Forecasting PM10 particulate matter is a two-sided challenge: on one hand, you need precise point estimates at monitoring stations; on the other, a continuous spatial map is essential to track the evolution of dust plumes, especially during storms that sweep across large areas. Chemical transport models (CTMs) like those from the Copernicus Atmosphere Monitoring Service (CAMS) deliver gridded fields but accumulate local biases, while graph neural networks (GNNs) track station-level performance well over short horizons but cannot produce spatial outputs. OmniPMNet, a fusion model based on Convolutional Conditional Neural Processes (ConvCNP), reconciles both approaches within a single shared spatial representation.

At its core, a terrain-aware Gaussian set convolution lifts irregular GNN station forecasts onto a regular grid, where a multi-scale Spatial Source Attention (SSA) module blends them with CAMS outputs. An omni-query readout then decodes this representation into consistent predictions – whether at station points or grid cells – over a 108-hour horizon. Evaluated on 1,618 monitoring stations across China throughout 2024, OmniPMNet matches the best GNN's station-level accuracy (mean absolute error 21.14 versus 22.00 µg/m³) and cuts CAMS error by 30%, while simultaneously delivering the spatial maps that GNNs cannot. Gains are sharpest in the high-concentration tail, where the 90th-percentile MAE drops by 9% relative to the GNN and 25% relative to CAMS, and during dust episodes, where it improves categorical detection while tracking evolving spatial trajectories.

Why does any of this matter for those designing on-premise AI stacks? The answer lies in the architecture itself. ConvCNPs belong to the neural process family, relatively lightweight models that for inference do not need enterprise GPUs: they can run on consumer hardware or even CPUs. That means an entire PM10 forecasting system can be hosted entirely on local servers, without sending data to external cloud platforms. This is far from a trivial detail when dealing with environmental data, often considered sensitive for public policy and subject to regulations like GDPR in Europe.

Secondly, the omni-query capability – querying the model for a specific station or an entire map – closely mirrors the evolution of Large Language Models that handle diverse prompts without changing architecture. OmniPMNet shows this paradigm also applies to environmental monitoring, eliminating the need to maintain separate systems for point and spatial forecasts. A regional agency could thus manage a single on-premise model, train it on local data (without sharing it with third parties), and serve real-time forecasts to decision-makers and citizens, with full sovereignty over the infrastructure.

Then there is a second-order effect: models like OmniPMNet reduce reliance on centralized cloud services even for environmental forecasting, a market until now dominated by vendors that lock data into proprietary platforms. For developing nations or public bodies with tight IT budgets, the ability to deploy an on-premise system on modest hardware with open-source tooling could offer a concrete alternative to vendor lock-in. The benefits are not just economic: keeping data within national borders avoids unauthorized access risks and aligns operations with increasingly strict data residency requirements.

The experiments described rely on Chinese stations and CAMS data, but the approach is generalizable. The structural point is that AI for environmental nowcasting is moving toward infrastructural autonomy, and architectures like ConvCNP make local deployment not just possible but cost-effective in terms of TCO. Those already running on-premise servers for LLMs will find the dynamics familiar: compact models, updatable without the cloud, capable of handling heterogeneous queries without external services. OmniPMNet, with its focus on particulate matter, scores a point for AI that doesn't need to leave home.