Classical models fail to converge on raw PM2.5 data: a quantum neural network reaches 0.60 correlation, about double the LSTM

A quantum neural network still learned PM2.5 dynamics from raw, undecomposed environmental time series, while classical models failed to converge on stable representations from the same data. Earlier practice decomposed the signal or engineered features before handing it to a classical network; here the raw time series goes straight into the model, compared against a multilayer perceptron and an LSTM on identical inputs. On raw data the depth-4 quantum neural network reached a Pearson correlation of 0.60, against 0.29 for the LSTM and 0.30 for the MLP, and it outperformed both classical baselines across all lookback periods.

Combining a five-level discrete wavelet transform with the quantum circuit produced the study's highest average correlation and lowest average error over the six-hour horizon. Wavelet decomposition expands each raw feature into six input dimensions and requires a proportional increase in qubit capacity, while classical models also improve markedly on wavelet features, narrowing the gap between the two approaches. The 172-qubit, depth-4 quantum neural network on wavelet data with lookback 2 recorded an average correlation of 0.73 and an NMSE of 0.488, the best reported in the study.

The quantum model's uncertainty intervals expand with pollution spikes instead of producing the overconfident predictions seen in the classical baselines. Classical baselines often gave confident but erroneous point predictions at extreme peaks, whereas the quantum model used quantile regression to derive 95% prediction intervals. On hourly data from two roadside monitoring stations in Guildford during 2019, the quantum model's 95% prediction intervals widened dynamically to encompass extreme PM2.5 peaks.

The work evaluates quantum machine learning on real urban air-quality data, forecasting PM2.5 from hourly 2019 measurements at two roadside monitoring stations in Guildford, UK, over horizons up to six hours. The authors list benchmarking on physical quantum hardware as the next step and recommend comparing the DynexQML framework against a wider array of classical models such as temporal convolutional networks and transformers.

The authors state that the quantum neural network's performance was obtained on an emulated platform and that the results do not account for the noise models, limited connectivity or decoherence inherent to physical quantum devices. They also note that with wavelet features the classical multilayer perceptron closely matched the quantum model, with the quantum advantage appearing mainly on raw, unprocessed data.