Observations about the probabilistic representation of uncertainty for maximum corrosion pit depths

It is traditional in engineering, corrosion science and applied statistics to apply extreme value statistics to represent the variability in the maximum (or ‘extreme’) depth of corrosion pits and to use this to extrapolate in time and for larger surface areas. The most commonly employed extreme value distribution is Gumbel, with a history of application to pitting of aluminium and stainless steel extending back some 50 years. Indeed, pitting corrosion representation is often held as one of the prime areas of application of extreme value statistics. It seems only natural, therefore, to apply it also to represent the extreme pitting depth for structural steels. However, while this may be correct superficially, recent detailed examination of actual field data for mild steel pitting caused by marine immersion conditions casts doubt on the validity of this approach. The external data is more consistent with a bi-model probability distribution. The possible reasons for this are described, including the high likelihood that the extreme pit depths are not statistically independent events, not even asymptotically. This means that theoretically, at least, extreme value theory is not applicable. A new approach based on elementary concepts in structural reliability is proposed and some observations about the practical implications are made.