Pinwheel
Member
First name (team leader)
Kgomotso
Last name
Lekola
Organisation name
Pinwheel
Organisation type
Small & Medium Enterprise or Startup
Organisation location
South Africa
Model
Model name
LNBC
Number of individuals supporting model development:
1-5
Maximum number of Central Processing Units (CPUs) supporting model development or forecast production:
< 8
Maximum number of Graphics Processing Units (GPUs) supporting model development or forecast production:
< 4
How would you best classify the IT system used for model development or forecast production:
Single node system
Model summary questionnaire for model LNBC
Please note that the list below shows all questionnaires submitted for this model.
They are displayed from the most recent to the earliest, covering each 13-week competition period in which the team competed with this model.
Which of the following descriptions best represent the overarching design of your forecasting model?
- Post-processing of numerical weather prediction (NWP) data.
- Statistical model focused on generating quintile probabilities.
- An empirical model that utilises historical weather patterns.
What techniques did you use to initialise your model? (For example: data sources and processing of initial conditions)
LNBC is initialized with NWP subseasonal forecasts from public APIs (GFS and ECMWF IFS). We use a global sampling of locations and interpolate to the 1.5° competition grid. Only data with timestamps on or before the Thursday 00 UTC initialization date are used. Initialization and preprocessing steps are standard for subseasonal forecasting.
If any, what data does your model rely on for real-time forecasting purposes?
Open-Meteo GFS API (gfs_seamless) – 16-day forecasts
Open-Meteo Seasonal API – ECMWF IFS subseasonal (EC46), 45-day forecasts
Variables: 2 m temperature, total precipitation, mean sea level pressure.
What types of datasets were used for model training? (For example: observational datasets, reanalysis data, NWP outputs or satellite data)
LNBC is not a trained ML model. It is a deterministic physics engine. No training datasets are used.
Validation and tuning use:
ERA5 (via Open-Meteo) for backtests
IBTrACS (e.g. Hurricane Katrina) for extreme-event checks
Live GFS for stability tests
Please provide an overview of your final ML/AI model architecture (For example: key design features, specific algorithms or frameworks used, and any pre- or post-processing steps)
LNBC is NOT a machine learning model. It is a deterministic physics engine based on the Riccati-Gevrey dissipation physics. No training datasets are used.
Validation & Tuning Datasets:
ERA5 Reanalysis (via Open-Meteo Historical API) – Used for backtesting stability improvements
IBTrACS Hurricane Database – Used for extreme-event validation (e.g., Hurricane Katrina case study)
Live GFS Forecasts – Used for real-time stability benchmarking
Note: These datasets are used solely for validation and parameter tuning, not for training weights or learning pattern
Have you published or presented any work related to this forecasting model? If yes, could you share references or links?
No peer-reviewed publications yet. LNBC is documented in internal technical reports and code.
Before submitting your forecasts to the AI Weather Quest, did you validate your model against observational or independent datasets? If so, how?
Validation Datasets:
Live GFS Forecasts – 4 global locations, 16-day horizon, stability gain: 41.69%
Hurricane Katrina (IBTrACS) – Extreme-event case study, validated trajectory prediction
ERA5 Backtests – Historical reanalysis comparison for multi-horizon accuracy
Validation Metrics:
Stability Improvement: 41.69% reduction in forecast instability (σ)
Projected RPSS: +0.66 (based on stability-to-skill conversion)
Probability Conservation: 100% compliance (all quintiles sum to 1.0 ± 1e-6)
Did you face any challenges during model development, and how did you address them?
Main challenges were converting a deterministic physics core into probabilistic quintile forecasts and ensuring full global coverage. We addressed these with ensemble dressing and interpolation. Implementation details are proprietary.
Are there any limitations to your current model that you aim to address in future iterations?
Current Limitations:
Spatial Coverage: 54.2% LNBC-processed, 45.8% climatology fallback (due to sampling constraints)
Ensemble Size: Limited to 51 members (ECMWF IFS constraint)
Computational Cost: Real-time processing of 312 locations takes ~15 minutes
Fixed Parameters: The model is purely physics-based with no learning component. All parameters (dissipation coefficients, energy barriers, interpolation weights) are currently hand-tuned based on theoretical constraints.
Planned Improvements:
Near-Term (Operational):
Increased Sampling Density: Target 80%+ LNBC coverage with optimized API batching (implemented via LNBC_HIGH_DENSITY mode)
Multi-Model Ensembles: Integrate GFS, ECMWF, and UKMO for larger ensemble sizes (implemented via LNBC_MULTI_MODEL mode)
Adaptive Interpolation: Replace nearest-neighbor with physics-aware spatial smoothing (implemented via LNBC_ADAPTIVE_INTERP)
Real-Time Optimization: GPU acceleration for sub-minute processing (implemented via LNBC_GPU mode)
Long-Term (ML-Hybrid Architecture): 5. ML-Optimized Parameter Tuning: While the core physics engine remains deterministic, we plan to implement machine learning for:
Adaptive dissipation coefficients (α, τ, δ) based on atmospheric regime (tropical vs. polar, stable vs. chaotic)
Dynamic energy barrier thresholds (L) optimized per location and season
Learned interpolation weights for physics-aware spatial smoothing
Topological Optimization: ML-based optimization of the manifold topology:
Automatic detection of singularities (e.g., polar vortex boundaries, jet stream cores)
Learned regularization strategies for different atmospheric features
Data-driven selection of optimal NWP ensemble members for blending
Implementation Strategy: The ML components would serve as a meta-optimizer for the physics engine, not replace it. The Riccati-Gevrey Lekola Log-Difference Laws remain enforced, but their parameters become adaptive rather than fixed. This preserves physical consistency while leveraging data-driven insights.
Are there any other AI/ML model components or innovations that you wish to highlight?
Physics-Based Stability Enforcement
First subseasonal model to apply Riccati-Gevrey energy constraints
Achieves 41.69% stability improvement without ML training
Z3-verified state machine ensures operational integrity
Zero illegal state transitions (S0→S1→S2→S3→S4→S5)
Deterministic physics core + statistical ensemble dressing
Preserves physical consistency while meeting probabilistic requirements
Who contributed to the development of this model? Please list all individuals who contributed to this model, along with their specific roles (e.g., data preparation, model architecture, model validation, etc) to acknowledge individual contributions.
Kgomotso Lekola - Mathematical formulation (Riccati-Gevrey Lekola-Log Difference Barrier Law)
Submitted forecast data in previous period(s)
Please note: Submitted forecast data is only publicly available once the evaluation of a full competitive period has been completed. See the competition's full detailed schedule with submitted data publication dates for each period here.
Access forecasts data