You are here

risk assessment

Nuclear accidents and risk assessments

Nuclear Monitor Issue: 

A new study published in Physics and Society analyses 174 nuclear accidents between 1946 and 2014 that resulted in loss of human life and/or more than US$50,000 of property damage (in 2013 dollars). The accidents involved nuclear energy at the production/generation, transmission, and distribution phase (nuclear power plants, uranium mines, enrichment/reprocessing/MOX plants, manufacturing plants, transportation by truck or pipeline, etc.)1

The authors − academics Spencer Wheatley, Benjamin Sovacool and Didier Sornette − state that the rate of nuclear accidents meeting their criteria decreased from the late 1970s, decreased further after Chernobyl (April 1986), and since then has been fairly stable at around 0.002 to 0.003 events per plant per year (roughly one accident per year worldwide meeting their criteria). The distribution of damage size dropped after the Three Mile Island accident (March 1979) − the median damage size became approximately 3.5 times smaller.

The worst accidents do not show any clear patterns. The authors note that "the term "dragon-king" has been introduced to refer the situation where extreme events appear that do not belong to the same distribution as their smaller siblings."

Based on their statistical calculations, the authors estimate a 50% chance of a Fukushima event (or larger) in the next 50 years, a Chernobyl event (or larger) in the next 27 years, and a Three Mile Island event (or larger) in the next 10 years. However they note that "there is tremendous estimation uncertainty associated with these estimations."

A more detailed version of the research, along with the list of 174 accidents, will be published at a later date.

Probabilistic risk assessment

Wheatley, Sovacool and Sornette question the accuracy of probabilistic risk assessment (PRA), which requires the definition of failure scenarios to which probabilities and damage values are assigned. They note that statistical/empirical analyses of nuclear accidents have "almost universally" found that PRA "dramatically underestimates the risk of accidents", and they point to research demonstrating that PRAs are "fraught with unrealistic assumptions, severely underestimating the probability of accidents".

Likewise, Princeton University physicist M.V. Ramana challenges "misleading" PRAs such as Areva's estimate for its EPR of one core-damage incident per reactor in 1.6 million years, and Westinghouse's claim that for its AP1000 reactors the core melt frequency is roughly one incident per reactor in two million years.2

Ramana writes:

"There are both empirical and theoretical reasons to doubt these numbers. A 2003 study on the future of nuclear power carried out by the Massachusetts Institute of Technology points out that "uncertainties in PRA methods and data bases make it prudent to keep actual historical risk experience in mind when making judgments about safety." What does history tell us? Globally, there have been close to 15,000 reactor-years of experience, with well-known severe accidents at five commercial power reactors − three of them in Fukushima.

"However, as Thomas Cochran of the Natural Resources Defense Council explained in his recent testimony to the US Senate, depending on how core damage is defined, there are other accidents that should be included. The actuarial frequency of severe accidents may be as high as 1 in 1,400 reactor-years. At that rate, we can expect an accident involving core damage every 1.4 years if nuclear power expands from today's 440 commercial power reactors to the 1,000-reactor scenario laid out in the MIT study. In either case, though, our experience is too limited to make any reliable predictions.

"Theoretically, the probabilistic risk assessment method suffers from a number of problems. Nancy Leveson of MIT and her collaborators have argued that the chain-of-event conception of accidents typically used for such risk assessments cannot account for the indirect, non-linear, and feedback relationships that characterize many accidents in complex systems. These risk assessments do a poor job of modeling human actions and their impact on known, let alone unknown, failure modes."

Ramana notes that conclusions about overall accident probabilities derived from PRAs are "far from dependable". He notes that before the Chernobyl accident, B.A. Semenov, the head of the International Atomic Energy Agency's safety division, said that "a serious loss-of-coolant accident is practically impossible" with Chernobyl-type reactors.

Ramana concludes:

"The lesson from the Fukushima, Chernobyl, and Three Mile Island accidents is simply that nuclear power comes with the inevitability of catastrophic accidents. While these may not be frequent in an absolute sense, there are good reasons to believe that they will be far more frequent than quantitative tools such as probabilistic risk assessments predict. Any discussion about the future of nuclear power ought to start with that realization."

The Fukushima disaster illustrated one of the weaknesses of PRAs − the difficulty of modeling common-cause failures. Fukushima illustrated another problem − PRAs do not account for complacency, corruption, slack regulation etc.

He Zuoxiu, a member of the Chinese Academy of Sciences and researcher at the CAS Institute of Theoretical Physics, wrote in a 2013 article:

"The world's 443 nuclear power plants have been running for a total of 14,767 reactor-years, during which time there have been 23 accidents involving a reactor core melting. That's one major accident every 642 reactor years. But according to the design requirements, an accident of that scale should only happen once every 20,000 reactor years. The actual incidence is 32 times higher than the theory allows.

"Some argue this criticism is unfair. After all, 17 of those 23 accidents were caused by human error − something hard to account for in calculations. But human error is impossible to eliminate, and cannot be ignored when making major policy decisions.

"Even if we set aside the accidents attributed to human error, technical failings have caused core melting once every 2,461 reactor-years. That's still more than eight times the theoretical calculation."


1. Spencer Wheatley, Benjamin Sovacool and Didier Sornette, April 2015, 'Of Disasters and Dragon Kings: A Statistical Analysis of Nuclear Power Incidents & Accidents', Physics and Society,

2. M. V. Ramana, 19 April 2011, 'Beyond our imagination: Fukushima and the problem of assessing risk', Bulletin of the Atomic Scientists,

3. He Zuoxiu, 25 Oct 2013, 'Chinese nuclear disaster 'highly probable' by 2030',