MUTACO: Simulator-Efficient Response Matching in Stochastic Nonlinear Systems via Compatibility-Aware Sequential Design

arXiv:2610.04619 · cond-mat.stat-mech, physics.comp-ph, physics.data-an · Submitted 2026-10-03 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.

Kai: We're looking at the paper: "MUTACO: Simulator-Efficient Response Matching in Stochastic Nonlinear Systems via Compatibility-Aware Sequential Design".

Mira: Here's the gist of the paper. MUTACO introduces an uncertainty-aware sequential design method for response matching that requires only noisy forward evaluations and learns the multicomponent response together with its predictive uncertainty The Core Problem and System Model Inverse problems in nonlinear dynamical systems play a central role in contemporary physics and scientific computing.

Kai: In many settings, the forward dynamics can be modeled explicitly, while recovering the underlying parameters from macroscopic observations remains a challenging inverse problem.

Mira: This difficulty is amplified in stochastic and disordered systems, where thermal fluctuations, quenched heterogeneity, or unresolved microscopic degrees of freedom make observable responses variable even at fixed parameter values.

Kai: The work focuses on the inverse characterization of disordered networks of stochastic nonlinear phase oscillators, where the task is to infer statistics of the local potential and interactions from macroscopic rotation and synchronization measurements under multiple driving conditions.

Title and authors: Kai: So we're looking at this paper called "MUTACO: Simulator-Efficient Response Matching in Stochastic Nonlinear Systems via Compatibility-Aware Sequential Design". Mira, what's the main idea behind that title?

Mira: It sounds like it’s about using a sequential design method to match responses, but instead of needing perfect models, it only needs noisy forward evaluations. The core concept is this uncertainty-aware sequential design that learns the whole response and its uncertainty at once.

Lev: I'm thinking about what that means practically for running these kinds of inverse problems on real hardware. If we can get away with just noisy simulations, that opens up a lot of possibilities because we don't need to spend massive amounts of time building perfect gradients or adjoints first.

Kai: Exactly. It focuses on inferring the statistical properties—like the local potential and interactions—from macroscopic rotation and synchronization curves under multiple driving conditions. It’s not about finding every single microscopic parameter individually, but rather identifying the ensemble statistics that govern how those parameters create a response that matches what we measure.

Mira: That distinction is key. They are trying to identify configurations of ensemble parameters whose macroscopic rotation and synchronization responses reproduce prescribed target behavior across multiple driving conditions. It’s mapping the statistical landscape instead of just fitting a single point in parameter space, which makes sense given the disorder in these systems where thermal fluctuations and quenched heterogeneity make observable responses variable even at fixed parameter values.

Lev: If the goal is reproducing macroscopic behavior from ensemble statistics, then MUTACO seems designed to handle that complexity without requiring an analytically differentiable forward model or any adjoint construction. That’s a big win for simulation-based inference on hard problems because those tools are often unavailable in physics simulations.

Kai: Right. So we're talking about using the structure of the multicomponent response and its predictive uncertainty directly during the sequential acquisition process. It builds a correlated multitask Gaussian process surrogate for the full vector of rotation and synchronization responses, which lets them learn correlations among observables and driving conditions together with their posterior uncertainty.

Mira: That sounds like they are building a smart map that knows where it's unsure, not just where it's certain about the response itself. The way they structure the acquisition strategy—alternating between target-directed refinement when a compatible region is found and uncertainty-driven exploration otherwise—is what makes it sequential design rather than just one big search.

Lev: From an engineering standpoint, that alternating strategy is smart because it allocates your limited simulation budget intelligently. If you find a promising spot, you refine it carefully using a logarithmic barrier to stay near the compatibility boundary, but if nothing good is found, you direct the next simulation toward regions of largest posterior response uncertainty.

The paper's summary: Kai: So what does the paper actually claim about MUTACO? It seems to be proposing this method for inverse characterization in noisy, heterogeneous many-body systems by requiring only noisy forward evaluations. Mira, can you put that in terms of what the system is trying to achieve?

Mira: The system is trying to infer ensemble statistics of the local potential and interactions from macroscopic rotation and synchronization curves measured under multiple driving conditions. It’s about identifying those ensemble parameter configurations whose responses reproduce the prescribed target behavior across all those different driving conditions.

Lev: But the paper also talks about a crucial distinction they make: that proximity in the space of parameter statistics does not necessarily imply proximity in the measured collective response. This is something I think is really important when we look at how these systems behave under noise and disorder, where that mismatch between what generates the response and what we actually measure can be quite large.

Kai: Right, so MUTACO achieves two things: it gets a good estimate of the ensemble parameters, and it also gets a low mean parameter-recovery error even though parameter recovery isn't their primary optimization objective. They show that this method achieves the lowest mean held-out response error among tested methods under strict budgets of only forty and one hundred forward simulations <ref:2610.04619#pg3,achieves the lowest mean held-out response error among>.

Mira: And they show that MUTACO also achieves lower mean parameter-recovery error than the compared methods, which is a significant result because it shows that response-focused inversion and parameter identification should be distinguished in stochastic oscillator networks. The joint response–parameter analysis shows that small parameter error and small response error are not interchangeable.

Lev: That distinction is what I’m interested in—it clarifies the role of identifiability here. The target vector only contains coarse rotation and synchronization observables measured at a finite set of bias conditions, and parameter vectors separated in the four-dimensional inference space can still be close in response space because they produce similar collective behavior.

Kai: And that's where the local-sensitivity analysis comes in, showing how these observables constrain different parameter directions very unevenly. The weak CVK direction provides a concrete explanation for that part of the response–parameter mismatch they observed above.

The paper's improvements: Mira: Now let's talk about the specific improvements they suggest over existing methods, because that’s where the method really earns its keep. The main improvement is using this compatibility-aware sequential design framework which uses a correlated multitask Gaussian process surrogate to define an uncertainty dependent compatibility criterion relative to the prescribed target.

Lev: That means they are organizing their sequential search based on learned correlations and posterior uncertainty, rather than just blindly searching the space or relying on scalarization techniques. It lets them alternate between target-directed refinement when a compatible region is found and uncertainty-driven exploration otherwise, which is much more efficient for limited simulation budgets.

Kai: And then they have this continuous terminal optimization policy after the simulator budget is exhausted, which minimizes the mean absolute mismatch between the physical posterior-mean response and the prescribed target over the entire admissible parameter domain. This yields a measurable improvement of twenty-eight point five percent in mean held-out physical response MAE compared to other restricted evaluation policies.

Mira: That continuous terminal search is interesting because it means you don't just pick points visited during sequential acquisition; you optimize continuously over the whole parameter space using the learned posterior information. It shows that this policy gives lower held-out response errors than simply selecting among points visited during sequential acquisition.

Lev: From an overhead perspective, that continuous optimization doesn't require any additional simulator evaluations to run it, which is good because we’re already tight on budget. It keeps the search going where it has the most potential, and that seems like a way to squeeze more information out of every simulation call.

Kai: So overall, MUTACO is designed so you don't need an analytically differentiable forward model or any adjoint construction. You just need a queryable stochastic simulator returning a multicomponent response; the method operates directly from finite noisy evaluations of that response, making it applicable to simulator-mediated inverse problems where reproducing observable collective behavior is more relevant than uniquely identifying the microscopic parameters that generated it.

Conclusion: Kai: So we've gone through MUTACO—the simulator-efficient response matching in stochastic nonlinear systems via compatibility-aware sequential design. To wrap up, what does this all mean for us working with these complex oscillator networks?

Mira: It means we can tackle inverse problems where the generating microscopic parameters are hard to identify, and focus instead on reproducing the macroscopic behavior they produce across multiple driving conditions using just a limited set of noisy simulations. The key is distinguishing between response reconstruction and parameter identification.

Lev: And for us, it suggests that we don't need to wait for perfect analytical tools like adjoints or differentiable models to solve these problems; we can build inference methods directly on finite noisy evaluations of the multicomponent stochastic response.

Kai: The paper shows that MUTACO achieves the lowest mean held-out response error among tested methods under strict budgets, and it also achieves lower mean parameter-recovery error than other compared methods, despite not optimizing parameter recovery.

Mira: So we can use this framework to navigate the search space intelligently by using learned correlations and posterior uncertainty to organize that search, alternating between targeted refinement and uncertainty-driven exploration. It’s a way to manage those limited simulation budgets effectively in these difficult inverse problems where the response statistics are what truly matter.

Lev: I just want to say that having a continuous terminal optimization policy that minimizes the mean absolute mismatch between the physical posterior-mean response and the prescribed target over the entire admissible parameter domain is a strong addition because it yields a measurable improvement of twenty-eight point five percent in mean held-out physical response MAE.

Kai: Yeah, so we’re looking at this MUTACO paper as a framework that lets us leverage simulator-mediated inference even when analytical tools are missing, provided we have the necessary structure in the parameter-to-response map to learn from the budget.

Mira: That's right. We need that structure to be learnable from the available simulation budget before we can use MUTACO effectively on these kinds of problems.

Lev: So, it’s a method that requires only a queryable stochastic simulator returning a multicomponent response; it doesn't need an analytically differentiable forward model or governing equations exposed to the optimizer.

Department of Physics, Bilkent University · Department of Electrical and Computer Engineering, Texas A&M University

cond-mat.stat-mech, physics.comp-ph, physics.data-an

Submitted: 2026-10-03

Updated: 2026-10-29

Comments: 29 pages, 8 figures, 5 tables. v2: Clarified model interpretation; updated references and AI-use acknowledgment. Numerical results unchanged

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

Importance score: 90/100

The gist: MUTACO introduces an uncertainty-aware sequential design method for response matching that requires only noisy forward evaluations and learns the multicomponent response together with its predictive

Terminology

Summary

MUTACO introduces an uncertainty-aware sequential design method for response matching that requires only noisy forward evaluations and learns the multicomponent response together with its predictive uncertainty

The Core Problem and System Model

Inverse problems in nonlinear dynamical systems play a central role in contemporary physics and scientific computing. In many settings, the forward dynamics can be modeled explicitly, while recovering the underlying parameters from macroscopic observations remains a challenging inverse problem. This difficulty is amplified in stochastic and disordered systems, where thermal fluctuations, quenched heterogeneity, or unresolved microscopic degrees of freedom make observable responses variable even at fixed parameter values. The work focuses on the inverse characterization of disordered networks of stochastic nonlinear phase oscillators, where the task is to infer statistics of the local potential and interactions from macroscopic rotation and synchronization measurements under multiple driving conditions.

The MUTACO Framework

MUTACO is a sequential-design method for response-focused inversion that requires only stochastic input–output evaluations of the forward simulator. It is distinguished by using the structure of the multicomponent response and its predictive uncertainty directly during sequential acquisition. The method builds a correlated multitask Gaussian process surrogate for the full vector of rotation and synchronization responses, allowing correlations among observables and driving conditions to be learned together with their posterior uncertainty.

Sequential Acquisition Strategy

The MUTACO search alternates between target-directed refinement when a compatible region has been identified and uncertainty-driven exploration otherwise. When a candidate is compatible, it is refined using a logarithmic barrier that penalizes approach to the compatibility boundary. If no finite strict-interior candidate is found, MUTACO switches to uncertainty-driven exploration by directing the next simulation toward regions of largest posterior response uncertainty.

Performance and Comparative Analysis

MUTACO achieves the lowest mean held-out response error among the tested methods under strict budgets of only 40 and 100 forward simulations. Notably, although parameter recovery is not its optimization objective, MUTACO also achieves lower mean parameter-recovery error than the compared methods. The comparison shows that proximity in the space of parameter statistics need not imply proximity in the measured collective response.

Distinction Between Objectives

The results show that response-focused inversion and parameter identification should be distinguished in stochastic oscillator networks. The joint response–parameter analysis shows that small parameter error and small response error are not interchangeable. MUTACO retains the multicomponent response and uses learned correlations and posterior uncertainty to organize sequential search, whereas scalar qLogNEI compresses each simulator evaluation to a single target-error value before fitting its surrogate. The distinction is particularly relevant under severe simulator budgets, consistent with the larger performance gap observed at B = 40 than at B = 100.

Local Response Sloppiness

The local-sensitivity analysis shows that these observables constrain different parameter directions very unevenly. This response–parameter mismatch clarifies the role of identifiability in the present problem: the target vector contains only coarse rotation and synchronization observables measured at a finite set of bias conditions, and parameter vectors separated in the four-dimensional inference space can therefore remain close in response space. The weak CVK direction provides a concrete explanation for part of the response–parameter mismatch observed above.

Conclusion

MUTACO requires only a queryable stochastic simulator returning a multicomponent response; it does not require an analytically differentiable forward model, accessible response Jacobians, adjoint construction, or governing equations exposed to the optimizer. The method operates directly from finite noisy evaluations of a multicomponent stochastic response, making it applicable to simulator-mediated inverse problems in which reproducing observable collective behavior is more relevant than uniquely identifying the microscopic parameters that generated it. The present formulation already accommodates stochastic variability in the forward response; changes in network size or temperature, as well as additional measurement noise, primarily modify the response statistics rather than the structure of the inverse problem. The relevant requirement is that the parameter-to-response map contain structure that can be learned from the available simulation budget.

Acknowledgments

A.C.K. is deeply grateful to Emma Mitchell for first drawing his attention to the broader importance of inverse problems in physics and engineering, and to Dan MacKinlay and Laurence Davies for introducing him to simulation-based inference for inverse problems and for their generous time and many discussions. This publication has been produced with support from the TÜBİTAK 2232/A International Fellowship for Outstanding Researchers Program (Project No. 124C530); A.C.K. and U.C.T acknowledge support under this project. The work of İ.O.K. and C.T was supported in part by TÜBİTAK Grant No. 124E065.

Data and Code Availability

The data underlying the reported results and the associated simulation, inference, and analysis code are available from the corresponding authors upon reasonable request. Public code and data repository links will be provided in the final version of the manuscript. The numerical calculations reported in this paper were performed at TÜBİTAK ULAKBİM, High Performance and Grid Computing Center (TRUBA resources).

Appendix A: Gaussian-Process Theory Background

We use both operator inner-product notation and, where convenient, a compact index notation. A Gaussian process assigns a probability measure to a space of functions. In this representation, k is the two-point correlation function of the field, and k−1 is the corresponding quadratic, or precision, operator.

Appendix B: MUTACO Gaussian-Process and Acquisition Implementation

The multitask GP has a task-dependent constant mean, E[f p(u)] = mp. The posterior covariance is denoted as Σe n(u). The latent posterior mean is given by µen(u) = m(u) + KuXC−1 (ye − M).

Appendix D: Terminal-Domain Ablation

The primary MUTACO procedure separates sequential data acquisition from the terminal recommendation. The continuous terminal optimization requires no additional simulator evaluations. The ablation shows that the continuous terminal search provides a measurable improvement beyond simply selecting among points visited during sequential acquisition.

Appendix C: Baseline Methods and Numerical Implementation

The scalar qLogNEI baseline reduces each multicomponent simulator response to a scalar target-matching error before fitting its surrogate. The Sobol search is a nonadaptive baseline that does not depend on the observed simulator responses. The comparison tests MUTACO against an objective-aligned scalarization of the same response-matching problem under the same simulator budget.

Table IV Summary

The terminal recommendation policies used in the 120-target benchmark comparison are summarized in Table IV. The MUTACO policy uses the MAE of posterior-mean physical response and continuous optimization over Θ. The Sobol baseline selects the observed physical response MAE from evaluated points only.

Table V Summary

The numerical design and simulator-call accounting for the 120-target benchmark are summarized in Table V.

Improvements for AI systems

  1. Response-focused inversion system that requires only noisy forward evaluations to infer statistical properties of local potential and interactions from macroscopic rotation and synchronization curves under multiple driving conditions. This system can achieve the lowest held-out response error among the tested adaptive-search, space-filling, and simulation-based inference methods while yielding lower parameter-recovery error than the compared methods.

  2. Compatibility aware sequential design that uses a correlated multitask Gaussian process surrogate to define an uncertainty dependent compatibility criterion relative to the prescribed target. This allows for a search strategy that alternates between targetdirected refinement when a compatible region is found and uncertainty-driven exploration otherwise, guiding the allocation of limited simulation budget.

  3. Continuous terminal optimization policy that minimizes the mean absolute mismatch between the physical posterior-mean response and the prescribed target over the entire admissible parameter domain after a simulator budget is exhausted. This policy yields lower held-out response errors than restricted evaluation policies, as shown by an improvement of 28.5% in mean held-out physical response MAE.

  4. A method capable of distinguishing between response reconstruction and parameter identification by showing that small parameter error and small response error are not interchangeable. The system can identify that distinct ensemble-parameter configurations can generate closely similar collective behavior over the measured bias conditions.

Sources

Related papers