Earthflow is a Physics AI™ geospatial analytics platform purpose-built for solar energy site assessment. It ingests data from 70+ authoritative sources, processes it through 19 parallel Physics AI™ modules, and produces 1,180+ environmental, financial, and engineering fields plus 200+ computed KPIs per site — all in under 1 minute. This document covers both the technical architecture (Chapters 1–2) and the Physics AI methodology — how Earthflow injects established physical equations into its AI layer (Chapter 3) — making every output defensible and auditable.
The platform spans 7 architectural layers — from external data ingestion through security, storage, the Physics AI™ Engine, application services, AI/agent intelligence, to a rich interactive frontend reachable via web, Slack, Teams, Email, Telegram, and Signal — all running on Google Cloud Platform with Firebase services.
The Earthflow platform is organized into seven functional layers. Data flows upward from external sources through security, storage, the Physics AI™ Engine, application services, and the AI/agent layer before reaching the user interface — available through web, Slack, Teams, Email, Telegram, Signal, SMS, and REST/webhook APIs. Each layer is independently scalable and communicates via well-defined APIs.
This diagram traces the path of a site analysis request from user interaction through the entire platform — showing how data sources connect to analysis modules, how results flow through services, and how the AI layer powers the frontend experience.
Earthflow is branded as the first Physics-Informed Agentic AI for solar energy. Chapters 1 and 2 show how the platform is built. This chapter explains why its intelligence is different — how established physical equations are injected into Earthflow's AI layer, why that makes every output defensible and auditable, and how Earthflow is not merely applying textbook physics but advancing it with novel hybrid approaches and modern deep learning inputs.
Most AI systems learn patterns from data. Physics AI™ starts from equations that already encode centuries of validated science, and uses AI only where equations alone cannot answer the question — typically to supply inputs (from imagery, satellite data, or sensors) or to recognize objects (from pixels). The result is a system that is both explainable — every number traces to a formula and a real dataset — and scalable, because AI extracts those inputs at planetary scale.
Earthflow’s methodology connects with two established research patterns in the broader Physics AI landscape:
Earthflow sits deliberately between two poles — defensible enough for reinsurance audit, practical enough to operate at portfolio scale.
Earthflow uses four distinct patterns to bind AI to physics. Each module in the platform is one or a combination of these modes.
The formula is the boss; AI feeds its inputs from real observations.
The deterministic equation remains the system of record. AI/ML supplies one or more of its inputs at scale from imagery, reanalysis, or sensor networks — inputs that classical workflows would otherwise estimate from sparse field surveys or regional defaults.
ML proposes; equations dispose.
The Earthflow Detect family of deep-learning models identifies candidate objects from satellite and aerial imagery — solar panels, substations, inverters, transmission infrastructure. Each candidate is then validated by a physical or geometric constraint before it is accepted as truth.
Net pattern: ML expands the search space; physics shrinks it back to the truth.
Agency without hallucination.
Cirra (Claude Sonnet 4.6) reasons in natural language but cannot fabricate values. Its 25+ tools each encode either an authoritative dataset call or a domain formula. Every output is traceable to a real source or a published equation. If a question has no tool that can answer it, the agent says so — it does not invent.
This is the agentic half of Physics-Informed Agentic AI: autonomy whose action space is physically grounded by construction.
No reasoned guesses anywhere in the pipeline.
Earthflow’s foundational rule: every constant, threshold, and regional default must come from a citeable authoritative dataset (NOAA, USGS, SSURGO, NIFC, LANDFIRE, NLDN, SMAP, JRC, FEMA, NREL, SPC). When a primary source is unavailable, the platform falls back to a physics-equivalent source — never to a reasoned approximation. This is the data-side discipline that pairs with the equation-side discipline of Modes 1–3.
Earthflow operates 19 parallel Physics AI™ modules. Each one pairs an established physical equation with modern AI/ML inputs — and in nearly every case extends the classical formulation beyond textbook practice. The 19 modules below cover the platform’s foundational environmental, structural, and meteorological perils — including hail kinetic-energy & module-fragility (FM Global PRG 18-1), ASCE 7-22 wind & tornado design loads, HAZUS-MH seismic, IEC 62305 lightning, foundation classification, road access, and permitting-layer coverage; additional modules (financial, regulatory, grid-interconnection) are documented in dedicated methodology reports. The italicized phrases describe Earthflow’s adaptations: the novel ways the equation is computed, refreshed, or refined.
| # | Module | Equation & Earthflow Adaptation | AI / ML / Deep Learning | Mode | Why Earthflow’s Hybrid Wins |
|---|---|---|---|---|---|
| 1 | Erosion | RUSLE 5-factor (Wischmeier & Smith, 1978; Renard et al., 1997) — adapted: C-factor refreshed at 10-m Sentinel-2 cadence rather than textbook 30-m static cover tables; season-aware, not annual mean; LS-factor from sub-arcsecond DEM derivatives. | Sentinel-2 NDVI → C-factor via deep-learning vegetation segmentation; SSURGO multi-fallback access for K-factor. | 1 | Live, land-cover-grounded C-factor; auditable formula, season-aware, asset-specific. |
| 2 | Precipitation | Wischmeier R-factor + NOAA Atlas 14 IDF curves + Modified Fournier Index — adapted: gridded multi-source reconciliation (Daymet 1-km, GridMET 4-km, NEXRAD radar archive, Open-Meteo) replaces nearest-gauge interpolation; site-level R-factor and IDF recomputed from observed daily series. | NEXRAD dual-pol QPE with ML-calibrated estimators; PINN-constrained deep-learning downscaling to sub-km, enforcing mass-conservation of precipitation totals. | 1 + 4 | Site-level rainfall statistics from observed data, not regional default curves. |
| 3 | Surface fire | KBDI (Keetch & Byram, 1968) + ERC (NWCG) — driven daily by live soil moisture and observed fuel state, not seasonal climatology; reconciled against a 5-year rolling ignition history. | LANDFIRE FBFM40 + SMAP soil moisture + MODIS / VIIRS active-fire (deep-learning satellite analysis). | 1 | Drought physics + observed fuel + recent-ignition feedback at site scale. |
| 4 | Peat / sub-surface fire | NIFC subsurface ignition probability — extended below ground via a coupled peat-depth + moisture-decay model; subsurface fire risk is rarely computed in industry tools. | Peat-depth raster + SMAP moisture; deep-learning SAR-based peat-boundary refinement (pipeline). | 1 | Subsurface risk with physics-bounded probability — no extrapolation beyond data. |
| 5 | Flood | HAND (Rennó, 2008) + Manning’s equation — reconciled across three independent layers: hydraulic theory, observed inundation, and legal flood zones. | JRC Global Surface Water (deep-learning-segmented Landsat) + MERIT Hydro (DL-corrected drainage network) + FEMA NFHL. | 1 + 4 | Hydraulic theory anchored to both measured inundation and regulated zones. |
| 6 | Seismic | USGS NSHM PSHA (Cornell, 1968) — refined per-asset using site-class derived from local soil rasters, beyond NSHM’s default coarse site grid. | Site-class extraction from Vs30 / SSURGO via supervised classification. | 1 | Site-specific PSHA without manual geotechnical workflow. |
| 7 | Wind exposure | IEC 61400-1 design wind speed — computed at the asset from blended 11-km reanalysis and 2-km wind-atlas downscaling, with extreme-value statistics extended via deep learning. | ERA5-Land + NREL Wind Toolkit; PINN-style deep-learning downscaling with boundary-layer physics constraints for extreme-value tails. | 1 | Asset-specific design wind from observed climate, not regional defaults. |
| 8 | Hail | Significant Hail Parameter (Cohen et al., 2006) + NOAA Storm Events historical frequency — adapted: county-resolution storm climatology spatially smoothed to site using physiographic priors; convective-environment indices integrated for forward risk. | ML spatial interpolation of historical events; deep-learning radar signatures (NEXRAD dual-pol) for hail discrimination. | 1 + 2 | Reinsurance-grade hail exposure at asset granularity, not county averages. |
| 9 | Tornado | Significant Tornado Parameter (Thompson et al., 2003) + EF-scale + SPC tornado climatology — adapted: county-level SPC catalog mapped to site with damage-path-width weighting and recurrence-interval Bayesian updating. | ML smoothing of sparse tornado tracks; deep-learning radar for tornado debris signature detection (research-stage). | 1 + 4 | Asset-specific recurrence interval beyond county-level statistics. |
| 10 | Lightning | Cloud-to-ground flash density (NLDN; Cummins et al., 1998) + IEEE 1410 structural protection criteria — adapted: site-specific strike density integrated with structure height and grounding for arrester sizing. | ML temporal-spatial smoothing of NLDN data; deep-learning correlation with convective indices. | 1 | Asset-protection-grade flash density without manual NLDN queries. |
| 11 | Soil moisture | SMAP tau-omega radiometer retrieval (O’Neill et al., 2021); exponential-filter root-zone derivation — adapted: SMAP 9-km enhanced product deep-learning-downscaled to ∼1-km using Sentinel-1 SAR ancillary data. | PINN-based deep-learning SAR × microwave fusion with water-balance constraints; data-assimilation networks for root-zone estimation. | 1 | Asset-scale soil moisture, not 9-km satellite footprint. |
| 12 | Vegetation health | NDVI (Rouse, 1974), EVI (Huete, 2002), NDWI, NBR burn-ratio — adapted: phenology-aware multi-index time-series anomaly detection at Sentinel-2 10-m resolution, with Landsat fusion for 1972–present continuity. | Deep-learning vegetation segmentation and land-cover classification; transition / encroachment detection via temporal CNNs. | 2 | Vegetation state from imagery, classified by trained networks, anchored to indices. |
| 13 | Slope stability | SHALSTAB infinite-slope Factor of Safety (Montgomery & Dietrich, 1994) — adapted: site-level FoS coupling DEM-derived slope, SSURGO cohesion / unit weight, and live SMAP moisture for rainfall-triggered failure. | Deep-learning DEM downscaling for sub-meter slope; ML landslide-susceptibility refinement. | 1 | Physics-based stability with live moisture — beyond static hazard maps. |
| 14 | Solar irradiance | Bird & Hulstrom clear-sky model (1981) + Perez transposition + NREL NSRDB GHI/DNI/DHI — adapted: spectral-mismatch correction by module technology (c-Si, CdTe, bifacial); back-side irradiance modeling for bifacial systems. | Satellite irradiance with ML cloud masking; deep-learning short-horizon irradiance nowcast (roadmap). | 1 | Module-aware bankable irradiance, not generic GHI lookup. |
Three properties fall out of this discipline — properties that purely-ML systems and purely-physics simulators cannot deliver together.
Reinsurance underwriting, regulator review, EPC engineering sign-off — all require that a number can be traced to its source. Pure-ML scores cannot pass this test. Earthflow’s outputs can: every field traces back to either a published equation, an authoritative dataset, or both.
A model that learned patterns from California solar farms may behave unpredictably on a Texas peatland. Equation-orchestrated outputs do not have this problem — RUSLE works the same in Bakersfield and in Beaumont. The AI inputs adapt to local conditions; the physics does not drift.
“Agentic AI” usually means letting an LLM choose actions freely. Earthflow’s agentic layer is constrained by construction: Cirra can only act through tools whose outputs are physically grounded. The result is autonomy you can trust to underwrite a reinsurance contract or schedule a $20M construction milestone.