TERRABYTEAnalytics
PROBABILISTIC RISK ANALYSIS PLATFORM

Probabilistic Physical-Risk Analysis for Complex Engineered Systems

Identify and quantify the system states that create risk—without exhaustively simulating every possible contingency.

See How It Works
Focus on the states that matter
Certified bounds, not sampling estimates
Make risk-informed decisions with confidence

Transmission Network

Substation Line Generator

Probabilistic Uncertainty

Risk of Overload

(Annual Risk)

L-101L-205L-310 L-412L-503 10⁻⁶10⁻⁴10⁻³10⁻² Annual Risk

Annual Risk

(Per Element)

2.3×10⁻² Expected Annual Risk L-101 L-205 L-310 L-412 Other

Consequence Distribution

(System Impact)

10⁰10⁻²10⁻⁴10⁻⁶ 10⁰10²10⁴10⁶ . Impact Impact

Risk Heat Map

(Probability of Overload)

High Low

THE PROBLEM

Two hard problems. One costly outcome.

1

Repeated Risk Analysis Under Changing Conditions

Weather, loads, outages, topology, and sensor data constantly shift—driving endless re-analysis.

Weather
Loads
Outages
Topology
Sensor Observations
2

High-Fidelity Simulation Under Combinatorial Uncertainty

AC power flow, outages, thermal behavior, CFD, and large scenario spaces create a combinatorial explosion.

AC Power Flow
OPF
Thermal Simulation
CFD
Large Scenario Spaces

HOW IT WORKS

From uncertainty to actionable insight in three steps.

1

Organize Uncertainty

Model uncertainty across drivers, correlate scenarios, and reduce the space to what drives scenario representation.

2

Select High-Value Simulations

Use intelligent filtering to choose the minimal set of high-fidelity simulations that maximizes information.

3

Quantify Risk Efficiently

Aggregate simulation results to produce probabilistic risk metrics with quantified confidence.

INDUSTRIES WE SERVE

Built for the most complex, consequence-critical systems.

Electric Transmission Systems

Helping utilities and operators make better risk-informed decisions for a reliable grid.

  • Investor-Owned Utilities
  • Transmission Operators
  • ISO / RTOs
  • Planners & Engineers
  • Reliability Teams

Data-Center Power & Cooling Resilience

Rank the competing explanations against the measurements you already have, say what the record can and cannot identify, and name the one measurement worth adding.

  • Sensor error vs equipment degradation vs control behavior
  • Which sensor to add next, ranked before purchase
  • Is the as-built record wrong?
  • Outage, weather and capacity scenarios

WHAT WE MEASURE (PROOF)

What sampling misses, measured.

5,028/5,028
Critical states captured. Matched-budget Monte Carlo missed 4,183 of them.
probabilistic_contingency_risk
2.4×
Coverage of the critical set at matched evaluation budget, versus the best of 12 baselines, on the congested ×4 system.
contingency_risk_at_scale
27%
Tighter bound than the Monte Carlo confidence interval at equal budget — certified, not sampled.
probabilistic_contingency_risk
100,000
Monte Carlo draws still reporting zero probability where our method returns a bounded non-zero interval.
probabilistic_contingency_risk
12
Independent baselines benchmarked — importance sampling, subset simulation, screening and hybrids.
contingency_risk_at_scale

Every number above is computed from committed benchmark results in our study repository by website/metrics/generate_metrics.py; the study named on each card is the file it comes from. Reference systems are RTS-GMLC and congested replicas of it.

The coverage advantage is largest on congested systems with many probable critical states, and modest on an uncongested base case (0.96 against 0.92 for the best baseline on RTS-GMLC itself). Our method buys completeness and a certified bound, not raw speed: in these comparisons it spends more wall-clock than Monte Carlo to do it.

DATA CENTERS: DIAGNOSIS, MEASURED

160×
Gap between the best and worst of 16 candidate sensors, ranked by expected information gain before any is installed. The top pick removes 743 of 944 milli-nats of diagnostic doubt.
active_sensing
98%
Residual drop when a mis-recorded cooling-unit-to-row link is corrected (evidence ratio 1012). A second suspected error is declared unresolvable, and a discriminating test is recommended instead of a guess.
topology_detective
184×
How far a converged, ±4 % estimate moved across four physically reasonable model formulations, in units of its own standard deviation. Sixteen days of 1 Hz data from a real server room with unmetered cooling.
unmetered_cooling

Same repository, same rule: each card names the committed study it is read from. The server-room record is a public deposit (Melgaard et al. 2026, doi:10.5281/zenodo.19398358); the other two are synthetic facilities built to be decisive, and their reports say so.

Read the full benchmarks — including where our method loses

WHY TERRABYTE

Risk insight without brute-force.

Terrabyte Analytics builds probabilistic physical-risk analysis for networked engineered systems, combining system physics, scenario intelligence, and selective high-fidelity simulation. Monte Carlo tells you what it happened to draw. We tell you what you cannot afford to miss, and prove the bound.

  • Physics-respecting and grid-aware
  • Probabilistic, transparent, and auditable
  • Built for scale and operational use
  • Actionable insights for better decisions

Get in touch.

Questions about the method, the benchmarks, or whether any of it applies to your system.

ricardo.marquez@terrabyteanalytics.com