Quantum annealing takes on urban PM2.5 nowcasting: raw-data correlation doubles, 172 qubits pair with wavelet decomposition

PM2.5 nowcasting gains a route that skips hand-built feature engineering: a quantum neural network can learn dynamics directly from raw, undecomposed hourly pollutant series. The dominant approach couples signal decomposition such as empirical mode decomposition or discrete wavelet transforms with CNN and RNN architectures, buying accuracy at the cost of sharply higher computational complexity and energy use. On raw data from the two iSCAPE roadside stations in Guildford, the classical MLP and LSTM failed to converge to stable representations, with NMSE of 1.236 and 1.299 and correlations near 0.30; the depth-4 quantum network reached a correlation of 0.599 on the same data, about double the LSTM.

Model scale extends to 172 logical qubits and delivers the study's best forecasting fidelity on wavelet-decomposed, high-dimensional inputs. Near-term quantum hardware is limited by noise, connectivity and decoherence, so quantum machine learning has largely stayed at small or toy problem sizes. Five-level discrete wavelet decomposition expands each original feature's input dimension six-fold; the 172-qubit, depth-4, lookback-2 configuration recorded an average correlation of 0.729 and NMSE of 0.488 over the full six-hour window, against 0.695 and 0.589 for the classical MLP and 0.703 and 0.563 for the LSTM.

The quantum layer trains by translating gate circuits into an energy minimization problem rather than differentiating directly in gate parameter space. Gate-based quantum circuits typically rely on continuous quantum evolution and parameterized-gate gradients, which become constrained by hardware coherence time as they scale. DynexQML encodes the circuit as a time-independent Hamiltonian mapped to a QUBO and derives updates by finite difference in quantum energy space; the 172-qubit depth-4 circuit inflates to more than 2,200 interconnected annealing variables through gate translation overhead, with the ground state found by neuromorphic emulation.

The results cover hourly data from two roadside monitoring stations in Guildford, UK, Sutherland Memorial Park and Stoke Park, with a forecast horizon of up to six hours and inputs including PM2.5, CO, NO2, O3, air temperature, relative humidity and atmospheric pressure. All quantum computation ran on the Dynex neuromorphic quantum annealing emulation platform, and the authors list benchmarking on physical quantum hardware as the next step.

The authors state that these results do not account for the noise models, limited connectivity or decoherence inherent to physical quantum devices, so whether the emulated performance translates into a true physical quantum advantage remains to be validated. They also note that once wavelet decomposition performs part of the feature decomposition on the classical side, the gap between the quantum model and the classical MLP narrows markedly, and that circuit depth relates non-linearly to performance, with depth 3 beating depth 4 on lookback 1 with wavelets, suggesting diminishing returns when parametrization outpaces temporal context.