An Annealing-based Quantum Neural Network with Discrete Wavelet Transforms for Enhanced Air Pollution Nowcasting

AI summary

Overview

Research area: Quantum machine learning applied to environmental science — specifically quantum neural networks (QNNs) for short-term air pollution forecasting (PM2.5 nowcasting).

Technical level: Advanced. The paper assumes familiarity with parameterized quantum circuits, ansätze, QUBO formulations, quantum annealing, and discrete wavelet transforms, alongside standard deep learning benchmarking practice.

Scope: The paper presents a hybrid gate-based/annealing QNN framework, translated onto the Dynex neuromorphic quantum computing platform, and benchmarks it against classical MLP and LSTM baselines for 6-hour-ahead PM2.5 prediction using urban air quality data from Guildford, UK.

What This Paper Is About

Air pollution, and fine particulate matter (PM2.5) in particular, is a major public health threat, and accurate short-term forecasting is difficult because environmental time series are non-stationary and full of extreme spikes. The authors ask whether a quantum neural network — made scalable by translating gate-based circuits into optimization problems solved on a neuromorphic annealing platform — can forecast PM2.5 more robustly than standard classical deep learning models, especially on raw, unprocessed data. They also test whether adding discrete wavelet decomposition to the inputs changes that comparison.

Key Contributions

1. A scalable hybrid QNN framework that couples variational quantum circuits with neuromorphic quantum annealing, intended to work around the qubit limits of NISQ-era hardware.
2. A demonstration that the QNN is resilient to input volatility, outperforming classical MLP and LSTM baselines on raw, unprocessed temporal environmental data.
3. A scalability validation up to 172 logical qubits (translated into approximately 2,228 annealing variables for the depth-4 case), with state-of-the-art predictive fidelity when combined with multi-level Discrete Wavelet Transforms.
4. An observation of a non-linear relationship between model complexity and performance: increasing qubit count or lookback period did not produce proportional gains.

Main Findings

• Raw-data regime favours the QNN: On raw inputs, the classical MLP and LSTM baselines did not converge on stable representations, with NMSE values above 1.2 and weak correlations (ρ < 0.35). The depth-4 QNN reached a Pearson correlation of ρ = 0.599 at Lookback 1 (16 qubits), roughly double the LSTM's ρ = 0.288.
• Wavelet features lift all models: Applying a five-level DWT improved performance across model types and narrowed the gap between classical and quantum approaches. The 172-qubit depth-4 QNN (Wavelet Dataset, Lookback 2) recorded the study's highest average correlation (ρ = 0.729) and lowest mean error (NMSE = 0.488) across the full 6-hour horizon.
• Depth does not scale monotonically: On Wavelet data with Lookback 1, Depth 3 gave the best balance (ρ = 0.760, NMSE = 0.496), while Depth 4 was slightly worse (ρ = 0.708, NMSE = 0.545). The authors suggest excessive parametrization without sufficient temporal context may cause diminishing returns or overfitting in the quantum landscape.
• Classical MLP is competitive once DWT is applied: With wavelet-processed features, the MLP nearly matched the QNN (for example, ρ = 0.718 versus 0.760 at Lookback 1, and ρ = 0.695 versus 0.729 at Lookback 2), which the authors interpret as DWT classically performing the decomposition the QNN appeared to handle internally.
• Better uncertainty quantification: Using 95% prediction intervals derived from quantile regression, the QNN's uncertainty bounds expanded dynamically to encompass extreme spikes, whereas classical baselines often produced overly confident but erroneous predictions.
• Statistical significance: Shapiro-Wilk tests rejected normality of the error distributions (W in [0.533, 0.955], p ≪ 0.05, peaking at p = 1.20 × 10⁻¹²¹). Kruskal-Wallis tests showed distinct error populations across configurations (p = 0.000), and pairwise Wilcoxon Signed-Rank tests confirmed the depth-4 QNN's superiority in the Wavelet/Lookback 2 setting against all other variations and classical baselines (p = 0.000).
• Translation overhead quantified: The 172-logical-qubit depth-4 circuit was estimated to map onto roughly 172 + 688 + 1,368 = 2,228 annealing variables, exceeding the coherence and connectivity limits of current physical gate-based QPUs.

Methodology in Plain English

The authors built a quantum neural network in the standard variational style. Each input feature is assigned to one qubit: Hadamard gates put all qubits into superposition, then a Y-rotation encodes each normalized feature value into its qubit's angle. A trainable "ansatz" then applies repeated layers of single-qubit rotations (Rz, Ry, Rz with trainable angles) followed by a circular pattern of CNOT gates between neighbouring qubits. Measurements of the qubits' expectation values are passed into a small classical fully connected layer with ReLU activation, which produces predictions for the chosen forecast horizon.

To get around the fact that real quantum hardware cannot handle circuits of this size, the circuit is converted into a Quadratic Unconstrained Binary Optimization problem — a Hamiltonian whose ground state encodes the circuit's output. This is solved using DynexQML, which turns ordinary GPUs into emulated quantum annealers through memristive emulation. Training uses energy-based gradients computed by finite differences, with parameters updated by gradient descent; quantum sampling happens at epoch boundaries rather than continuously.

The data came from two iSCAPE air-quality monitoring stations in Guildford, UK (Sutherland Memorial Park and Stoke Park), with hourly measurements of PM2.5, CO, NO2, O3, air temperature, relative humidity and atmospheric pressure, plus sine/cosine encodings of hour, day and month and static latitude, longitude and altitude. Extreme spikes up to 7,448 μg/m³ (against a dataset average of about 7.957 μg/m³) were clipped at 1,000 μg/m³ to stabilise training. July was held out as the test set for both stations; the rest was split 80% training and 20% validation, with min-max scaling to [0, 1] fitted only on the training subset.

Two parallel input streams were compared: raw features and wavelet-decomposed features. A five-level DWT was applied to each continuous feature, producing five detail signals and one approximation signal and expanding the input space six-fold, with the mother wavelet chosen per feature by minimizing reconstruction RMSE. The classical baselines were an MLP with 2,025 trainable parameters and an LSTM with 64 hidden units and 61,885 trainable parameters, both trained for 300 epochs with Adam at a learning rate of 0.001, using MSE loss and tracking MAE.

Why This Matters

Impact on research: The paper argues that quantum machine learning has been stuck on toy problems and that the field needs benchmarks at industrially relevant scale before claims about advantage or sustainability can be meaningful. It demonstrates a hardware-agnostic emulation route to testing 172-qubit-class algorithms today, and it reframes the debate from speculative "quantum advantage" toward measuring a concrete "energy-to-solution" ratio. Its most provocative result is that the QNN helps most precisely where classical models struggle — on raw, noisy, high-entropy data — implying quantum models might reduce dependence on domain-specific feature engineering. Notably, the paper does not report actual energy consumption figures for either quantum or classical training, so the sustainability argument is framed as a research agenda rather than a measured result.

Real-world applications:

• Short-term urban air quality warning systems that give residents and authorities hours of lead time before PM2.5 spikes.
• Public health monitoring and exposure-reduction advisories, given that no level of PM2.5 exposure is considered safe.
• Traffic and industrial emissions management, where nowcasts can inform temporary interventions.
• Environmental monitoring networks in data-scarce settings, where models that cope with raw, unprocessed signals reduce instrumentation and preprocessing burden.

Industry relevance: Cloud and GPU providers offering quantum-annealing emulation as a service; environmental consultancies and smart-city platform vendors; hardware and framework developers needing realistic algorithmic benchmarks to de-risk future-scale quantum investment.

Future Directions

• Measuring actual energy-to-solution for the 172-qubit QNN against classical baselines, which the paper frames as the next necessary step but does not itself report.
• Determining when added circuit depth helps versus overfits — the non-monotonic depth results (Depth 3 best at Lookback 1 with wavelets, Depth 4 best at Lookback 2) leave the depth/lookback trade-off unresolved.
• Testing whether the QNN's apparent internal handling of volatile raw signals generalises beyond PM2.5 and beyond this pair of Guildford stations.
• Validating these emulation-based results on physical gate-based quantum hardware as qubit counts, connectivity and coherence improve.

Target Audience

Quantum machine learning researchers and practitioners interested in real-world, large-qubit benchmarks; environmental and atmospheric scientists working on air quality forecasting; and AI sustainability researchers and industry strategists evaluating whether quantum methods are ready for industrially relevant problems. Readers without a background in variational circuits, QUBO formulations or wavelet decomposition will find the methodology sections demanding.

Authors’ abstract

Air pollution poses a significant threat to public health, exacerbated by the rapid industrialization and urbanization of modern society. While traditional deep learning offers predictive capabilities, recent advancements in quantum machine learning (QML) present a compelling alternative. Here, we introduce a scalable hybrid framework that integrates variational quantum circuits with quantum annealing. Leveraging a neuromorphic quantum computing platform to transcend the qubit limitations of the NISQ era, we model particulate matter (PM$_{2.5}$) dynamics using high-density urban data from the UK. We demonstrate that our Quantum Neural Networks (QNNs) exhibit an intrinsic resilience to input volatility, significantly outperforming classical baselines (MLP and LSTM) on raw, high-entropy environmental data. Furthermore, we successfully scaled the model to 172 qubits, observing a synergistic enhancement when coupled with wavelet decomposition that achieves state-of-the-art predictive fidelity. These findings establish a critical benchmark for quantum utility, proving that scalable, neuromorphic quantum machine learning can now address complex, non-linear real-world problems.