Ask HN: Genetic algorithms for detecting Single Point Of Failure

I've always been curious at the level of sophistication of the simulation tools wielded by our peers in other natural sciences like physics. It appears that they are able to really simulate a multitude of complex interactions and consequences through both mathematics and computational models. Perhaps I'm being extremely naive, but I'm yet to come across a comprehensive enough tool (or set of tools) that can model a "perturbation" (request, input etc) rippling through an entire software system that could ultimately cause some severe failure. This failure would be the result of some hidden Single Point of Failure or a result of a congruence of multiple side-effects that end up in a system-wide halt or performance degradation, and it could be modeled by initializing a sense of "state" per component of your system (ex: When my web server is servicing 5000 requests, AND my message queue has applied some backpressure and is 80% full, AND there's some heavy asynchronous processing happening in a few other machines for crunching matrices, AND one of my DB's has gone down and my Read replica is going to stand in its place temporarily, AND there's some arbitrary network jitter -> BOOM!)

I understand that architects try to think through these cases on their own, but as you add more components to your system and interface with more open source software that you may not understand as deeply as its authors, it appears that there needs to be a more scalable way to 'discover' hidden weaknesses. Could we use genetic algorithms to model each of these subsystems' "state" and throw some kind of perturbations (requests) its way to see how our system reacts? Is there much value in doing this, and if so, then why haven't we had our computers simulate computers and subsystems in the realm of software development and architecture?