On the Theoretical Limits of Learning Unconstrained Interatomic Potentials

arXiv:2610.15078 · cs.LG, physics.chem-ph · Submitted 2026-10-10 · Read on arXiv

Listen

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: "On the Theoretical Limits of Learning Unconstrained Interatomic Potentials".

Jane: The gist: Machine-learned interatomic potentials are shown to suffer from fundamental limitations where minimizing energy-and-force matching alone does not guarantee accurate sampling,

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

Title and authors: Tom: So we're looking at this paper called "On the Theoretical Limits of Learning Unconstrained Interatomic Potentials".

Jane: It sounds a bit heavy, but basically, it's asking if just matching the energy and the force is enough to get a good simulation.

Lu: Exactly. The authors are showing that for unconstrained learning classes, this matching alone isn't enough for accurate sampling when you go into high dimensions.

Tom: It seems like they set up this fundamental problem right there, page one of THIS PAPER, where they state that prediction errors that decay exponentially with dimension can coexist with nearly maximal total variation distance between the learned and target distributions <ref:2610.15078#pg1,prediction errors that decay exponentially with dimension can coexist with nearly maximal>.

Meng: Wait, so even if our AI model is super accurate on the numbers—the energy and forces are almost perfect—the actual way it samples atoms in time can still be totally wrong.

Jane: That's the core idea they're pointing out. They identify two big failure modes: kinetic trapping and equilibrium mismatch.

Tom: Kinetic trapping is when relaxation to an accurate learned equilibrium takes exponentially long times, even if you start your simulation from a perfect training configuration, which is shown in theorems four point one and four point two of THIS PAPER <ref:2610.15078#pg1,relaxation to an accurate learned equilibrium>.

Lu: And then there's the second one, which I think is really interesting: equilibrium mismatch. This happens when the stationary distribution itself that the AI learns is just far away from the true target distribution, and they show this can persist even in local, symmetry-preserving models as system size increases.

Meng: From a deployment standpoint, that's worrying. If we can't guarantee where the simulation ends up settling, then we don't have a reliable way to trust our results at scale.

Tom: Right, so they look at how different structural priors—like locality or symmetry—affect these failures in theorems like Appendix H of THIS PAPER.

Jane: They suggest that while things like locality might hide some errors initially, the analysis shows those hidden errors can still grow with system size, which is a bit of a warning sign for us.

Lu: It’s important to note that even under these structural constraints, they show that per-atom bounds don't necessarily grow with N if you use local invariant hypothesis classes.

Tom: They also establish some sufficient conditions for when sampling *is* accurate, Theorem four point six in THIS PAPER, which gives us global energy-control conditions and additional requirements for rapid relaxation <ref:2610.15078#pg1>.

Title and authors: Jane: That’s a relief because it tells us the specific mathematical hurdles we need to clear to make these models reliable.

Lu: The paper also provides concrete examples of these failures through simulations, like the Einstein Crystal Model, where they show kinetic trapping leading to a surrogate line at twelve point five THz compared to the true crystal prediction.

Meng: That kind of specific discrepancy is what we need to see in production testing if we want confidence in our potentials.

Tom: And for something more general, like the FCC Argon Crystal, they show that even with small held-out force errors, as you increase the degrees of freedom—the dimension sweep—the sampling discrepancy increases by several orders of magnitude.

Jane: So what this means for us is that we can't just focus on minimizing energy and force errors in isolation if we want guaranteed accuracy across different system sizes.

Lu: The paper concludes that smooth, conservative surrogates with dimension-independent Hessian bounds can achieve exponentially small population energy and force errors while still inducing dynamics whose distribution remains far from the target.

Meng: That’s a strong statement because it means even when the model is locally good, the global sampling can be fundamentally flawed without more constraints.

Jane: It suggests that we need more than just matching energy and force to ensure reliable molecular dynamics sampling in these unconstrained MLIPs.

Tom: So, the main message of "On the Theoretical Limits of Learning Unconstrained Interatomic Potentials" is that architectural constraints are necessary for reliable sampling when you're not explicitly controlling all the errors.

Lu: That's right, and it points us toward designing better models with explicit physical priors rather than just relying on unconstrained optimization objectives.

Meng: For production, this means we have to look beyond just fitting the training data perfectly; we need to think about how the learned dynamics will actually behave when things get complex.

Tom: Before we wrap up, let's look at how they propose fixing these issues. The paper suggests using stiffness thresholds to overcome exponential trapping in unconstrained MLIPs by picking a stiffness threshold alpha greater than gamma squared over two, like alpha1 and alpha2.

Jane: That’s a specific engineering knob they suggest tuning to keep the simulation running smoothly in those tricky high-dimensional spaces.

Lu: And then there's controlling equilibrium mismatch by enforcing a lower bound on predicted energy using condition (b) of Theorem four point six, where there is a set S with the target distribution having at least one minus eta and the predicted energy V hat is less than or equal to V plus delta prime on that set <ref:2610.15078#pg1>.

Title and authors: Meng: That means we're looking for regions where the AI is confident about its energy prediction relative to what it should be.

Tom: And finally, structure-aware sampling using local invariant hypothesis classes, like Hloc(rc, L), which provide per-atom bounds that don't grow with system size.

Jane: So, in short, they are telling us that we need to add these specific types of architectural rules to the learning process if we want our MLIPs to be trustworthy across different sizes.

Lu: It seems like the theoretical foundation here is solid, showing exactly where the limitations lie when you leave things unconstrained.

Meng: I just wonder how much computational overhead those structural constraints add when we try to deploy them in a real-time environment.

Tom: That's a fair point, Meng. The paper doesn't detail the exact cost increase for those constraints, but it sets the theoretical bar for what kind of control we need over our models.

Jane: It’s about moving from just fitting data to ensuring physical reality in the simulation dynamics.

Lu: So, looking at "On the Theoretical Limits of Learning Unconstrained Interatomic Potentials", the authors really lay out a rigorous proof that matching energy and force isn't enough on its own for high-dimensional systems.

Meng: It’s a warning about relying too much on data fit without considering the underlying dynamics.

Tom: That’s right, and they give us clear directions for what those necessary constraints look like—global energy control or local invariant classes.

Jane: We're definitely going to be looking at these structural priors more closely as we build the next generation of materials simulation tools.

Lu: It’s a necessary step toward making MLIPs robust enough for complex, real-world applications where sampling accuracy is paramount.

Meng: Yeah, it makes sense that we need that explicit control over the stationary distribution.

Tom: So to wrap up this discussion on "On the Theoretical Limits of Learning Unconstrained Interatomic Potentials", the core message is that unconstrained MLIPs can have exponentially small prediction errors while still having dynamics far from the target if you don't impose structure.

Jane: It’s a reminder that for high-dimensional problems, we need to be intentional about how we constrain our learning process.

Lu: That’s the essence of it, and it gives us a clearer path for designing next-generation potentials that are both accurate and reliable during simulation.

Meng: We'll keep watching how they apply these structural ideas in their future work.

The paper's summary: Tom: So, we just covered how energy and force matching isn't enough for accurate sampling in high dimensions for these unconstrained interatomic potentials. Now we’re going to look at what the authors actually found when they put all that math together on page one of this paper.

Jane: Okay so I guess the big summary is that these machine-learned potentials have a fundamental problem when you don't tell them how to behave structurally. They show that you can have very small errors in energy and force, but the actual way the simulation samples atoms over time can still be wildly wrong.

Tom: Exactly. It’s like the model is good at predicting where an atom *should* be based on its neighbors right now, but it gets stuck in a bad spot because it can't explore the whole system correctly.

Jane: And they break that down into two specific failures, kinetic trapping and equilibrium mismatch. Kinetic trapping means even if you start the simulation from a perfect training state where everything matches exactly, it can take an exponentially long time to actually relax to the right equilibrium distribution.

Tom: That’s when Theorem four point one and Theorem four point two come in, showing that the total variation distance between what your simulation is doing and what it should be is still huge over those long times.

Jane: Then there's equilibrium mismatch, which is even weirder because it shows that even if the AI eventually settles down to *some* stationary state, that stationary state itself can be completely wrong when the system gets really big.

Tom: They prove this happens in local, symmetry-preserving models where small per-atom errors just compound and eventually lead to a massive distributional discrepancy as you increase the system size N.

Jane: It’s a serious warning because it means we can’t just trust our simulations at scale if we only focus on minimizing the energy and force error terms.

Tom: So, what does this mean for us? It suggests that simply fitting data perfectly isn't enough to guarantee that the resulting dynamics will behave like real physics across different system sizes.

Jane: It points toward needing some kind of structural constraint built into the learning process itself if we want reliable sampling, which leads us into how they look at those specific priors next.

The paper's improvements: Tom: So we’ve seen where the problem lies, and now we're looking at what the authors are actually suggesting to fix it in this paper. Basically, they give us three concrete ways to impose structure on our models so they can sample reliably in these high-dimensional spaces.

Jane: Okay so if we just stop using unconstrained learning classes, what are the specific solutions they propose? Are we talking about better training data or some kind of mathematical trick?

Tom: It’s more about adding constraints to the learning process itself. They suggest kinetic trapping mitigation via stiffness thresholds. That means you choose a stiffness threshold alpha greater than gamma squared over two, like alpha one and alpha two, to help the system overcome that exponential trapping in unconstrained potentials.

Jane: So it’s like tuning a dial on the model during training to make sure it doesn't get stuck in those deep energy wells that we talked about earlier. What about equilibrium mismatch? That’s the second fix they propose.

Tom: They suggest controlling equilibrium mismatch using one-sided global energy bounds. This means enforcing a lower bound on predicted energy, specifically by using condition b of Theorem four point six, where there's a set S with the target distribution having at least one minus eta and the predicted energy V hat is less than or equal to V plus delta prime on that set.

Jane: That sounds like they’re telling us to find regions in the model's prediction space where we can be confident about the energy, even if we don't have perfect accuracy everywhere. It’s a practical way to handle that discrepancy.

Tom: And then there’s structure-aware sampling using local invariant hypothesis classes. These are things like Hloc(rc, L), which give you per-atom bounds that don't grow with the system size N, even when the global equilibrium discrepancy approaches one exponentially fast in N.

Jane: So these structural constraints—stiffness thresholds and local invariant classes—are aimed at making sure that no matter how big our system gets, we maintain some level of predictable behavior.

Tom: Right, because if you can control those specific error mechanisms, the paper suggests the model can achieve exponentially small population energy and force errors while still inducing dynamics whose distribution remains close to the target.

Jane: That means we move away from just chasing zero error everywhere and start thinking about how to make our models physically constrained during learning.

Conclusion: Tom: So we’re wrapping up this look at "On the Theoretical Limits of Learning Unconstrained Interatomic Potentials". The main point is that just matching energy and force isn't enough when you don't have structural constraints on these models in high dimensions.

Jane: Exactly. The authors show that without those rules, you get kinetic trapping and equilibrium mismatch, no matter how small your prediction errors are.

Tom: Right, it means we can’t just focus on getting the training data fit perfectly if we want our simulations to stay close to the real physics across different system sizes.

Jane: It suggests that for these machine-learned potentials to be reliable tools, we need a design target beyond just minimizing error metrics on the training set.

Tom: Lu, you’ve been deep in the theory on this stuff. What’s your take on how big of a deal these constraints are for representation learning?

Lu: I think it fundamentally changes how we view the representation space; it shows that we need to embed physical intuition directly into the model's architecture, not just rely on raw data fitting.

Jane: Meng, from a deployment standpoint, what does this mean for the cost and stability of these models in production? Does adding these constraints make them too slow or too expensive to run?

Meng: It adds complexity to the training pipeline initially, sure. But if we can get it right on the first try, those structural priors could save us from having to constantly retrain because of sampling failures at scale.

Tom: Lalam, what does this result mean for how you handle literature and connect these theoretical limits back to what’s actually happening in the research community?

Lalam: Based on the literature I see, this paper is very clear about the gap between energy matching and actual physical sampling; it’s a strong warning that unconstrained potentials have inherent limitations.

Jane: So we’re getting a clearer picture of what to look for when evaluating new MLIPs—we need to check if they have those specific constraints built in.

Tom: We need to ensure the next generation of models isn't just accurate on paper but actually robust during long-term dynamics.

Jane: That’s the big shift, moving from data fitting toward physics-aware learning architectures.

Prajwal Ananth, ♔Shuwen Yue

Cornell University · R. F. Smith School of Chemical and Biomolecular Engineering

cs.LG, physics.chem-ph

Submitted: 2026-10-10

Updated: 2026-10-10

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 80/100

The gist: The gist: Machine-learned interatomic potentials are shown to suffer from fundamental limitations where minimizing energy-and-force matching alone does not guarantee accurate sampling, leading to

Key concepts

Exponentially Ill-Conditioned in Dimension: This means that as the number of dimensions (system size) increases, the relationship between how well a model predicts energies/forces and how accurate its resulting simulation samples becomes degrades exponentially. Even tiny errors in prediction can lead to massive errors in sampling accuracy when the system is large.
Kinetic Trapping: This failure occurs when a simulation gets stuck in a local energy minimum that is not the true physical equilibrium. Even if the model perfectly matches energies and forces at the start, reaching the correct equilibrium state requires an exponentially long time, making accurate sampling practically impossible within reasonable simulation lengths.
Equilibrium Mismatch: This happens when the stable state predicted by a machine-learned potential does not match the true physical stationary distribution of the system. The authors show this can persist even in local models, where small errors per atom accumulate into a large discrepancy between what the model predicts as stable and what is actually physically stable.

Summary

The gist: Machine-learned interatomic potentials are shown to suffer from fundamental limitations where minimizing energy-and-force matching alone does not guarantee accurate sampling, leading to kinetic trapping and equilibrium mismatch in high dimensions

Fundamental Limitations of Unconstrained Matching

The paper demonstrates that for unconstrained hypothesis classes, the relationship between prediction error and sampling accuracy is exponentially ill-conditioned in dimension In particular, arbitrarily small population energy and force errors can coexist with nearly maximal total variation distance between the learned and target distributions. This limitation is established through two primary failure mechanisms.

High-Dimensional Failure Mechanisms

The authors identify two distinct sampling failure modes that can persist even under exponentially small prediction errors.

** Kinetic trapping:**

This occurs when relaxation to an accurate learned equilibrium requires exponentially long times, even when simulations are initialized from training configurations at which energies and forces are matched exactly. This is shown in Theorem 4.1 and Theorem 4.2.

** Equilibrium mismatch:**

This failure mode occurs when the learned stationary distribution itself remains far from the target. The authors show that this can persist in local, symmetry-preserving models where small per-atom errors coexist with a distributional discrepancy that approaches its maximum as system size increases.

Roles and Limitations of Structural Priors

The analysis extends to various structural priors to characterize sampling guarantees. The paper characterizes the sampling guarantees and limitations of conservative, local, permutation-invariant, and E(3)-symmetric hypothesis classes.

** Locality and Symmetry:**

The analysis shows that locality and symmetry can potentially reduce the scale of hidden errors but do not prevent their amplification with system size. In Appendix H, it is shown that locality could potentially raise the error that can be hidden from e−omega(d) to e−omega(ν).

Sufficient Conditions for Accurate Sampling

The work establishes conditions necessary for reliable sampling.

** Global Energy Control:**

Theorem 4.6 identifies global energy-control conditions (Theorem 4.6) that guarantee accurate equilibrium sampling and stronger conditions needed for rapid relaxation. Condition (i) provides a uniform error bound, while condition (ii) requires accuracy only on the support of the target and a global lower bound Vˆ ≥ V − δ.

Experimental Illustrations

Controlled experiments are included to illustrate these theoretical failures.

** Einstein Crystal Model:**

The authors construct an Einstein-crystal model that directly exhibits the kinetic-trapping mechanism of Theorems 4.1 and 4.2. The simulation shows The trapped surrogate’s line is at 12.5 THz compared to the true Einstein crystal's analytic prediction.

** FCC Argon Crystal:**

For an FCC argon crystal, the experiment shows that small held-out force error need not translate into accurate finite-time sampling as the number of degrees of freedom increases. The simulation results show that the dimension sweep, while the sampling discrepancy increases by several orders of magnitude.

The impossibility results for unconstrained MLIPs provide theoretical support for imposing structure beyond energy-and-force matching and identify design targets for such constraints. The results are existential and worst case and therefore do not establish how frequently these failure modes arise in practice.

How it works

The relationship between prediction error and sampling accuracy is exponentially ill-conditioned in dimension. This means that arbitrarily small population energy and force errors can coexist with nearly maximal total variation distance between the learned and target distributions.

Kinetic Trapping under Exponentially Small Error

Theorem 4.1 establishes a lower bound on the total-variation distance between the distribution generated from a displaced initialization and the equilibrium distribution of an Einstein crystal. The construction introduces a smooth basin far from µ, where πtar assigns mass e−omega(d). This leads to the conclusion that the distribution induced by the dynamics remains at total variation distance 1 − e−omega(d) from πtar throughout this time horizon.

Equilibrium Mismatch for Arbitrary Initialization

Theorem 4.3 establishes a stronger failure mode where the learned stationary distribution is itself incorrect. The construction involves subtracting a smooth well of depth ∆, which is constructed such that the well must be deep enough that eβ∆ compensates for the exponentially small prior mass that πˆ would otherwise assign to the ball. This results in TV(ˆπ, πtar) ≥ 1 − e−omega(d).

Sufficient Conditions for Accurate Sampling

Theorem 4.6 provides sufficient conditions for accurate sampling. Condition (i) imposes a uniform error bound, which rules out both failure modes. Condition (ii) weakens this requirement in the direction relevant for practice, requiring accuracy only on the support of the target and replacing global accuracy with a global lower bound Vˆ ≥ V − δ.

Symmetry- and Locality-aware Hypothesis Classes

The paper considers structural priors to see if they eliminate these failures. The analysis shows that locality can potentially raise the error that can be hidden from e−omega(d) to e−omega(ν). The construction in Theorem H.5 shows that with locality, the per-atom bounds do not grow with N.

Experimental Illustrations

The authors include controlled simulations to illustrate the sampling failures predicted by their theory. The experiments are intended as finite-time illustrations of the phenomena captured by the theory rather than numerical realizations of the theorem constructions.

Conclusion

The paper concludes that "Smooth, conservative surrogates with dimensionindependent Hessian bounds can interpolate the training data exactly, achieve exponentially small population energy and force errors, and still induce dynamics whose distribution remains far from the target. The authors suggest that additional constraints beyond energy-and-force matching may be important for reliable MD sampling".

Equilibrium Mismatch for Arbitrary Initialization

Theorem 4.3 establishes a stronger failure mode where the learned stationary distribution is itself incorrect.

Improvements for AI systems

Bold header: Kinetic Trapping Mitigation via Stiffness Thresholds

The improved system can sample from a target distribution with a total variation distance of at least 1 − e −omega(d) for time horizons up to T ≤ e−c⋆d, by choosing a stiffness threshold α ≥ maxγ2/2, α1, α2. This allows the system to overcome exponential trapping in unconstrained MLIPs.

Bold header: Equilibrium Mismatch Control via One-Sided Global Energy Bounds

The AI can ensure that the stationary distribution of its learned dynamics is close to the target by enforcing a lower bound on predicted energy, specifically by using condition (b) of Theorem 4.6 where there is a set S with πtar(S) ≥ 1 − η and Vˆ ≤ V + δ′ on S.

Bold header: Structure-Aware Sampling via Local Invariant Hypothesis Classes

The system can incorporate architectural constraints by using local invariant hypothesis classes, such as Hloc(rc, L), which provide per-atom bounds that do not grow with system size. This allows the model to maintain accuracy in local environments even when the global equilibrium discrepancy approaches one exponentially fast in N.