QPI-DeepONet-MAC: A Scalable and Stable Hybrid Classical-Quantum Architecture for Physics-Informed Deep Operator Networks

arXiv:2610.01824 · quant-ph · Submitted 2026-10-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: Today's paper: "QPI-DeepONet-MAC: A Scalable and Stable Hybrid Classical-Quantum Architecture for Physics-Informed Deep Operator Networks".

Mira: The gist General operator learning for parametric partial differential equations (PDEs) is a fundamental challenge at the intersection of artificial intelligence and physicsbased modeling.

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

Title and authors: Kai: So we're diving into the title and who’s behind this work, "QPI-DeepONet-MAC: A Scalable and Stable Hybrid Classical-Quantum Architecture for Physics-Informed Deep Operator Networks." Mira The name itself tells you a lot about what they are trying to achieve. It’s about scaling up physics models using a hybrid classical and quantum approach within the framework of physics-informed deep operator networks.

Kai: That’s right. They are clearly focused on making these complex physical systems, described by partial differential equations, solvable using this specific architecture that mixes classical and quantum elements together. Lev From a researcher's view, it sounds like they are trying to build a system that can handle the complexity of the physics while also managing the computational challenges inherent in learning models for those physics.

Mira: Exactly. The paper is motivated by the fact that standard PI-DeepONets, which use governing equations to learn solution operators, often run into optimization struggles and lose expressivity when dealing with high-dimensional settings. Kai So, they’re proposing a modification to address those known weaknesses in existing deep learning methods for these physical problems.

Lev: It sounds like the authors are trying to bridge that gap between the theoretical power of classical neural networks and the potential computational flexibility offered by quantum systems.

Mira: They introduce parameterized quantum circuits into this structure via multiplicative and additive couplings, which is what they call MAC, to couple classical representations with quantum expectation values. Kai So, the key mechanism isn't just throwing a QNN in there; it’s this specific coupling method that integrates the quantum information into the operator learning process.

Lev: I wonder how robust those couplings are when you start looking at real physical systems where noise is always present. Mira That’s a fair question, and they tackle that by introducing an informational regularization scheme based on quantum coherence and state fidelity to control the optimization trajectory.

Kai: So it’s not just about adding the quantum part; it's about how you manage the interaction between the classical learning part and that quantum component during optimization. That’s a subtle but important point.

The paper's summary: Kai: Now, let’s look at what they actually say in terms of their overall summary of the "QPI-DeepONet-MAC: A Scalable and Stable Hybrid Classical-Quantum Architecture for Physics-Informed Deep Operator Networks." Mira Basically, they’re confirming that this framework provides a theoretical foundation for incorporating parameterized quantum circuits into physics-informed operator learning.

Mira: They establish that the resulting framework is a principled way to approximate the solution operators of parametric PDEs. They show it can achieve universal approximation under appropriate assumptions for DeepONet architectures.

Kai: So, they’re saying that despite adding the quantum elements, you don't lose the ability to approximate any operator you want as precisely as possible. Lev That’s a strong claim because usually when you add complexity, expressivity starts to suffer in high-dimensional problems.

Mira: They formally define a map called DeepONet v˜♭,WB♭ ·,WT: R mM×one −→ CRM(D), which is the mathematical core of their framework <ref:2610.01824#pg1>. Kai This notation shows exactly how they are mapping the input data to the output space defined by that operator.

Lev: That’s a dense way to describe the mechanism—it’s showing how they rigorously define that approximation relationship between inputs and outputs, which is what allows them to claim arbitrary precision.

Mira: They show that for any input v in V with a discretization v˜♭ in R mM×one and coordinates y in D, there exists a set of weights and biases WB*♭, WT* such that DeepONetj approximates the operator defined by G: V −→ CRM(D) v seven−→ G (v). Kai This is what underpins their claim of arbitrary precision.

Lev: So they’re showing that the math holds up for approximating the operator defined by G: V M −→ CRM(D) v seven−→ G (v) with arbitrary precision, which is a pretty big statement in this context.

Mira: That’s the essence of their work—providing a rigorous mathematical proof that this hybrid approach is both expressively rich and trainable. Kai It really lays the groundwork for how we can integrate quantum computation into scientific modeling in a structured way.

The paper's improvements: Lev: Moving on to the actual improvements they suggest, what are they highlighting that make this architecture better than previous methods? Kai They emphasize mitigating optimization difficulties and barren plateaus specifically. Mira They show that they can avoid those exponentially vanishing gradients by deriving bounds on the gradients with respect to the quantum parameters, setting conditions for when those vanishing gradients won't happen.

Lev: That’s good because it tackles a major practical hurdle for anyone trying to run these models, whether it's on a small test set or a large one. Kai And they also give us this scaling criterion for trainability, which tells you exactly when the system is still stable even as the quantum components get bigger.

Mira: Furthermore, they introduce informational regularization based on quantum coherence and state fidelity to preserve quantum-state structure during training. This is a mechanism to penalize loss of coherence while controlling changes in the quantum state trajectory.

Kai: That means they aren't just hoping the model works; they have a way to actively guide it toward solutions that keep those critical quantum properties intact throughout the learning process. Lev That’s more than just a heuristic; it gives us a specific constraint on how the learning algorithm should behave.

Mira: They also include this total loss function: L T OT which includes terms for coherence, batch correlation, and PDE loss, plus a regularization term LINF. This LINF term is specifically designed to ensure convergence toward a state that preserves an entropy value Sp(ϱQ) > Sthreshold and a coherence Cp(ϱ(p)Q) > zero.

Kai: So they are essentially giving us tools—gradient bounds, scaling criteria, and regularization terms—to make this architecture robust against the inherent instabilities of quantum optimization.

Conclusion: Kai: To wrap up this discussion on "QPI-DeepONet-MAC," we’ve seen how the authors have built a principled hybrid framework that combines classical and quantum elements to handle physics operator learning with stability and expressivity. Mira They successfully show that this approach can remain expressive and trainable even as the quantum components increase in size, rather than just showing performance advantages over purely classical methods.

Lev: From my side, I think the most important part is that they’ve provided analytical guarantees about avoiding those barren plateaus, which gives it a solid foundation for running on real hardware. Kai And they're not just giving us the explicit rules on how to train it; they are showing that this hybrid QPI-DeepONet-MAC architecture can remain expressive and trainable through its structure, regardless of the size of the quantum components.

Mira: So, ultimately, it’s a principled way to integrate parameterized quantum circuits into physics-informed operator learning. Kai This paper gives us a solid theoretical underpinning for incorporating these circuits into scientific modeling in a structured way.

Lev: It sounds like the core contribution is establishing this hybrid framework as theoretically sound for approximating parametric PDE solution operators.

Kai: So, we’re done with this paper and ready to move on to the next one in our queue. Mira This QPI-DeepONet-MAC work provides a strong theoretical base for using quantum computing to tackle these kinds of operator learning problems.

Said Lantigua, José Valencia, Gilson Giraldi, Renato Portugal, Jonas Maziero

Departament of Physics, Center for Natural and Exact Sciences, Federal University of Santa Maria · National Laboratory for Scientific Computing (LNCC) · Department of Mathematics, State University of Londrina (UEL)

quant-ph

Submitted: 2026-10-01

Updated: 2026-10-01

Comments: 47 pages, 3 figures

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

Importance score: 82/100

The gist: The gist General operator learning for parametric partial differential equations (PDEs) is a fundamental challenge at the intersection of artificial intelligence and physicsbased modeling.

Key concepts

QPI-DeepONet-MAC Architecture
This is a hybrid model that merges classical neural networks (DeepONet) with parameterized quantum circuits (QPI). It uses multiplicative and additive couplings to combine classical representations with quantum expectation values, creating a new way to represent the operator.
Multiplicative and Additive Couplings (MAC)
MAC refers to the method used in this architecture where classical neural representations are coupled with quantum expectation values. This coupling introduces a quantum-dependent additive term that enriches the resulting operator representation, improving its ability to model complex physics.
Barren Plateaus Mitigation
This technique addresses a problem where gradients in deep learning models become exponentially small during training, making the model untrainable. The paper shows conditions and scaling criteria to maintain detectable gradients and ensure the architecture remains trainable even as quantum components grow larger.
Quantum-Informational Regularization
This is a loss function scheme used during optimization to control the quantum component's dynamics. It penalizes the loss of quantum coherence while simultaneously constraining changes in the quantum state trajectory, helping to preserve important structural information during training.

Terminology

Summary

The gist General operator learning for parametric partial differential equations (PDEs) is a fundamental challenge at the intersection of artificial intelligence and physicsbased modeling.

QPI-DeepONet-MAC Architecture

The proposed QPI-DeepONet-MAC is a hybrid quantum-classical architecture that integrates parameterized quantum circuits into Physics-Informed Deep Operator Networks through multiplicative and additive couplings (MAC) of classical neural representations and quantum expectation values<ref:2610.01824#pg4>. This framework combines classical operator-learning representations with quantum expectation values, introducing a quantum-dependent additive coupling that enriches the resulting operator representation<ref:2610.01824#pg4>. The model is formally underpinned by a rigorous theoretical framework that ensures its expressivity, trainability, and information efficiency<ref:2610.01824#pg5>.

Theoretical Foundations and Expressivity

The architecture is designed to retain the universal approximation capability of the underlying DeepONet architecture under appropriate assumptions<ref:2610.01824#pg4>. The framework is mathematically defined by a map DeepONet v˜,WB ·,WT: R mM×1 −→ CRM(D), (9). It is established that for any v ∈ VM with discretization v˜ ∈ R mM×1 and coordinates y ∈ D, there exists a set of weights and biases WB∗, WT∗ that allow DeepONetj to approximate the operator defined by G: V −→ CRM(D) v 7−→ G (v), (17). This allows DeepONet to approximate the operator defined by G: V M −→ CRM(D) v 7−→ G (v) (·) with arbitrary precision<ref:2610.01824#pg9>.

Trainability and Barren Plateaus Mitigation

The paper derives bounds on the gradients with respect to the quantum parameters, providing conditions under which exponentially vanishing gradients can be avoided, together with a scaling criterion for trainability<ref:2610.01824#pg4>. The architecture actively mitigates the barren plateaus problem by maintaining detectable gradients and ensuring trainability even as the quantum components scale<ref:2610.01824#pg5>. Specifically, a scaling condition for trainability is derived based on the expression O hN BN T N BN T √ N BNB + 1 N BNB √ N T NT + N T NT √ N BNB ≲ O 1/ϵgrad<ref:2610.01824#pg6>.

Quantum-Informational Regularization

To regulate the quantum component during optimization, an informational regularization scheme is introduced based on the dynamics of quantum coherence and state fidelity<ref:2610.01824#pg4>. This scheme penalizes loss of coherence while controlling changes in the quantum-state trajectory, thereby providing a mechanism for preserving quantum-state structure during training<ref:2610.01824#pg4>. The total loss function is defined as L T OT = λIC L˜IC + λBC L˜BC + λPDE L˜PDE + λSOL SOL + LINF, (74).

Sensor Requirements and Observability

A lower bound on the number of sensors required to achieve a prescribed approximation accuracy for a class of well-posed parametric PDEs is derived<ref:2610.01824#pg4>. This result relates the required sampling density to the regularity and dimensionality of the underlying problem and to the dynamical properties of the corresponding evolution equation<ref:2610.01824#pg6>. The minimum number of sensors m required is given by m ≥ CLCInt/ϵ˜−ϵ d k expωtdk<ref:2610.01824#pg8>.

Conclusion

The analytical results provide a theoretical foundation for incorporating parameterized quantum circuits into physics-informed operator learning and establish QPI-DeepONet-MAC as a principled hybrid framework for the approximation of parametric PDE solution operators<ref:2610.01824#pg3>. The work demonstrates that the architecture can remain expressive and trainable as its quantum components increase in size, rather than establishing performance advantages over classical operator-learning methods<ref:2610.01824#pg6>.

Future Work

Numerical experiments are needed to assess the practical consequences of the theoretical results, including approximation accuracy, optimization stability, sensor requirements, and the behavior of the proposed quantum-informational regularization<ref:2610.01824#pg6>. The implementation of QPI-DeepONet-MAC on quantum hardware will require optimized circuit constructions, parameterizations, and training procedures adapted to the noise and resource constraints of current Noisy Intermediate-Scale Quantum (NISQ) devices<ref:2610.01824#pg6>. The authors plan to develop classical simulations and numerical benchmarks for this purpose<ref:2610.01824#pg6>.

Appendix A Summary

The appendix provides a detailed proof of the theorem 4.1 guaranteeing the universal approximation capabilities of our hybrid QPI-DeepONet-MAC model<ref:2610.01824#pg4>. This proof relies on rewriting the subtraction Gj − Gj in a specific form and regrouping terms to establish four facts that lead to the final inequality (42)<ref:2610.01824#pg4>.

Appendix D Summary

The appendix presents a fundamental result showing the lower bound on the gradient for quantum circuits constructed using local operators whose decay avoids the barren plateau problem<ref:2610.01824#pg4>. This proof concludes by obtaining (53), which is derived from combining the gradient bounds of the branch (E99) and trunk (E100) for the j-th component<ref:2610.01824#pg4>.

Appendix F Summary

The appendix determines the minimum number of sensors required for our hybrid QPI-DeepONet-MAC architecture to approximate the operator G with arbitrary precision<ref:2610.01824#pg4>. This proof concludes by obtaining expression (56), which mathematically represents the minimum number of sensors m required to guarantee system observability<ref:2610.01824#pg4>.

Appendix G Summary

The appendix provides a comprehensive summary of the symbols used throughout this work in Table G1<ref:2610.01824#pg10>. This table summarizes the nomenclature and symbols for the QPI-DeepONet-MAC architecture<ref:2610.01824#pg10>.

Appendix H Summary

The appendix provides a summary of the nomenclature and symbols used throughout this work in Table G1<ref:2610.01824#pg10>. This table summarizes the nomenclature and symbols for the QPI-DeepONet-MAC architecture<ref:2610.01824#pg10>.

Acknowledgments

First, I thank God the Father, infinite goodness, God the Son, redeemer of the world, and God the Holy Spirit, who—together with the Virgin Mary and Saint Joseph—are my protection, my source of inspiration, and my strength on this path of knowledge<ref:2610.01824#pg4>. I also thank my family, for they have made me who I am today<ref:2610.01824#pg4>.

Declarations

The authors declare no competing interests<ref:2610.01824#pg4>. The code and data supporting the findings of this study are available from the corresponding author upon reasonable request<ref:2610.01824#pg4>. Author Said Lantigua (SL) conceived the project; subsequently, authors José Valencia (JV), Gilson Giraldi (GG), Renato Portugal (RP), and Jonas Maziero (JM) contributed to the development of the research<ref:2610.01824#pg4>. SL and JV carried out the formal derivations that provide the mathematical foundation for the project, under the supervision of GG, RP, and JM<ref:2610.01824#pg4>. SL and JV actively worked on drafting the initial version of the manuscript as well as subsequent versions, which were improved and revised by GG, RP, and JM<ref:2610.01824#pg4>. The funding was provided through the National Council for Scientific and Technological Development (CNPq), which initially funded this work through the Institutional Capacity Building Program (PCI), process no. 301066/2025-6, granted by the Ministry of Science, Technology and Innovation (MCTI), and subsequently through the Junior Postdoctoral Fellowship (PDJ) program under process no. 150351/2026-7<ref:2610.01824#pg4>. The authors declare no competing interests<ref:2610.01824#pg4>.

Improvements for AI systems

  1. Improved training stability through informational regularization: The system can maintain quantum richness during optimization by utilizing a quantum-informational regularization scheme based on the dynamics of quantum coherence and state fidelity, which penalizes loss of coherence while controlling changes in the quantum-state trajectory.

  2. Enhanced expressivity in high-dimensional PDE learning: By integrating parameterized quantum circuits via a multiplicative-and-additive coupling (MAC) of classical neural representations and quantum expectation values, the model can retain the universal approximation capability of operator networks under appropriate assumptions.

  3. Mitigation of optimization difficulties: The hybrid architecture is designed to avoid training pitfalls by providing analytical guarantees, specifically deriving bounds on the gradients with respect to the quantum parameters, providing conditions under which exponentially vanishing gradients can be avoided and establishing a scaling criterion for trainability.

  4. Quantifiable sensor requirements: The model can determine the necessary data density for accurate approximation by deriving a lower bound on the number of sensors required to achieve a prescribed approximation accuracy, explicitly relating sampling requirements to the regularity and dimensionality of the input functions and to the dynamical growth of the underlying PDE.

  5. Guaranteed convergence toward quantum-rich solutions: A regularization term, LINF = λCOE X p∈IP 1 − Cp(ϱ(p)Q + λF ID X p∈IP 1 − Fpϱ(p-1)Q, ϱ(p)Q, ensures the model converges to a state that "preserves an entropy value Sp (ϱQ) > Sthreshold and a coherence Cp (ϱ(p)Q) > 0, ensuring that the final solution maintains the requisite quantum richness for the approximation of complex operators."

  6. Robustness against barren plateaus: The architecture actively mitigates optimization instability by showing that it mitigates the barren plateaus problem, with gradient bounds derived from established results such as E(ΞB∗,ΞT ∗) ∇L(v˜, y, OˆB, OˆT,WB∗,WT∗, ΞB∗, ΞT ∗) ≥ O 1 poly NB !.

  7. Optimal resource allocation for sensor placement: The system can calculate the minimum number of sensors m required to guarantee that G v˜, OˆB,WB∗, ΞB∗ ·, OˆT,WT∗, ΞT ∗ − G (v) (·) ∞,M ≤ ϵ˜, which provides a rigorous method for designing efficient sensor networks.

Sources

Related papers