You can imagine playing chess as a tree. You start with a game state, and every possible move is an edge to a new game state. Unfortunately, the amount of game states grows approximately exponentially (e.g. for every game state there are approximately 40 moves to play, thus every extra move multiplies #(game states) by 40, approximately). This neural engine is a trick so that Stockfish does not have to simulate all 40 moves every game states. The neural net outputs the moves that are most likely to be strong moves, and Stockfish will only consider those moves. Of course, this is a very basic explanation and there are many more optimisations Stockfish uses, though the ultimate goal of almost every optimisation is to reduce the amount of simulation Stockfish has to do.