One Domain. The Right Questions.
Each case study takes a single real-world domain and works it from multiple directions. Six to twelve investigations per domain, each using a genuinely different method at a different fidelity. Real data, real findings—not sliders and toy models.
Can PJM Handle Virginia's Data Centers?
67 million people, 13 states. Capacity market prices went from $29 to $333/MW-day. Four questions chained in sequence. The Monte Carlo changed the investment answer by $9B. Interruptibility is at least 5x cheaper than building gas plants. Every finding validated against PJM auction results and FERC/NERC reports.
Beyond 19 GW: Where Data Centers Actually Matter
PJM can absorb 19 GW system-wide. But ask where the load concentrates and Northern Virginia breaks at 1-2 GW — a 10x local problem hidden by the system-wide number.
Can Texas Keep the Lights On?
27 million people, isolated grid, no backup from neighbors. Same question, radically different structure. The answer depends on whether you model the grid as a whole or where the load actually lands.
Where Should Texas Build EV Chargers?
Texas EV registrations doubled from 235K to 472K in 2.3 years. Seven questions about where to build, how big, and what the grid can handle — validated against 3,976 real stations, DOE, and NREL benchmarks. The optimizer achieves 88.5% coverage with 56% fewer stations. The grid constraint — 430 lines of code checking 4,151 real substations — saves $68M over ten years.
Forever Chemicals in the Groundwater
A military base contaminated its aquifer with PFAS. Seven questions about the plume, the remedy, and the cost. The screening-level answer says P&T is cheaper. The full analysis reverses the decision. This is a physics problem, not an AI problem — PDEs and Monte Carlo, not neural networks.
What Predicts How Long We Live?
Does machine learning beat domain knowledge for predicting longevity? 64,000 real people across three continents (NHANES, CHNS, HRS) and 13 investigations spanning diabetes, inflammation, biological age, cognition, wealth, and sleep. Seven questions answered: ML wins on some (diabetes, heart disease), domain wins on others, and more data doesn’t always help. The honest finding: the right method depends on the question.
What Should Decision-Makers Have Known?
Kincade, Camp Fire, Dixie, Marshall — four real wildfires spanning 6,000 to 963,000 acres. Real NASA FIRMS satellite data, USGS elevation, ASOS weather, NIFC perimeters. Eight questions that emergency managers actually face: which communities are threatened, when does fire arrive, where should resources stage? The right model shifts evacuation triggers 1–7 hours earlier across 16 communities. Over-prediction validated against all four NIFC perimeters (mean 2.8x, CV=0.33). No ML. No RL. Physics + uncertainty quantification.
What Are Insurers Missing About Wildfire?
Three real fires, 316 properties, 200 Monte Carlo draws. Zone-based classification understates risk by 117%. The model catches 100% at Kincade, 14% at Marshall.