[nine_k]

I think this is similar to what happened in Fukushima, where the original architects of the Daiichi plant suggested building a much higher protective wall against tsunamis, but were forced to build a lower wall because such a catastrophic tsunami was deemed too improbable. (Then it came, of course.)

[gerdesj]

That may or may not be true but as per the article, building standards like that are defined in terms of 1 in 100 or 1 in 1000 year events etc. The years bit in Japan will be defined already in their standard for designing a protection for tsunami events. Now, I don't know what 1 in n years is used but I suspect that it would be 1 in 500 or 1000 years. If a 1 in 10000 years event turns up then the inevitable happens.

So all we have to look up is the standard and the event and compare. Even then you have to design to account for what to do and what happens in the event of your wall's failure and that is probably where things went really wrong.

You could always read things like this: https://www-pub.iaea.org/MTCD/Publications/PDF/Pub1710-Repor...

[huhnmonster]

Yeah, exactly. The thing is, where do you draw the line of what is statistically so unlikely that you can deem it improbable?

If we run multiple different climate models through some sort of Monte Carlo simulation, I suppose each model would output a different probability of such an event occuring. In the case where all models predict a very low probability, it may be easy to say that it is unlikely. But what if two predict a very low and one predicts a low probability? Is this now applicable and should we build to be able to sustain such an event?

These are hard questions and I currently do not see any way to get better data for the future as the different models still do not agree in many points

[MayeulC]

In that case it could be interesting to plan for future reinforcements, and continue to run the simulations as time passes and the situation evolves.

I don't know if it's due to models and computing environments not being stable enough, but I haven't heard of such a thing, while it is the obvious thing to do: even without reinforcing a dam, you could tell when it becomes dangerous to operate it.

[huhnmonster]

> In that case it could be interesting to plan for future reinforcements, and continue to run the simulations as time passes and the situation evolves.

While I agree that the idea is good, I imagine that at a point where you are planning to extend your dam by another 20m, you already have to implement the required changes in your foundation and general structure and additionally, turbines and emergency vents have to be upsized for the extended heights. So the only thing you are now missing during build is the extra 20m of dam, as the rest has to be built anyways. This leads back to my initial point of the right amount of over-/underprovisioning..

I am not sure if I understood the second point correctly, but you can obviously tell when a dam becomes too dangerous. This can either be because the foundation has set and you now have increased structural stress, the amount of water has increased beyond the maximum allowed levels or a multitude of other reasons, but they are actively monitored.

[gumby]

> Yeah, exactly. The thing is, where do you draw the line of what is statistically so unlikely that you can deem it improbable?

Important question, because otherwise your costs are absurd and something important (like a power source) never gets built.

Some fields, like road construction, have a cost model built in and at some point are willing to either pull the plug or decide that a few unlikely (originally autocorrected to "unlucky!") fatalities are acceptable.

[layoutIfNeeded]

There are branches of probability theory that deal with these questions rigorously.

https://en.m.wikipedia.org/wiki/Large_deviations_theory https://en.m.wikipedia.org/wiki/Extreme_value_theory

[supertrope]

That's what reinsurance actuaries are for. Of course extremely rare events that have not occurred in living memory are harder to model for than life insurance tables.