Computing for Safety Analysis

Modeling what could go wrong, how likely it is, and what the consequences would be, to show the plant is safe by design.

What safety analysis asks

Safety analysis systematically asks: what can fail, how can it fail, how likely is each failure, and what happens if it does. The goal is to show that no credible sequence of events leads to unacceptable consequences, and to identify where design changes or safeguards buy the most risk reduction.

The main tools

• Failure modes and effects analysis: enumerate component failures and their effects
• Fault trees: work backward from an undesired top event to its causes
• Event trees: work forward from an initiating event through success and failure of safeguards
• Consequence modeling: physics of what the failure actually does

Why fusion is different

Fusion has favorable safety physics: the reaction stops if fuel or confinement is lost, there is no runaway chain reaction, and the fuel inventory in the plasma is tiny. The dominant hazards are the tritium inventory, activated materials, and the large stored energy in magnets and coolant. Safety analysis focuses effort where these hazards actually live rather than importing fission assumptions wholesale.

def event_sequence(p_initiate, safeguards):
    # probability that all safeguards fail after an initiating event
    p = p_initiate
    for s in safeguards:
        p *= (1 - s.reliability)
    return p

Computing's role

Computation quantifies the trees, propagates uncertainty, and models consequences (tritium transport, thermal transients, magnet energy dumps). It turns a qualitative hazard list into numbers a reviewer can check. Crucially, the analysis must be reproducible: each result traces to inputs, code, and assumptions so it can be independently re-run.

Kronos framing

This work supports licensing evidence for the Hyperion breeder ahead of construction, which begins in the second quarter of 2027. It is design-stage analysis on machines that are not yet built, and it is kept honest by separating demonstrated results from modeled projections.