Embedded Files
Harmonizing Water Models
Harmonizing Water Models
Common Backbone: Lynker Spatial provides a platform for indexing, propagating, improving, and validating raster- and vector-based rainfall–runoff hydrologic models against the Lynker Modeling Fabric.
Model Skill Scorecards: Standardized, repeatable scorecards for validating model performance and assessing skill across space, time, and hydrologic regimes—supporting transparent comparison, benchmarking, and continuous improvement.
High-Fidelity Spatial Visualization: Real-time rendering of discharge magnitude and inundation potential. The platform dynamically renders millions of stream segments, delivering situational awareness from CONUS-scale overviews down to individual catchments.
The Conflation Problem Solved
The Conflation Problem Solved
Flowfabric bridges the gap between disparate model outputs and local hydrography.
We ingest gridded discharge products (ECMWF GloFAS, FLASH-CREST/HP/SAC) and vector-based solutions (Lynker's LSTM, AWI NextGen, NWM, USGS observations).
By indexing these inputs to the standardized Hydrofabric data model, we unlock granular interoperability, allowing for seamless comparison between a 10km grid cell model and a hyper-local stream gauge.
Network Expansion & Downscaling
Network Expansion & Downscaling
We don't just display model outputs; we enhance them. FlowFabric propagates flow from coarse model outputs through the dense Hydrofabric network.
This allows us to generate synthetic forecasts for tributaries and stream segments that were computationally infeasible to include in the original source models, effectively "downscaling" global predictions to local relevance.
Physics-Informed Machine Learning
Physics-Informed Machine Learning
Our ML integration extends beyond post-processing. We utilize machine learning to infer critical channel dimensions and hydraulic parameters across the modeling space.
These learned geometries actively drive our propagation algorithms and bias-correction pipelines, ensuring that both flow routing and final error adjustments are grounded in a refined, high-fidelity representation of the physical network.
Multi-model & Temporal Continuity
Multi-model & Temporal Continuity
From historical reanalysis to long-range horizons.
Hindcast: High-resolution FLASH products.
Short-Term: NWM and AWI NextGen.
Medium/Long-Range: NWM and ECMWF GloFAS.
Access a continuous stream of hydrologic insight across all time scales.
Scientific Validation & Skill Assessment
Scientific Validation & Skill Assessment
We continuously compute and publish skill metrics (KGE, NRMSE, R², PBIAS, etc.) for every ingested model. By analyzing spatial error distribution (as seen in the maps below), users can identify which model performs best in their specific "Region of Interest".
Heterogeneous Formulations and Diagnostic Model Selection:
Heterogeneous Formulations and Diagnostic Model Selection:
Hydrologic predictability varies significantly across diverse landscapes; a model that excels in the snow-dominated Rockies may struggle in the humid Southeast. Our interactive map identifies the highest-performing model configuration for each hydrographic region based on your selected skill metric. By analyzing these spatial performance patterns, forecasters can determine which model source provides the most reliable guidance for their specific area of interest, ensuring decisions are driven by the best available science.
The Power of KGE Decomposition:
The Power of KGE Decomposition:
We utilize Kling-Gupta Efficiency (KGE) not just as a single score, but as a diagnostic tool. By decomposing KGE into its three components, we provide actionable insights for model developers:
Variability (Flashiness): Ratio of standard deviations (σsim / σobs).
Ideal = 1.
α > 1: Model is too 'flashy' (peaks too high, recessions too fast).
α < 1: Model is too smooth or damped.
Volume Bias: Ratio of means (μsim / μobs).
Ideal = 1.
β > 1: Overestimation of flow volume (Wet Bias).
β < 1: Underestimation of flow volume (Dry Bias).
Timing & Shape: Measures how well the forecast timing and hydrograph shape match observations.
Ideal = 1.
Values < 1 indicate phase errors (peaks arriving too early/late) or noise.
Model Addition: Lynker's own model development
Model Addition: Lynker's own model development
A custom Lynker Spatial LSTM designed to be lightweight and easily tailored to specific applications without the heavy dependencies of frameworks like NextGen.
High-Fidelity Modeling: Utilizing state-of-the-art LSTM architectures, the model captures diverse hydrologic responses—such as the distinct April/May snowmelt pulse and orographic signals. Preliminary evaluations are encouraging, showing the model generalizes robustly from training basins to unseen areas.
Operational Efficiency: Once trained, the model is exceptionally computationally cheap; a 5-year simulation across 8,300 catchments can be completed in tens of minutes on a single g5.xlarge instance (approx. $1.01/hr), making it highly scalable.
FlowFabric API
FlowFabric API
Programmatic access to the entire National Hydrologic Intelligence data lake. Secure, authenticated endpoints allow for querying vectors for specific time windows.
Seamlessly integrate national-scale hydrology into your applications. The FlowFabric API is a RESTful, interface designed for high-throughput access to both indexed model outputs and derived hydraulic products.
Performance AND scale
Performance AND scale
Whether it's 3 or 33 million features, the Data Portal and language agnostic toolings provide rapid, consistent access to the data you need in operational timespans.
NWM Analysis & Assimilation Data for 2.7 million reaches for 3 hours
1.2 seconds
GLoFAS data for for 2.7 features for 1 day
27+million values in 3.2 Seconds
NWM Analysis & Assimilation Data for 8,330,202 reaches for 18 hours
49+million values in 1.8 seconds
NWM Retrospective Dataset for 2 features for 1 year (hourly forecast)
17,520 values in 14.38 Seconds
1.2 seconds
GLoFAS data for for 2.7 features for 1 day
27+million values in 3.2 Seconds
NWM Analysis & Assimilation Data for 8,330,202 reaches for 18 hours
49+million values in 1.8 seconds
NWM Retrospective Dataset for 2 features for 1 year (hourly forecast)
17,520 values in 14.38 Seconds
Unlock the data you need when you need it.
Page updated
Report abuse