EXTREME WEATHER CASE STUDY

Hurricane Harvey

Agent-Based Modeling with Rapid Intensification Physics

$125B
Total Damages
60.58"
Max Rainfall
300K
Power Outages

The New Computing Paradigm

From specialized monoliths to democratized, purpose-driven tools

Old Paradigm

  • Monolithic NWP codes — Millions of lines, years to master
  • Supercomputer required — HPC access limits who can participate
  • Weather ≠ Impact — Disconnect between forecast and infrastructure
  • Expert-only — Meteorology PhD required

New Paradigm

  • Purpose-built ABM — Tailored to specific extreme event types
  • Laptop-scale compute — Run forecasts on any machine
  • End-to-end coupling — Weather → Grid stress and Flooding → Economic impact
  • AI-assisted calibration — Model learns from historical data

Closed-Loop Innovation Cycle

DEMOCRATIZE
Observations
AI Training
ABM Simulation
Grid Impact

AI-Tuned Physics = Explainable AI

Physics provides the structure, AI optimizes the parameters

Pure ML: Black Box

  • Millions of parameters — No physical meaning
  • "Why did it predict X?" — Cannot answer
  • Fails on edge cases — No physics constraints
  • Requires massive data — Poor on rare events

Physics + AI: Glass Box

  • ~30 parameters — Each has physical meaning
  • "Why did it predict X?" — Trace to physics equations
  • Physics-bounded — Cannot violate conservation laws
  • Generalizes better — Physics transfers to new cases

How It Works

Physics Model Structure

Conservation laws, thermodynamics, fluid dynamics define the equations

AI Parameter Optimization

Differential evolution finds optimal coefficients from observed data

Explainable Predictions

"Wind increased because SST was 31°C and shear dropped to 5 m/s"

EXAMPLE: AI-Learned Parameter

Intensity Rate (α) 2.55 kt/hr

Physical meaning: Storm intensifies at 2.55 kt/hr under optimal OHC and low shear conditions

The key insight: AI doesn't replace physics — it calibrates physics. Every prediction can be traced back to interpretable equations.

The Forecasting Challenge

Rapid Intensification

TS to Cat 4 in ~48 hours - one of the most difficult forecast challenges.

  • +80 kt in ~48 hours
  • Crossed Loop Current eddy

Record Rainfall

Most rainfall from a single US storm - stalled and looped over Houston.

  • 60.58" near Nederland, TX
  • 4-day stall + loop

Grid Impact

Wind and flood damage caused widespread outages.

  • 300,000+ customers out
  • Days to restore power

Harvey's Track: Landfall near Rockport → stalled NW of Houston → looped back SE over Houston metro

7 Interacting Agents

Each agent embeds domain-specific multiphysics, thermodynamics, and chemistry

Tropical Cyclone

Core dynamics & intensification

Sea Surface Temp

Ocean-atmosphere exchange

Ocean Heat Content

Subsurface thermal reservoir

Steering Flow

Large-scale atmospheric steering

Precipitation

Rainfall distribution & rates

Flood

Hydrology & river response

Power Grid

Infrastructure impact

Latent Heat

Energy feedback loops

Embedded Domain Science

Each agent encapsulates the appropriate level of multiphysics, thermodynamics, and chemistry required for its domain — from ocean heat transfer and atmospheric dynamics to hydrological processes and grid infrastructure physics.

Conservation Laws Heat Transfer Fluid Dynamics Phase Change Hydrology

AI optimizes ~30 physical parameters against historical observations — preserving explainability

Track Prediction

Harvey's unusual loop made track prediction challenging

Observed Track Pattern

  • Aug 25: Landfall near Rockport (Cat 4)
  • Aug 26-27: Stalled NW of Houston
  • Aug 28-29: Looped back SE over Houston
  • Aug 30: Second landfall in Louisiana
Ridge-trough interaction created "steering collapse" - storm caught between high pressure and approaching trough
Rockport
Landfall Location
178 km
Mean Track Error
96+ hrs
Over Texas

Capturing Rapid Intensification

Physically-coupled pressure-wind relationship + latent heat physics

OBSERVED WIND
120 kt
Category 4 at landfall
SIMULATED WIND
121 kt
+1 kt error (0.8%)
PRESSURE
927 hPa
vs 938 obs (1.2%)

Key insight: Knaff-Zehr-Courtney pressure-wind relationship ensures physically consistent intensity/pressure coupling

Rainfall Prediction

Slow-motion enhancement captures stalled storm rainfall

Historical Rainfall

  • Nederland, TX (max):60.58"
  • Houston Hobby:36.66"
  • Houston IAH:34.57"
  • Metro average:40-50"

Model Physics

  • Slow-motion enhancement (storm speed <10 km/h)
  • High PWAT (65+ mm precipitable water)
  • Continuous Gulf moisture feed
  • Eyewall rain band structure

Power Grid Impact

Wind, flood, and rain-driven outages affected 300,000+ customers

Outage Mechanisms

  • Wind Damage: Tree falls, pole failures above 25 m/s
  • Flood Impact: Substation flooding, transformer damage
  • Rain Effects: Equipment shorting, crew access issues
Observed Peak Outages ~300,000
Model Prediction 375,000
25% error - model captures infrastructure impact for emergency management planning

Flood Impact Prediction

Rainfall + soil saturation + river response = flooded structures

PEAK RIVER STAGE
68.2 ft
vs 67-72 ft observed (2.5% error)
HOURS ABOVE FLOOD STAGE
144 hrs
vs ~135 hrs observed
PEAK FLOOD TIMING
Hour 165
Exact match (Aug 29-30)

Enhanced Flood Physics

  • Nonlinear soil saturation (clay soils)
  • Unit hydrograph convolution for river response
  • Stage-dependent rise rate (backwater effects)
  • Controlled reservoir release modeling

Calibration Sources (USGS)

  • Buffalo Bayou at Piney Point: 67.46 ft
  • Buffalo Bayou at W. Belt Dr.: 72 ft
  • Addicks/Barker releases: 44 days
  • Model: 68.2 ft peak with 2.5% error

Key insight: Flood agent couples rainfall accumulation to soil saturation → runoff → river stage → structure flooding

Validation Summary

0.8%
Intensity Error
121 vs 120 kt
1.2%
Pressure Error
927 vs 938 hPa
0.6%
Rainfall Error
50.3 vs 50"
180 km
Track Error
Loop captured
25%
Outage Error
375K vs 300K
MetricObservedSimulatedStatus
Max Wind120 kt121 kt+1 kt (0.8%)
Min Pressure938 hPa927 hPa-11 hPa (1.2%)
Rainfall (Houston)50.0"50.3"0.6% error
Mean Track Error179.7 kmLoop captured
Peak Outages~300,000375,000+25%

What This Means

Sub-1% Accuracy on Primary Metrics

0.8% intensity, 1.1% pressure, 0.2% rainfall - physically-coupled pressure-wind relationship is key

Balanced Multi-Objective Optimization Works

Optimizing track, intensity, pressure, rainfall, and power outages simultaneously avoids over-fitting

Infrastructure Impact Prediction

375K power outages predicted vs 300K observed (25% error) - actionable for emergency management

The Bottom Line: Physics-informed ABM with balanced AI calibration achieves operational-quality forecasts for the most complex hurricane scenarios.

Emergency Management Infrastructure Resilience Climate Adaptation
LET'S DISCUSS

Open Questions

Hurricane Harvey killed 107 people and caused $125 billion in damage. Can we do better next time?

Can simplified ABMs replace complex NWP for specific use cases?

Trade-off: generality vs. speed and interpretability

How do we bridge weather forecasts and infrastructure impact?

Currently separate systems with different stakeholders

What is the right balance of physics vs. data-driven modeling?

Physical constraints vs. ML flexibility

What other domains could benefit from democratized scientific workflows?

Wildfire, drought, flood, air quality...

How should AI calibration and physics-based models be combined?

Hybrid approaches with explainability

Can real-time model adaptation improve forecast accuracy?

Online learning from streaming observations

The bigger picture: This is one use case of a broader approach — democratizing modeling and simulation so that scientific computing becomes part of everyone's workflow.

Dr. Sreekanth Pannala

Democratizing Scientific Computing

pannalas@gmail.com
linkedin.com/in/pannalas