So, the rules were: give a table of data with 3 colums. first the temperature, then the pressure and at last the property we want to get, it could be entropy, density, viscosity and other stuff... Run the ANN ad if we get a R close to 1 (linear regression coefficient), then our equation was right.
in theory you can, but practically you just can't. the difference between the example you gave is that that first both input and output are small, and the result does not have to be 100% accurate, also the output can be reversed with a few possibilities.
for example if you are only allowed to multiply 2 numbers as an input to generate a single output
if you have
3 * 4 = 12
if you have
3 * ? = 12 > easy to solve the second input
if you have
? * ? = 12
you have a few options to try out
1* 12 = 12
2 * 6 = 12
3 * 4 =12
4* 3 = 12
----------------------------------
but doing this with a 256 bits you always have 2^256 combinations !
the strength of the hashing function is that you can't know where it came from.