- The paper introduces NEXUS, a compact neural architecture combining patch embedding, low-rank projections, parallel convolutional pathways, adaptive fusion, and spatial attention to forecast CO, NO, and SO₂ across Delhi NCR.
- NEXUS achieves R² scores of 0.9404 for CO, 0.9140 for NO, and 0.9521 for SO₂ with only 18,748 parameters, improving average R² by 6.95% over FEDformer while reducing parameters by 93.71%.
- The model trains in about 25 minutes and performs inference in 0.8 milliseconds per sample, supporting resource-constrained deployment while highlighting cold, calm conditions and winter morning periods as high-risk pollution regimes.
Overview and Motivation
This paper presents NEXUS (Neural Extraction and Unified Spatiotemporal architecture), a compact deep learning model for forecasting carbon monoxide, nitrogen oxide, and sulfur dioxide concentrations over the Delhi National Capital Region. The work is motivated by a persistent gap in the air quality forecasting literature: most recent neural forecasting architectures optimize accuracy while treating computational efficiency as an afterthought, which limits deployment in resource-constrained operational settings where such forecasts are most needed. The authors position NEXUS as addressing three gaps simultaneously—parameter efficiency, joint spatiotemporal modeling, and interpretability—using four years (2018–2021) of reanalysis data across sixteen spatial grids.
The central quantitative claim is that NEXUS achieves R² of 0.9404 for CO, 0.9140 for NO, and 0.9521 for SO₂ using only 18,748 parameters—a 93.71% reduction relative to FEDformer (298,080 parameters)—while improving average R² by 6.95% over that baseline. This pairing of higher accuracy with dramatically fewer parameters is the paper's strongest and most consequential result.
Data and Evaluation Setup
The study domain spans approximately 60 km × 60 km of Delhi NCR. Pollutant mass mixing ratios come from the Copernicus Atmosphere Monitoring Service (CAMS) at three-hourly resolution at four corner monitoring locations; meteorological variables (precipitation, surface net solar radiation, 10 m wind components, skin temperature) come from ERA5 reanalysis at hourly resolution across sixteen locations, aggregated to three-hourly intervals and spatially aligned via inverse distance weighting. After quality control and robust median/IQR normalization, the dataset comprises 15,392 spatiotemporal samples, or 61,568 individual sequences given four locations per sample.
Each input is a sliding window of 168 timesteps (21 days at 3-hourly resolution), chosen to capture weekly periodicity and seasonal transitions; the target is the next 3-hour timestep. Temporal partitioning assigns January 2018–December 2020 to training, January–June 2021 to validation, and July–December 2021 to testing, deliberately exposing the model to unseen seasonal regimes during evaluation. This is a methodologically sound choice, though it also means test performance reflects only one half-year cycle per pollutant, dominated by the monsoon-to-winter transition.
Two caveats bear on interpretation. First, the pollutant data are CAMS reanalysis fields rather than ground-based in situ measurements, so reported skill quantifies reconstruction of assimilated model output rather than true observational forecasting. Second, the spatial network is sparse—four pollutant monitoring points—which constrains how fine-grained the "high-resolution" spatial claims can be.
Architecture
NEXUS is built from three design principles grounded in atmospheric data structure: local spatiotemporal coherence justifies patch-based compression, dominant meteorological modes justify low-rank projections, and concurrent multi-scale processes justify parallel feature extraction. The pipeline proceeds through five stages:
- Patch embedding: overlapping patches of length p=4 with stride s=2 halve the effective sequence length without discarding boundary information.
- Low-rank projection: a factored linear map H=PW1W2+b through a bottleneck of rank r=32 exploits the low-dimensional manifold of atmospheric dynamics while regularizing against overfitting.
- NanoBlock processing: two stacked blocks, each running three parallel pathways—a small-kernel Conv1D pathway (CompactKernel), depthwise-separable convolution pathway (MicroConv), and multiplicative gating pathway (FusionGate)—fused via input-conditioned softmax weights computed from global pooled statistics.
- Weighted spatial pooling: learned, time-varying attention weights aggregate across the four locations rather than simple averaging.
- Prediction head: a shallow two-layer fully connected mapping producing simultaneous forecasts for all three pollutants, exploiting inter-species correlations.
Training uses Adam with exponential decay (η0=0.001, factor 0.95 every five epochs), batch size 64, MSE loss equally weighted across species and sites, dropout 0.1, L2 weight decay 10−4, Kaiming initialization, and early stopping with patience 10. Baselines (SCINet, Autoformer, FEDformer) were tuned by grid search on identical splits, mitigating tuning asymmetry concerns.
Across all six metrics (R², RMSE, MAE, sMAPE, IoA, NSE), NEXUS outperforms all three baselines on all three pollutants. Representative figures:
| Pollutant |
Metric |
SCINet |
Autoformer |
FEDformer |
NEXUS |
| CO |
R² / RMSE |
0.7537 / 0.5527 |
0.8840 / 0.3794 |
0.8890 / 0.3711 |
0.9404 / 0.2718 |
| NO |
R² / RMSE |
0.7203 / 0.6535 |
0.8549 / 0.4706 |
0.8477 / 0.4822 |
0.9140 / 0.3624 |
| SO₂ |
R² / RMSE |
0.7852 / 0.5423 |
0.9022 / 0.3659 |
0.8875 / 0.3924 |
0.9521 / 0.2560 |
RMSE reductions relative to SCINet reach 50.8% for CO. On efficiency, NEXUS trains in roughly 25 minutes versus 90 minutes for FEDformer and infers at 0.8 ms per sample versus 4.5 ms—the difference the authors identify as decisive for real-time deployment on commodity hardware. Residual diagnostics show near-zero mean bias, near-normal Q-Q behavior with slight tail deviations, and approximately constant variance across concentration levels, including during high-pollution episodes. Scatter plots yield regression slopes near unity (0.904–0.964).
A notable strength is that the model reproduces discrete episode-level spikes during the October–December winter period, not merely smoothed seasonal envelopes—an explicit failure mode of simpler models. The implication is that NEXUS retains sufficient temporal resolution to support episode-triggered emergency protocols rather than only average-condition advisories.
Ablation Study
Systematic single-component removals attribute performance as follows: removing patch embedding costs 8.77% average R² and increases parameters (to 24,320) and training time (to 38 minutes), demonstrating that patch compression improves accuracy and efficiency jointly. Removing the parallel NanoBlock pathways costs 6.28%, the low-rank projection 5.72%, weighted pooling 4.09%, and reducing to a single NanoBlock 7.49%. Parallel pathways outperform sequential ordering by 4.2% in R². Replacing NanoBlocks with LSTM layers (128 hidden units) reduces R² by 5.8% while more than doubling parameters to 45,312; substituting self-attention reduces R² by 3.4% while expanding to 156,928 parameters. The latter result supports the paper's argument that long-range attention offers no marginal benefit when predictive signal concentrates in the recent past and local spatial neighborhood. Hyperparameters (patch size, stride, hidden dimension 128, block depth 2, rank 32) were selected by grid search on validation data.
One limitation of the ablation is that interactions between components are not decomposed—only leave-one-out configurations are reported—so attribution of the full 6.95% average gain over FEDformer to specific mechanisms remains partial.
Physical and Policy-Relevant Findings
The interpretability analyses yield several quantified relationships. Temperature correlates negatively with all species (−0.561 for CO, −0.475 for NO, −0.700 for SO₂), establishing atmospheric stability—inversion intensity governed largely by overnight minima—as the dominant meteorological control; wind speed correlations are weaker (−0.277 to −0.358), identifying horizontal dispersion as secondary. Regime stratification shows cold-and-calm conditions (T<282 K, wind <1.13 m/s) produce concentrations exceeding six times those under warm-and-windy conditions, with cold-but-windy regimes intermediate—evidence that horizontal dispersion partially compensates for suppressed vertical mixing. Diurnal analysis identifies morning peaks (03:00–06:00) tied to traffic emissions trapped under shallow boundary layers, with NO showing the largest amplitude consistent with its short lifetime. Spatially, winter gradients exceed a factor of two between northwestern industrial areas and downwind regions under weak northwesterly flow, implying that regionally averaged exposure assessments systematically misrepresent peak exposures. Monthly means exceed 0.35 kg/kg in November–December versus below 0.05 kg/kg from March through September.
These findings carry direct operational implications stated within their respective sections: anticipatory emergency protocols timed to the autumn transition, diurnally resolved rather than daily-mean exposure standards, and neighborhood-level prediction for equitable intervention targeting.
Limitations and Open Questions
Several constraints qualify the results. The evaluation relies on CAMS reanalysis rather than surface observations, so absolute skill levels may not transfer to station-based forecasting where measurement noise and missing data are material. The pollutant network covers only four locations, limiting validation of the spatial heterogeneity claims despite sixteen-grid meteorology. The forecast horizon is a single 3-hour step; multi-step rollout performance, which the introduction motivates but does not evaluate, remains untested. Testing spans one half-year period (July–December 2021), so generalization across multiple unseen annual cycles—including interannual variability in biomass burning intensity—is not established. Finally, the adaptive fusion weights are analyzed only indirectly; a direct verification that learned gating aligns with hypothesized regime-dependent behavior (spatial mixing emphasis during coherent events) would strengthen the mechanistic interpretation.
Conclusion
NEXUS demonstrates that architectural choices matched to the structure of atmospheric data—patch embedding, rank-constrained projections, parallel multi-scale pathways with input-conditioned fusion—can improve forecasting accuracy while reducing parameter count by nearly two orders of magnitude relative to transformer baselines. With R² above 0.91 for all three pollutants at 18,748 parameters and sub-millisecond inference, the model is credibly deployable on constrained hardware. The accompanying physical analyses quantify the meteorological thresholds and spatial gradients relevant to Delhi NCR policy. The principal open questions concern validation against in situ observations, multi-horizon forecasting, and robustness across additional unseen annual cycles.