Physics-Guided AI Architecture
How we translate 10km atmospheric coarse grids into 100m street-level operational predictions while enforcing physical laws of energy conservation.
1. UNet Super-Resolution
A deep convolutional autoencoder with skip connections that ingests coarse numerical weather predictions and projects them onto 100m high-resolution priors (NDVI, NDBI, elevation slope/aspect, and surface albedo).
2. Bidirectional ConvLSTM
Captures thermal inertia and temporal lag. Buildings absorb solar radiation during the day and re-emit it at night, creating the nocturnal Urban Heat Island (UHI) phenomenon that generic models fail to predict.
3. Physics-Informed Loss
Standard neural networks hallucinate non-physical temperatures. Our loss function penalizes violations of surface energy balance equations (Sensible Heat + Latent Heat + Ground Flux = Net Radiation).
Atmospheric Downscaling Formulation
The target 100m street temperature field T{100m}(x, y, t) is modeled as a conditioned nonlinear transformation of the low-resolution NWP state T_low, static geospatial covariates G(x, y), and dynamic satellite states S(x, y, t):
Wet Bulb Globe Temperature (WBGT) Operational Formulation:
Calculated in accordance with ISO 7243 international standards for occupational heat stress:
The Low-Cost IoT Sensor Calibration Protocol
Competitors make the mistake of attempting city-wide dense sensor networks (costing hundreds of thousands of dollars to maintain). In contrast, we use sensors strictly for calibration, not continuous coverage:
Strategically deployed across high-rise, commercial, green parks, and informal settlement typologies.
Zero external wiring or local grid power dependency. 3-year autonomous battery cycle life.
Self-calibrating algorithms detect and isolate faulty sensor telemetry without corrupting the downscaled model weights.