Reply #134: Yes, my understanding is correct.

[eomer]

Most likely you are saying that when you're modeling an event that has already occurred then it is justifiable to fudge the model until you get it to duplicate the results of the real event. Agreed, but only up to a point and only if you're using the resulting model only in a certain way.

Ideally you should not fudge parameters that physics tells you are one thing but you need to be another just to get the model to work. If you have to do that then your model is internally wrong somehow. To take an obvious example, if one of the parameters in the model is the force of gravity then you could possibly get the model to do what you need it to do (to tilt or to collapse) by increasing the force of gravity from 1G to 1.5G. You're arguing, effectively, that that would be justifiable because, after all, it did collapse. But if you're needing to fudge the force of gravity then you really should instead go back to your model to figure out what's wrong with it. The factors that NIST fudged were not as obvious as the force of gravity but they were similar in that the values they got by modeling the physics of the components were one thing but then they had to arbitrarily change them to something else in order to get the model to do what they wanted.

That said, I do agree that in some cases fudging the model is exactly what you want to do anyway because you are only interested in getting the model to deliver behavior that resembles the behavior you see in the real world and you don't care whether the internals of the model are all individually true versus delivering a behavior that in the aggregate predicts the behavior that you see in the real world. In such a situation you might be willing to fudge the force of gravity to 1.5G because you know you're approximating some other un-modeled factor in the process and aren't concerned about the distinction between the force of gravity and the un-modeled factor.

But a model that delivers the right external results but has internal components that are only accurate in the aggregate rather than accurate individually is then a model that you can count on only in terms of the external results. You can't use such a model as proof that the internals are each individually a true representation of reality.

In other words, such a model could not tell you that columns weren't weakened by the use of explosives. For all you know, the weakening of columns by explosives (a piece of the model that is missing) is what is causing you to have to fudge other factors to get the model to tilt. For all you know, the real-life towers would never have tilted if it weren't for the explosives. If you just fudge the plane impact damage until you get the tower to tilt then you may well be fudging the plane impact damage until the extra amount is enough to act as a proxy for the damage that was in reality done by explosives.

I've done quite a lot of modeling (in my few decades of actuarial consulting) where we intentionally lumped several forces together because we didn't care about the distinction between them. This would have been factors such as the force of mortality, turnover, disability, salary scale, and investment earnings. We understood that the internal components were not truly represented individually but that the external results in the aggregate were all we cared about anyway (in those particular cases).

So the NIST model might be useful for some purposes even if it had to fudge certain factors beyond their reasonable range under the laws of physics. But proving that each internal factor is individually true isn't one of the things you can do with such a fudged model. In other words, you can't prove the towers collapsed just due to impact and fire. If you have to fudge the effect of impact then you may well be pushing it up because you need to account for the un-modeled force of explosives.