here that allows us to study these higher-order interactions by learning them directly from data in this way.

More episodes

← Home