Can anyone explain this exchange sacrifice?

If black captures at e4 with his bishop, and white recaptures with his knight, the black knight at d3 is hanging, and white also menaces Nd6, menacing royal fork at f7, capturing the rook at e8, as well as the pawn at b7.

Not to lose his knight, black plays his knight to f4, and then Nd6 is strong.

May need longer calculation.

Grab an engine and see the lines.

I calculated what you said, but I am pretty sure black should be better there (I might be missing something though) as after 19 Qd1 Nd3 20. Re4 Bxe4 21. Nxe4 Nf4 22. Nd6 Rf8 23. Nf7+ Rxf7 24. Bxf7, black has exd4 followed by Nf5 or Nd5 blockading the pawn (white has a very weak isolated d-pawn and I much prefer black due to simple targets and not too badly placed pieces or pawns - black can even throw in cxb5 to create another isolated weak pawn for white to deal with - black will also have the ability to create an outside passer - so I evaluated the line you gave as slightly/better for black).

I checked with Stockfish 17 and after 10-15 moves, white had a significant advantage (+3 territory), though as a human, I still couldn't see a way to utilise such an advantage or find a winning continuation for white.

I guess much later down the line, there would be crushing pressure for white - I guess, as Stockfish would have calculated every possible arising position tens of moves after the exchange sac, it would have evaluated every arising position as winning for white. It's a pretty wild exchange sacrifice - Re3 looks so much more playable even after the complications from Nxf2 (20. Re3 Nxf2).

In human terms: in the resulting position white's minor pieces are all amazing (the d6 knight, the bishop pair), while black's rooks are pathetic. The game won't last until the endgame where rooks would be good.

I guess taking the rook is too dangerous for black since it allows white to get a monster knight on d6.