Computing for Probabilistic Risk Assessment

Combining failure probabilities and consequences across many scenarios into an integrated, quantified picture of risk.

What it integrates

Probabilistic risk assessment (PRA) answers three questions together: what can go wrong, how likely is it, and how bad is it. It combines initiating-event frequencies, the reliability of safeguards, and the consequences of each outcome into a quantified risk picture, so that the many possible scenarios can be compared on a common scale and the dominant contributors identified.

The building blocks

• Initiating events and their frequencies
• Event trees: how safeguards succeed or fail after an initiator
• Fault trees: why each safeguard might fail
• Consequence models: what each end-state actually causes
• Uncertainty: distributions on every number, propagated to the result

Risk as frequency times consequence

Risk combines how often something happens with how severe it is. A frequent, mild event and a rare, severe one can carry similar risk, and PRA makes that comparison explicit rather than leaving it to intuition. The output is often a distribution over consequences with their frequencies, not a single risk number.

Uncertainty is central, not decorative

Every input to a PRA is uncertain, and those uncertainties compound. A PRA that reports point estimates without uncertainty is nearly meaningless, because the spread often dominates the story. Monte Carlo propagation of the input distributions gives a credible range and, importantly, shows which uncertain inputs drive the result, linking PRA to sensitivity analysis.

def risk(scenarios):
    # scenario: frequency (per year) and consequence measure
    return sum(s['frequency'] * s['consequence'] for s in scenarios)

Kronos framing

PRA for the Hyperion breeder focuses on the hazards that actually dominate fusion risk, tritium inventory, activated materials, and stored energy, rather than importing fission scenarios. It supports the licensing case ahead of construction in the second quarter of 2027, and it is kept reproducible so a regulator can re-run it.