[Paper Review] NEXUS: A compact neural architecture for high-resolution spatiotemporal air quality forecasting in Delhi National Capital Region

Rampunit Kumar, Aditya Maheshwari|arXiv (Cornell University)|Feb 23, 2026

Air Quality Monitoring and Forecasting0 citations

TL;DR

NEXUS is a compact neural architecture for high-resolution spatiotemporal forecasting of CO, NO, and SO2 in Delhi NCR, delivering state-leading accuracy with substantially fewer parameters.

ABSTRACT

Urban air pollution in megacities poses critical public health challenges, particularly in Delhi National Capital Region (NCR) where severe degradation affects millions. We present NEXUS (Neural Extraction and Unified Spatiotemporal) architecture for forecasting carbon monoxide, nitrogen oxide, and sulfur dioxide. Working with four years (2018--2021) of atmospheric data across sixteen spatial grids, NEXUS achieves R$^2$ exceeding 0.94 for CO, 0.91 for NO, and 0.95 for SO$_2$ using merely 18,748 parameters -- substantially fewer than SCINet (35,552), Autoformer (68,704), and FEDformer (298,080). The architecture integrates patch embedding, low-rank projections, and adaptive fusion mechanisms to decode complex atmospheric chemistry patterns. Our investigation uncovers distinct diurnal rhythms and pronounced seasonal variations, with winter months experiencing severe pollution episodes driven by temperature inversions and agricultural biomass burning. Analysis identifies critical meteorological thresholds, quantifies wind field impacts on pollutant dispersion, and maps spatial heterogeneity across the region. Extensive ablation experiments demonstrate each architectural component's role. NEXUS delivers superior predictive performance with remarkable computational efficiency, enabling real-time deployment for air quality monitoring systems.

Motivation & Objective

• Motivate accurate, real-time air quality forecasting in a dense megacity and address gaps in efficiency, joint spatiotemporal modeling, and interpretability.
• Develop a compact neural architecture that jointly forecasts CO, NO, and SO2 while leveraging meteorology and inter-pollutant correlations.
• Demonstrate parameter efficiency and real-time inference on commodity hardware in a challenging urban setting.

Proposed method

• Patch embedding to reduce temporal sequence length.
• Low-rank projection to learn a compact latent representation.
• NanoBlock with parallel pathways (CompactKernel, MicroConv, FusionGate) for multi-scale spatiotemporal feature extraction.
• Adaptive fusion with input-conditioned weights to combine pathway outputs.
• Weighted spatial pooling via learned attention over monitoring sites.
• Single-head prediction head mapping pooled features to three pollutants.

Experimental results

Research questions

• RQ1 Can a compact neural architecture jointly forecast multiple pollutants with high accuracy in a dense urban setting?
• RQ2 How much can parameter count be reduced without sacrificing predictive performance for air quality forecasts?
• RQ3 What is the impact of patch-based embedding, low-rank projections, and parallel NanoBlock pathways on spatiotemporal modeling quality?
• RQ4 How do meteorological conditions influence pollutant dispersion and model performance in Delhi NCR?

Key findings

• NEXUS achieves R2 of 0.9404 (CO), 0.9140 (NO), and 0.9521 (SO2).
• NEXUS uses only 18,748 parameters, substantially fewer than SCINet, Autoformer, and FEDformer.
• NEXUS delivers higher average R2 (0.9355) and lower RMSE/MAE than baselines, with faster training and inference.
• Patch embedding and low-rank projection are the largest contributors to performance and efficiency.
• Two stacked NanoBlocks with adaptive fusion improve robustness over single-pathway or sequential designs.
• Inference speed is 0.8 ms per sample on GPU, six times faster than FEDformer.