Q-NEXUS: Air Quality Forecasting Architecture

• Q-NEXUS architecture is a neural framework for high-resolution spatiotemporal air quality forecasting that integrates patch embeddings, low-rank projections, and adaptive fusion.
• It employs NanoBlocks with parallel convolutional pathways—CompactKernel, MicroConv, and FusionGate—to efficiently decode complex atmospheric dynamics under tight parameter constraints.
• Empirical evaluations show Q-NEXUS achieves R² values >0.91 for key pollutants while operating in real-time on commodity hardware, outperforming larger models.

Q-NEXUS architecture, formally introduced as a neural framework for high-resolution spatiotemporal air quality forecasting in the Delhi National Capital Region, is designed for maximal efficiency and expressivity with a compact parameter count. It leverages explicit structural priors, incorporating patch embeddings, low-rank projections, and adaptive fusion—allowing the model to decode complex atmospheric dynamics while maintaining real-time computational performance. Q-NEXUS is alternatively referred to as the NEXUS (Neural Extraction and Unified Spatiotemporal) architecture in the literature (Kumar et al., 23 Feb 2026).

1. Architectural Overview and Data Pipeline

The Q-NEXUS model processes multivariate, gridded time series data: four years of 3-hourly atmospheric observations across four spatial monitoring locations, each with three pollutant measurements (CO, NO, SO₂) and five meteorological variables, for a total feature dimensionality $D = 8$. The temporal look-back window is $T = 168$ (21 days). The core data flow is as follows:

• Input tensor $\mathbf{X} \in \mathbb{R}^{L \times T \times D}$ ($L = 4$ monitors)
• Patch embedding via 1D unfolding: $$P = \mathrm{Unfold}(\mathbf{X}; p=4, s=2) \in \mathbb{R}^{L \times T' \times (pD)}$$, with $p=4$, $s=2$, $T' = 83$
• Low-rank projection: $$H = P W_1 W_2 + b \in \mathbb{R}^{L \times T' \times d'}$$, with $W_1 \in \mathbb{R}^{32 \times 32}$, $T = 168$0, $T = 168$1
• Two stacked NanoBlocks, each performing adaptive fusion over compact, micro, and gated convolutional pathways
• Attention-weighted spatial pooling: $T = 168$2
• Fully connected prediction head: $T = 168$3 (one-step-ahead forecasts for CO, NO, SO₂ at each monitor)

2. Spatiotemporal Patch Embedding and Low-Rank Encoding

Q-NEXUS encodes temporal dependencies via patch-wise feature aggregation. Patches of length $T = 168$4 (12 hours) are extracted from each temporal segment; stride $T = 168$5 ensures overlap and temporal continuity. Each patch vector, $T = 168$6, is then projected into a low-dimensional latent via a two-stage low-rank factorization: $T = 168$7 where $T = 168$8 reduces dimensionality to $T = 168$9, and $\mathbf{X} \in \mathbb{R}^{L \times T \times D}$0 expands to $\mathbf{X} \in \mathbb{R}^{L \times T \times D}$1. This structure introduces a denoising bottleneck and regularization, empirically tuned for maximal $\mathbf{X} \in \mathbb{R}^{L \times T \times D}$2.

3. Adaptive Fusion and Parallel Convolutional Pathways

Each NanoBlock within Q-NEXUS applies three parallel transformations per embedded patch sequence:

• CompactKernel: standard 1D convolution of kernel size $\mathbf{X} \in \mathbb{R}^{L \times T \times D}$3
• MicroConv: depthwise (per-channel) convolution $\mathbf{X} \in \mathbb{R}^{L \times T \times D}$4, followed by pointwise (1x1) convolution
• FusionGate: elementwise gating, $\mathbf{X} \in \mathbb{R}^{L \times T \times D}$5

The outputs $\mathbf{X} \in \mathbb{R}^{L \times T \times D}$6 are merged by a learned data-dependent attention: $\mathbf{X} \in \mathbb{R}^{L \times T \times D}$7

$\mathbf{X} \in \mathbb{R}^{L \times T \times D}$8

with $\mathbf{X} \in \mathbb{R}^{L \times T \times D}$9 a two-layer MLP. This mechanism allows dynamic adjustment of spatiotemporal receptive fields and context-aware weighting of feature streams.

4. Loss, Optimization, and Hyperparameterization

The Q-NEXUS training objective is MSE-loss with $L = 4$0-regularization: $L = 4$1 where $L = 4$2 is batch size, $L = 4$3 (sites), $L = 4$4 (species), and $L = 4$5. The Adam optimizer is used with scheduled decay, batch size 64, ReLU Kaiming initialization, dropout 0.1 after each NanoBlock, and early stopping after 10 epochs without improvement. Total parameter count is 18,748.

5. Empirical Performance and Comparative Studies

Q-NEXUS achieves $L = 4$6 for CO, $L = 4$7 for NO, and $L = 4$8 for SO₂, outperforming SCINet (35,552 params, $L = 4$9), Autoformer (68,704 params, $$P = \mathrm{Unfold}(\mathbf{X}; p=4, s=2) \in \mathbb{R}^{L \times T' \times (pD)}$$0), and FEDformer (298,080 params, $$P = \mathrm{Unfold}(\mathbf{X}; p=4, s=2) \in \mathbb{R}^{L \times T' \times (pD)}$$1) while using fewer parameters and delivering 2–6× faster inference (0.8 ms/s) (Kumar et al., 23 Feb 2026). Ablation studies demonstrate the necessity of patch embedding ($$P = \mathrm{Unfold}(\mathbf{X}; p=4, s=2) \in \mathbb{R}^{L \times T' \times (pD)}$$2), low-rank bottleneck ($$P = \mathrm{Unfold}(\mathbf{X}; p=4, s=2) \in \mathbb{R}^{L \times T' \times (pD)}$$3), and parallel/adaptive fusion pathways ($$P = \mathrm{Unfold}(\mathbf{X}; p=4, s=2) \in \mathbb{R}^{L \times T' \times (pD)}$$4). Omission of any major submodule yields marked degradation.

Model	Parameters	Avg $$P = \mathrm{Unfold}(\mathbf{X}; p=4, s=2) \in \mathbb{R}^{L \times T' \times (pD)}$$5	Inference Time
Q-NEXUS	18,748	0.9355	0.8 ms/s
SCINet	35,552	0.7531	2.1 ms/s
Autoformer	68,704	0.8804	1.7 ms/s
FEDformer	298,080	0.8747	4.5 ms/s

6. Integration of Physical Features and Interpretability

Although no physical thresholds are explicitly encoded, Q-NEXUS internalizes key meteorological controls through learned fusion. Post hoc correlation analysis reveals negative associations between pollution and temperature/wind speed, with “cold‐calm” ($$P = \mathrm{Unfold}(\mathbf{X}; p=4, s=2) \in \mathbb{R}^{L \times T' \times (pD)}$$6 K, $$P = \mathrm{Unfold}(\mathbf{X}; p=4, s=2) \in \mathbb{R}^{L \times T' \times (pD)}$$7 m/s) regimes showing up to $$P = \mathrm{Unfold}(\mathbf{X}; p=4, s=2) \in \mathbb{R}^{L \times T' \times (pD)}$$8 higher pollutant concentrations versus “warm, windy” counterparts. These forms of feature attribution confirm that model pathways such as FusionGate and the low-rank bottleneck adaptively condition on global meteorological state.

7. Domain Deployment and Significance

Q-NEXUS demonstrates its suitability for operational, city-scale air quality forecasting. Its compact design enables real-time deployment on commodity hardware, with empirical studies confirming robust predictive skill and pronounced capacity for capturing seasonal and episodic pollution responses. The architecture provides a reference implementation for neural spatiotemporal forecasting under tight parameter and latency constraints, and evidences generalizability for atmospheric and environmental domains requiring efficient hierarchical time series encoding (Kumar et al., 23 Feb 2026).