🤖 machine learning

Conformalized Quantum DeepONet Ensembles for Scalable Operator Learning with Distribution-Free Uncertainty

This paper introduces Conformalized Quantum DeepONet Ensembles, a framework that utilizes Quantum Orthogonal Neural Networks and Superposed Parameterized Quantum Circuits to achieve scalable operator learning with linear inference complexity and rigorous, distribution-free uncertainty quantification for safety-critical applications.

Original authors: Purav Matlia, Christian Moya, Guang Lin

Published 2026-05-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Purav Matlia, Christian Moya, Guang Lin

Original paper licensed under CC BY 4.0 (http://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

Imagine you are trying to predict how a complex machine, like a power grid or a weather system, will behave in the future. Scientists usually use massive computer simulations to do this, but they are slow and expensive. To speed things up, researchers build "surrogate models"—smart AI shortcuts that learn the rules of the machine and can predict the future almost instantly.

However, these AI shortcuts have two big problems:

They get too slow too fast: As the details get finer, the computer work explodes, like trying to count every grain of sand on a beach one by one.
They are overconfident: They give you an answer but can't tell you how much they might be wrong. In safety-critical situations (like keeping a power grid from crashing), not knowing the risk is dangerous.

This paper introduces a new solution called Conformalized Quantum DeepONet Ensembles. Think of it as a "super-charged, self-checking prediction team" that uses the strange laws of quantum physics to solve these problems.

Here is how it works, broken down into simple concepts:

1. The Quantum Speed-Up (The "Magic Lens")

Traditional AI models are like a librarian who has to walk down every single aisle of a massive library to find a book. As the library grows, the time it takes grows exponentially (quadratically).

The authors use a Quantum DeepONet, which is like giving that librarian a "magic lens." Instead of walking down aisles one by one, the quantum lens can look at the whole library at once. This changes the speed from "walking every aisle" to "glancing at the whole room," making the predictions much faster and scalable, even for very detailed simulations.

2. The Prediction Team (The "Ensemble")

Even with a fast quantum model, a single AI can still be wrong or unsure. To fix this, the researchers don't just use one model; they use a team of models (an ensemble).

The Analogy: Imagine asking 10 different experts to guess the weather. If they all agree, you are confident. If they disagree wildly, you know there is uncertainty.
The Problem: Usually, running 10 different quantum models requires 10 times the quantum hardware (qubits), which is impossible with current technology. It's like trying to build 10 separate super-computers in your garage.

3. The "Superposition" Trick (The "Ghost Team")

This is the paper's cleverest trick. Instead of building 10 separate quantum computers, they use something called Superposed Parameterized Quantum Circuits (SPQCs).

The Analogy: Imagine you have a single, magical coin. Instead of flipping it 10 times to get 10 results, you flip it once, and because of quantum magic (superposition), it lands on "Heads" and "Tails" (and all the variations in between) simultaneously.
The Result: They can run all 10 "experts" (models) inside a single quantum circuit at the same time. This saves massive amounts of hardware resources while still getting the benefit of having a diverse team of predictors.

4. The Safety Net (The "Conformal Guarantee")

Having a team of experts is great, but how do you know the range of their answers is trustworthy? The paper adds a layer called Conformal Prediction.

The Analogy: Imagine the team of experts gives you a forecast. The Conformal method acts like a "safety inspector" who looks at past mistakes. It says, "Based on how we've done before, I can guarantee with 90% certainty that the real answer will fall inside this specific range."
The Benefit: This gives a mathematically rigorous "safety tube" around the prediction. It doesn't just guess; it tells you exactly how wide the margin of error is, without needing to assume the data follows a specific pattern (distribution-free).

5. The Hybrid Approach (The "Best of Both Worlds")

The researchers realized that some parts of the calculation are noisy on quantum computers, while others are fine.

The Strategy: They built a Hybrid system. They let the classical (regular) computer handle the parts that are easy and don't need quantum speed, and they only use the quantum "magic lens" for the heavy lifting.
The Result: This removes the "noise" (errors) from the quantum parts that aren't necessary, making the final prediction cleaner and more accurate.

What Did They Prove?

The authors tested this system on two types of challenges:

Fake Math Problems: They solved complex equations (like how heat moves or how waves travel) to prove the math works.
Real Power Grids: They tested it on real-world data from electrical power systems.

The Results:

The system was accurate.
The "safety tubes" (uncertainty ranges) were reliable, hitting their target coverage rates (e.g., being right 90% of the time) even when the quantum computer was "noisy" (imperfect).
The "Superposition" trick worked: they got the benefits of a large team of models without needing 10 times the hardware.

In Summary

This paper presents a new way to build AI for complex physical systems. It combines quantum speed (to handle big data), team-based prediction (to understand uncertainty), and a mathematical safety net (to guarantee reliability). It solves the problem of quantum computers being too small to run large teams of models by using a "ghost team" trick, making it possible to have fast, safe, and trustworthy predictions for things like power grids.

1. Problem Statement

The paper addresses two fundamental limitations in current neural operator learning (specifically DeepONets) for high-dimensional dynamical systems, such as power grids and partial differential equations (PDEs):

Scalability Bottleneck: Classical DeepONets rely on dense neural layers with $O(n^2)$ computational complexity. This quadratic scaling makes them prohibitive for safety-critical applications requiring fine discretizations and massive Monte Carlo evaluations (e.g., $10^4$ timesteps).
Unreliable Uncertainty Quantification (UQ): Safety-critical deployments require rigorous uncertainty estimates. Standard methods like Bayesian Neural Networks or Monte Carlo dropout lack distribution-free guarantees. While conformal prediction offers such guarantees, applying it to quantum models is challenging because:
- Naïve quantum ensembles (running $L$ models in parallel) scale qubit requirements linearly ( $O(L \cdot n)$ ), creating a hardware bottleneck.
- Relying solely on quantum measurement shot noise for uncertainty is insufficient, as it captures hardware noise rather than true epistemic uncertainty (model diversity).

2. Methodology

The authors propose Conformalized Quantum DeepONet Ensembles, a framework that integrates scalable quantum inference, resource-efficient ensembling, and rigorous uncertainty calibration.

A. Scalable Quantum Operator Inference

Quantum Orthogonal Neural Networks (QOrthoNNs): The core architecture replaces classical dense layers with parameterized quantum circuits (PQCs) using Reconfigurable Beam Splitter (RBS) gates.
Complexity Reduction: By operating in the unary subspace and utilizing RBS gates, the inference complexity is reduced from $O(n^2)$ to $\tilde{O}(n)$ .
Training Strategy: To avoid hardware noise during training, the model is trained classically. The quantum operations are mathematically equivalent to exact classical orthogonal transformations. The optimization occurs in the rotation angle space ( $\theta$ ), ensuring strict orthogonality and avoiding barren plateaus, before being deployed to quantum hardware for inference.

B. Resource-Efficient Ensembling

To capture epistemic uncertainty without linearly scaling qubit requirements, the authors employ two strategies:

Hybrid Classical-Quantum Architecture: The DeepONet consists of a "Branch" network (processes input functions, evaluated once per scenario) and a "Trunk" network (processes query coordinates, evaluated $M$ $M$ times).
- The Branch is replaced with a classical dense network. Since it is evaluated only once, the $O(n^2)$ cost is amortized, and the dominant quantum noise source is eliminated.
- The Trunk remains quantum to preserve the $\tilde{O}(n)$ speedup for high-frequency evaluations.
Superposed Parameterized Quantum Circuits (SPQCs): Instead of running $L$ $L$ ensemble members in parallel (requiring $L \times n$ $L \times n$ qubits), SPQCs encode all $L$ $L$ models into a single circuit.
- Using address qubits in a uniform superposition, the circuit executes state preparation, unitary evolution, and measurement exactly once for the entire ensemble.
- This reduces qubit requirements to $n + 1 + \lceil \log_2(L) \rceil$ , enabling scalable ensemble execution.

C. Distribution-Free Uncertainty Quantification

Adaptive Conformal Prediction: The framework combines the ensemble's epistemic uncertainty (standard deviation of predictions) with conformal prediction.
Calibration: A nonconformity score is calculated as the absolute error normalized by the local ensemble standard deviation. An empirical quantile ( $\hat{q}$ ) is derived from a calibration set.
Guarantee: This yields prediction sets $C_\alpha(u, y)$ with a strict, distribution-free guarantee that the true value falls within the set with probability $1-\alpha$ , provided data exchangeability holds.

3. Key Contributions

Scalable Quantum Inference: Demonstrated a reduction in operator inference complexity from $O(n^2)$ to $\tilde{O}(n)$ using QOrthoNNs, enabling fine-grained discretization.
Hardware-Efficient Ensembling: Introduced a hybrid architecture and SPQCs to compress ensemble execution, solving the linear qubit scaling problem inherent in parallel quantum ensembles.
Rigorous UQ: Integrated adaptive conformal prediction with quantum ensembles to provide mathematically guaranteed, distribution-free uncertainty bounds.
Robust Evaluation: Validated the framework under realistic quantum noise models (depolarizing, readout errors, thermal relaxation) on synthetic PDEs and real-world power system dynamics.

4. Experimental Results

The framework was evaluated on synthetic benchmarks (Antiderivative, Advection equations) and real-world power system datasets (Voltage-to-Voltage, Voltage-to-Power).

Synthetic Benchmarks:
- Achieved low relative $L_2$ errors (e.g., 0.46% for Antiderivative, 2.28% for Advection).
- Maintained empirical coverage close to the target 90% (e.g., 92.13% and 88.99%) under ideal simulation.
- Noise Resilience: Under realistic noise models (IBM Brisbane, Torino, Marrakesh), the conformal framework maintained coverage $\ge 90\%$ , proving robustness against hardware drift.
Real-World Power Systems:
- Successfully handled offline (post-mortem analysis) and online (real-time forecasting) tasks.
- Online Voltage Prediction: Achieved 90.11% coverage despite violating the exchangeability assumption (due to temporal dependencies), serving as a rigorous stress test.
Architecture Comparisons:
- Hybrid vs. Fully Quantum: The Hybrid (Classical Branch + Quantum Trunk) configuration significantly outperformed fully quantum setups in accuracy and uncertainty width by eliminating the noisy branch bottleneck.
- SPQC vs. Naïve Parallelism: SPQCs matched the predictive accuracy and uncertainty calibration of standard ensembles while drastically reducing qubit usage (e.g., 7 qubits vs. 20 qubits for a 4-member ensemble).

5. Significance and Outlook

Practical Pathway: This work establishes a viable path for deploying quantum machine learning in safety-critical infrastructure by solving the dual problems of scalability and reliability.
Resource Efficiency: The SPQC approach fundamentally changes the resource scaling of quantum ensembles from linear to logarithmic in terms of qubits, making large-scale uncertainty quantification feasible on near-term hardware.
Future Directions: The authors note that while current guarantees rely on data exchangeability, future work must address non-exchangeable streaming data (e.g., via sequential conformal prediction). Additionally, the anomalous observation that noise sometimes improved uncertainty bounds in shallow circuits warrants further investigation into "noise-induced regularization."

In summary, the paper presents a unified framework that leverages quantum speedups for operator learning while using classical-compliant ensembling and conformal prediction to ensure the rigorous safety guarantees required for real-world deployment.