I have seen one paper on decompiling C code and other software-based reverse engineering (of sorts) been done using machine learning (or claims on doing so).
I am interested in finding out if the same idea can be used to simulate the working of an actual mechanical machine, assuming that the input and outputs are known, but the machine and it's internal conditions are considered as a "blackbox".
The idea is to assume that the internal working of this machine is constant under normal working conditions. The inputs and outputs given to the ML algorithm is obtained at normal working conditions, with a few borderline cases. If such an algorithm can be developed, it should be able to take in similar inputs (ones that cannot be given in real life or risk ruining the machine) and predict the output.
The reasoning behind such a question is due to the fact that, the number of parameters involved is in the tens of thousands, To hard code a suitable simulation software (within budget & time constraints) is not practical. The actual working of the machine is not important; we are interested only in how it performs for certain inputs.