I don’t know that human calculation is so fundamentally different than computer evaluation, it’s just that it’s much, much slower. The human GM’s eval of any given position, without thinking ahead at all, is probably better than most engines. The problem is that strong evaluation isn’t worth much compared to a weaker evaluation that is nevertheless several candidate moves wider and many plies deeper (you might argue AlphaZero disproves this, but Stockfish was regularly beating the neural network engines even before its own neural network, for example).

Humans see fewer candidate moves. They regularly miss quiet moves that are the strongest. They often calculate to a certain depth (even a very shallow one) and stop based on gut instinct. But it’s still just trees of moves with some eval function.

I don’t think it’s fundamentally that hard to mimic, and it would actually be genuinely interesting for didactic purposes. But for fairly obvious reasons it’s not a priority outside of a couple of projects.