⚛️ quantum physics

Quantum Convolutional Neural Networks for Groundwater Heat Plume Prediction: A Surrogate Modeling Approach

This paper proposes a Quantum Convolutional Neural Network (QCNN) as a surrogate model for predicting groundwater heat plume dynamics in Munich, demonstrating that while classical networks currently offer superior accuracy, the QCNN achieves competitive performance on quantum simulators and shows promising improvements under error-mitigated hardware conditions.

Original authors: Danyal Maheshwari, Julia Pelzer, Miriam Schulte

Published 2026-06-23

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Danyal Maheshwari, Julia Pelzer, Miriam Schulte

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

The Big Picture: Predicting Underground Heat Waves

Imagine the ground beneath the city of Munich is like a giant, invisible river of water. People are installing heat pumps (like giant air conditioners for the ground) to heat and cool their homes. These pumps inject warm or cold water into the ground, creating "heat plumes"—think of them like invisible ink stains spreading through a wet sponge.

The city needs to know exactly how far these stains will spread and how hot they will get. If they spread too far or get too hot, they could mess up the natural balance of the underground water or interfere with other pumps.

The Problem: The "Slow Cooker" vs. The "Microwave"

To figure out where these heat plumes go, scientists usually use super-accurate computer simulations. However, these simulations are like slow cookers. They take a very long time to run. If you want to test 1,000 different scenarios (like moving the pump slightly or changing the soil type), you'd have to wait forever.

To solve this, scientists use "surrogate models." These are like microwaves. They aren't as perfect as the slow cooker, but they give you a result almost instantly. The goal of this paper is to see if a new type of "microwave"—powered by Quantum Computers—can do the job better than the old ones.

The Experiment: Training a Quantum Student

The researchers built a special type of Artificial Intelligence called a Quantum Convolutional Neural Network (QCNN). You can think of this as a student trying to learn a map of the underground heat.

The Input (The Lesson): Instead of feeding the computer a massive, detailed map of the ground, they gave it a tiny, simplified summary: just two numbers (how slippery the ground is and how hard the water is pushing).
The Output (The Test): The computer had to guess four things: How long the heat stain is, how wide it is, how hot the hottest spot gets, and exactly where that hot spot is.
The Classroom (The Hardware): They tested this student in four different "classrooms":
- The Perfect Classroom (Ideal Simulator): A computer simulation where everything works perfectly, with no mistakes.
- The Noisy Classroom (Noisy Simulator): A simulation that mimics a real, imperfect computer.
- The Real World (IBM Quantum Computer): They actually ran the code on a real quantum processor in Kyiv (IBM's 127-qubit chip).
- The Real World with Safety Nets (Error Mitigation): The same real computer, but with special tricks to fix mistakes as they happen.

The Results: How Did the Student Do?

1. The Perfect Classroom:
In the ideal simulation, the Quantum student did very well. It learned the rules quickly and could predict the heat stain's shape and temperature accurately. It proved that the idea works.

2. The Real World (Without Safety Nets):
When they ran the test on the actual IBM quantum computer without any help, the student got very confused. Real quantum computers are like tuning forks in a hurricane; they are extremely sensitive to noise (vibrations, heat, electrical interference). The results were messy and inaccurate. The "heat stain" it predicted looked distorted and wrong.

3. The Real World (With Safety Nets):
When they added "error mitigation" (the safety nets), the student improved significantly. The predictions became much clearer, and the shape of the heat stain looked more like the real thing. It wasn't perfect, but it was much better than before.

4. The Comparison (The Old Way):
Throughout all these tests, they compared the Quantum student to a Classical Neural Network (a standard AI running on a regular laptop).

The Winner: The Classical AI won every time. It was faster, more accurate, and didn't get confused by the noise.
The Takeaway: The Quantum AI is currently a "promising student" that shows potential, but it hasn't beaten the "veteran teacher" (the Classical AI) yet.

The Conclusion

The paper concludes that while Quantum Computers are exciting and show they can learn these patterns in theory, they are currently too noisy and error-prone to replace the standard tools used by engineers today.

However, the fact that the Quantum model improved significantly when they added error-correction tricks is a good sign. It suggests that as quantum computers get better and less "noisy" in the future, they might one day become a powerful tool for predicting groundwater temperatures. For now, though, the old-fashioned computers are still the best choice for the job.

Technical Summary: Quantum Convolutional Neural Networks for Groundwater Heat Plume Prediction

Problem Statement
The management of shallow groundwater heat pumps (GWHPs) in urban environments like Munich requires precise modeling of subsurface temperature fields to prevent efficiency losses and ecological damage. While high-fidelity numerical simulations (e.g., using Pflotran) provide accurate predictions of heat plume dynamics, they are computationally expensive and unsuitable for real-time applications or the optimization of large numbers of heat pump constellations. Conversely, existing analytical models (such as the Linear Advective Heat transport Model, LAHM) are fast but lack accuracy, particularly in heterogeneous aquifers or complex flow scenarios. Deep learning surrogate models, specifically Convolutional Neural Networks (CNNs), have emerged as a compromise, offering high speed with reasonable accuracy. However, the integration of Quantum Machine Learning (QML) into this domain remains unexplored. This work investigates whether Quantum Convolutional Neural Networks (QCNNs) can serve as effective surrogate models for predicting groundwater heat plume characteristics, addressing the scalability constraints of current Noisy Intermediate-Scale Quantum (NISQ) hardware.

Methodology
The study employs a hybrid quantum-classical approach to model heat plumes generated by a single open-loop GWHP.

Dataset Generation: A dataset of 352 samples was generated using the Pflotran software. The simulations cover a domain of $100 \times 1280$ meters with a resolution of 5 meters ( $20 \times 256$ cells). Input variables include hydraulic gradients and permeability fields, while the output is a spatially resolved temperature field.
Dimensionality Reduction: To comply with the qubit limitations of current quantum hardware, the high-dimensional simulation data was compressed. The input was reduced from spatial fields to two scalar variables (permeability and pressure gradient). The output temperature field was reduced to four representative parameters: the X and Y extents of the 1-Kelvin isoline, the maximum groundwater temperature, and the spatial position of this maximum.
QCNN Architecture: The proposed QCNN consists of three main stages:
1. State Preparation: Classical data is encoded into quantum states using an Instantaneous Quantum Polynomial (IQP) encoding scheme. This Hamiltonian-inspired strategy utilizes commuting gates to embed classical information into the phases of the quantum state.
2. Quantum Processing: The architecture includes a quantum convolutional layer and a quantum pooling layer. The convolutional layer uses parameterized unitary blocks (single-qubit rotations and entangling gates like CNOT) to extract localized features. The pooling layer reduces the Hilbert space dimensionality via measurement and partial trace operations, analogous to classical max- or average-pooling.
3. Readout and Classical Post-Processing: The final quantum state is measured to obtain expectation values, which are fed into a classical fully connected neural network (implemented in PyTorch) to perform the final regression and map quantum features to the four target scalar values.
Training and Evaluation: The model was trained using the Adam optimizer with a Mean Squared Error (MSE) loss function. Experiments were conducted across four distinct execution regimes:
1. Ideal statevector simulator.
2. Noisy simulator with realistic noise models.
3. IBM's 127-qubit Kyiv quantum processor (raw hardware).
4. IBM Kyiv processor with advanced error-mitigation techniques (including dynamical decoupling, Twirled Readout Error Extinction, Zero-Noise Extrapolation, and Probabilistic Error Cancellation).

Key Results
The performance of the QCNN was benchmarked against a classical CNN baseline across the different execution environments:

Ideal Simulation: On the statevector simulator, the QCNN demonstrated the ability to learn the mapping from the four input features to the four target descriptors, achieving competitive performance with low MSE.
Noisy Simulation: The noisy simulator showed a slight degradation in performance compared to the ideal case but maintained a stable learning trajectory, indicating that moderate noise levels are tolerable.
Real Hardware (Raw): On the raw IBM Kyiv processor, the QCNN struggled significantly. The MSE was markedly higher than in simulations, and the model failed to generalize, producing distorted heat plume reconstructions. This highlights the devastating impact of gate imperfections, decoherence, and crosstalk on non-trivial regression tasks in the NISQ era.
Real Hardware (Mitigated): The application of advanced error-mitigation techniques on the Kyiv backend resulted in a noticeable improvement. The training and testing losses decreased, and the visual reconstruction of the heat plume became more consistent with the target patterns. However, the performance did not fully recover to the level of the ideal simulator.
Classical Baseline: The classical CNN consistently outperformed all quantum variants across all settings, achieving the lowest MSE and the most accurate spatial reconstructions.

Significance and Claims
The paper posits that while QCNNs are not yet ready to replace classical surrogate models for groundwater heat plume prediction, they represent a promising direction for future research. The study demonstrates that:

Feasibility: QCNNs can theoretically learn complex mappings from compact input features to spatial descriptors, provided the hardware is noise-free.
Hardware Limitations: Current NISQ hardware noise severely degrades performance for regression tasks, rendering raw quantum execution ineffective for this specific application.
Role of Mitigation: Error-mitigation techniques are essential for extracting meaningful signals from noisy quantum devices, partially recovering performance and enhancing the physical interpretability of the results.
Future Viability: As quantum hardware matures and more efficient mitigation and compilation techniques are developed, QCNNs may become viable for high-dimensional hydrothermal prediction, particularly when integrated into hybrid quantum-classical workflows.

The authors conclude that the primary value of this work lies in establishing a testbed to understand quantum architecture behavior under realistic device conditions, rather than claiming immediate superiority over classical methods.