💻 computer science

Conformalized Quantum DeepONet Ensembles: Towards Scalable Operator Learning with Distribution-Free Guarantees

This paper introduces Conformalized Quantum DeepONet Ensembles, a framework that leverages Quantum Orthogonal Neural Networks to achieve scalable operator learning with reduced computational complexity while employing split conformal calibration to provide rigorous, distribution-free uncertainty guarantees for safety-critical applications.

Original authors: Purav Matlia, Christian Moya, Guang Lin

Published 2026-10-16

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Purav Matlia, Christian Moya, Guang Lin

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of modern science and engineering, many problems involve predicting how complex systems will behave over time. Think of a power grid managing electricity flow, or a weather model tracking a storm. These systems are governed by laws of physics that can be written down as equations, but solving those equations for every possible scenario is often too slow and expensive for real-time decision-making. To speed things up, scientists have developed "surrogate models," which are simplified, learned versions of these complex systems. These models can make predictions almost instantly, acting as a fast-forward button for simulations. However, there is a catch: when these models are used for safety-critical tasks, like keeping a power grid stable, knowing what the model predicts is not enough. Engineers also need to know how much they can trust that prediction. If a model is unsure, it needs to say so, and it needs to provide a reliable range of possible outcomes. Furthermore, as these systems become more detailed, the computer power required to run the models grows so fast that it can become impossible to use them at all.

A team of researchers at Purdue University has tackled these two problems—speed and reliability—by combining two distinct fields: quantum computing and a statistical method for measuring uncertainty. They created a new framework called a "Conformalized Quantum Deep Operator Network." In simple terms, this is a type of artificial intelligence designed to learn the rules of how systems change, but it is built using the principles of quantum mechanics to run faster, and it is wrapped in a statistical safety net to ensure its predictions are trustworthy. The researchers did not just build a single model; they built a group of these models working together, a technique known as an ensemble, to better gauge how uncertain they are. Their work shows that while these quantum models are still in the early stages of development, they can achieve high accuracy and provide mathematically guaranteed safety margins, even when simulated on noisy, imperfect hardware.

The core of the problem the researchers addressed is that traditional artificial intelligence models, which use standard computer chips, struggle when asked to simulate complex, continuous systems. These models rely on layers of calculations that become exponentially more expensive as the system gets more detailed. If you want to predict the behavior of a system with high precision, the time it takes to run the simulation can grow so large that it defeats the purpose of having a fast model. The researchers turned to quantum computing to solve this bottleneck. They utilized a specific type of quantum neural network, which uses the unique properties of quantum bits, or qubits, to perform calculations in a way that is theoretically much more efficient for these specific types of problems. Instead of the standard computer chips that require a quadratic amount of work to process data, their quantum approach scales much more gently, offering a potential path to handling the massive amounts of data required for fine-grained simulations.

However, simply making a model faster does not make it safe to use in critical situations. In fact, quantum computers are currently prone to errors and noise, which can make their outputs seem random or unreliable. To address this, the researchers combined their quantum models with a technique called conformal prediction. This is a statistical method that does not rely on guessing the probability of an error based on past data distributions. Instead, it uses a separate set of data to calibrate the model, creating a prediction interval that is guaranteed to contain the true answer a certain percentage of the time, regardless of how the data is distributed. By training multiple quantum models independently and then using this calibration method, the team created a system that not only predicts outcomes but also draws a box around those predictions, stating with mathematical certainty that the true value lies inside that box.

The researchers tested this framework on several different challenges. First, they used synthetic benchmarks, such as a mathematical task involving the integration of functions and a simulation of a wave moving through space. In these ideal, noise-free simulations, the system performed with high accuracy, achieving a relative error of less than half a percent for the integration task and about two percent for the wave simulation. More importantly, the system successfully met its reliability targets. When the researchers asked the model to provide a 90 percent confidence interval, the actual results fell within that interval roughly 88 to 92 percent of the time, closely matching the goal. This demonstrated that the method could provide the rigorous safety guarantees needed for real-world applications.

To ensure the findings were robust, the team also simulated the effects of real-world hardware imperfections. Quantum computers today are not perfect; they suffer from "noise" that can corrupt calculations. The researchers tested their models under various levels of simulated noise, including specific error patterns found in real quantum processors from major manufacturers. Even with these imperfections, the system remained remarkably stable. The prediction intervals adjusted appropriately, and the coverage rates stayed near the target levels. In some cases, the presence of noise even seemed to have a subtle, unexpected effect on the model's behavior, though the researchers noted this was preliminary and required further study. They also explored different ways to run the models, such as splitting the work between classical and quantum parts of the computer, or using a technique to run multiple models on a single quantum circuit to save resources. These experiments showed that while there are trade-offs between speed, accuracy, and the number of quantum bits required, the framework is flexible enough to adapt to different hardware constraints.

Finally, the researchers applied their method to a real-world scenario: power system dynamics. They trained the models to predict how voltage and power flow in an electrical grid would change in response to disturbances. This is a critical task for preventing blackouts and maintaining grid stability. The models were tested on both offline tasks, where the entire history of a system is analyzed at once, and online tasks, where predictions must be made in real-time as new data arrives. In these tests, the quantum ensemble models achieved relative errors ranging from about 4 percent to 12 percent, depending on the complexity of the task. Crucially, the uncertainty estimates remained reliable, with the actual outcomes falling within the predicted ranges nearly 90 percent of the time, just as the theory suggested. This proved that the framework could handle the messy, high-stakes data of a real power grid, not just clean mathematical puzzles.

The work presented in this study is a significant step forward in the field of scientific machine learning, but it comes with important caveats. The researchers are careful to note that their results regarding speed are based on theoretical analysis and simulations, not on running the models on a physical quantum computer that is faster than a classical one. The current quantum hardware is not yet powerful enough to outperform classical supercomputers for these tasks in a practical, real-world setting. The primary achievement here is the demonstration of a complete framework that successfully combines quantum efficiency with rigorous statistical guarantees. The researchers have shown that it is possible to build a system that is both resource-aware and uncertainty-aware, providing a blueprint for how these technologies might be used together in the future.

As the technology matures, this approach could become a vital tool for engineers and scientists who need to make fast, reliable decisions in complex environments. By proving that quantum models can be calibrated to provide trustworthy safety margins, the researchers have removed a major barrier to the adoption of these technologies in safety-critical fields. The study does not claim to have solved the problem of quantum computing noise or to have built a machine that is ready for immediate deployment in a power plant. Instead, it offers a clear, mathematically sound path forward, showing that when quantum learning is paired with robust statistical methods, it can deliver predictions that are not only fast but also honest about their own limitations. This combination of speed and trustworthiness is the key to unlocking the potential of quantum machine learning for the most important challenges in science and engineering.

Technical Summary: Conformalized Quantum DeepONet Ensembles

Problem Statement
Operator learning offers a pathway to fast surrogate modeling for high-dimensional dynamical systems, such as those found in power systems. However, existing classical approaches face two critical bottlenecks: (1) Scalability, driven by the quadratic $O(n^2)$ computational cost of dense neural layers in architectures like DeepONet, which becomes prohibitive when evaluating operators across thousands of query coordinates for uncertainty quantification; and (2) Reliability, as standard uncertainty quantification (UQ) methods often lack rigorous, distribution-free guarantees required for safety-critical deployment. While quantum machine learning has proposed architectures to reduce layer complexity, they typically yield noisy point predictions without robust epistemic uncertainty estimates.

Methodology
The authors propose the Conformalized Quantum Deep Operator Network (DeepONet) Ensemble, a framework integrating three core components:

Quantum Orthogonal Neural Networks (QOrthoNNs): The authors replace classical dense hidden layers with QOrthoNNs. These layers utilize Reconfigurable Beam Splitter (RBS) gates operating in a unary subspace to perform orthogonal transformations. The architecture consists of a branch network (encoding input functions) and a trunk network (encoding query coordinates), both implemented as Quantum DeepONets.
Ensemble-Based Uncertainty Quantification: To capture epistemic uncertainty, the framework employs an ensemble of $L$ independently trained Quantum DeepONets. Predictions are aggregated to compute a mean $\mu(u)(y)$ and a standard deviation $\sigma(u)(y)$ , which serves as a proxy for model disagreement.
Split Conformal Calibration: The ensemble outputs are calibrated using split conformal prediction. A nonconformity score is defined as the absolute error normalized by the ensemble standard deviation. Using a calibration dataset, a threshold $b_q$ is computed to construct prediction intervals $C_\alpha(u, y) = [\mu(u)(y) \pm b_q \tilde{\sigma}(u)(y)]$ .

Key Contributions

Resource-Aware Complexity Analysis: The paper refines the complexity analysis of Quantum DeepONets (building on Xiao et al., 2025). By accounting for classical parallelism and quantum resources, the authors establish an explicit resource regime where QOrthoNNs offer an asymptotic running-time improvement over classical dense layers. Specifically, under a work-depth model where the number of classical processors $P$ scales as $P = o(n/\log n)$ , the quantum depth-shot cost of $\tilde{O}(n)$ outperforms the classical parallel time of $O(n^2/P + \log n)$ . This is a theoretical complexity result, not a measured wall-clock speedup.
Finite-Sample, Distribution-Free Guarantees: The authors prove a finite-sample lower bound on the query-averaged marginal coverage for a new input–output function pair. Unlike standard conformal prediction which assumes exchangeability of individual points, this proof assumes exchangeability of complete input-output pairs (units), allowing for arbitrary dependence among query coordinates within a single unit. The bound is given by $\frac{r}{(K+1)M}$ , where $K$ is the number of calibration units, $M$ is the number of queries per unit, and $r$ is a quantile index.
Exploratory Execution Strategies: To manage the trade-off between runtime (sequential execution) and qubit requirements (parallel execution), the paper explores:
- Hybrid Architectures: Partitioning the network such that the branch (computed once per input) is classical, while the trunk (computed per query) remains quantum.
- Multiplexed Execution: A proof-of-concept implementation inspired by Superposed Parameterized Quantum Circuits (SPQCs) that uses address-conditioned routing to execute multiple ensemble members within a single circuit, reducing qubit usage from $O(Ln)$ to $O(n + \log L)$ at the cost of increased circuit depth.

Experimental Results
The framework was evaluated on synthetic operator benchmarks (Antiderivative and Advection equations) and real-world power-system dynamics (Transient-to-Transient, Causal Forecasting, and Causal Pointwise Transient operators).

Ideal Simulations: Under noiseless conditions (simulated via mathematically equivalent classical orthogonal networks), the ensembles achieved high predictive accuracy (e.g., 0.46% relative $L_2$ error for the Antiderivative task) and empirical coverage near the 90% target (e.g., 92.13%).
Noise Simulations: Experiments included gate-based depolarizing noise and device-calibrated composite noise models for IBM Eagle, Heron, and Brisbane processors. The framework maintained coverage near the target across various shot counts and noise strengths. In some online voltage prediction tasks, increasing noise levels resulted in narrower intervals and higher coverage, suggesting a potential (though preliminary) noise-induced regularization effect.
Hybrid and Multiplexed Performance: Hybrid classical-branch architectures yielded the lowest error by removing quantum noise from the shared representation. The multiplexed SPQC-inspired circuit successfully tracked the performance of a standard ensemble with significantly fewer qubits (9 qubits vs. 28 for $L=4$ ), validating the qubit-depth trade-off.

Significance and Claims
The paper claims to connect resource-aware scalability analysis with finite-sample uncertainty guarantees for quantum operator learning. It demonstrates that:

QOrthoNNs can theoretically provide an asymptotic hidden-layer speedup over classical dense layers within specific resource regimes.
Conformal calibration can be successfully applied to quantum operator ensembles to provide rigorous, distribution-free coverage guarantees for query-averaged predictions, even when queries within a function are dependent.
Hybrid and multiplexed execution strategies offer viable pathways to manage the computational and hardware constraints of quantum ensembles.

The authors explicitly state that their complexity claims are asymptotic and layer-level, not end-to-end runtime guarantees on current hardware. Similarly, the empirical results are based on simulations (ideal and noisy) rather than physical quantum hardware execution, and the observed noise-induced regularization effects are noted as preliminary and task-specific. The work serves as a proof-of-concept for scalable, uncertainty-aware quantum operator learning.

Technical Summary: Conformalized Quantum DeepONet Ensembles

More like this