<28 ... Rxb2 29 Kxb2 Na4+ 30 Kc1 Nxc5 31 Nxe6 Nxe6>

click for larger view

What engine were you using, to what depth did you let it analyze, what was your setting for MPV (to display multiple "Principal Variations"), and what moves does your engine show after 29...Qa6? This is the position after 27...Rxb2:

click for larger view

I typically run analyses using several engines and setting MPV=3 or MPV=5 to let the engine display its top 3 or top 5 moves. In this case I used MPV=3 (it's somewhat faster at getting to a greater search depth) and the game continuation 27...Rxb2 28.Rxb2 Qa2# was only it's third "best" move. And it only considered this mate-in-2 as its 3rd best move because most other alternatives to 29.Rxb2 result in a mate-in-1 with either 29.Qa2# or 29.Rxb1#. This is one form of the "horizon effect". In a lost position, specially a forced mate, an engine will typically choose the continuation that delays the loss as much as possible, regardless of the number of pieces that it has to sacrifice in order to push the forced mate beyond its search horizon.

The two alternatives to allowing a forced mate-in-2 are 28.Qf8+ and 28.Qc8+, both deflecting the Rb8 from supporting the Rb2 and, in the later case, deflecting the Black queen (if 28...Qxc8) from the mate on a2. And, particularly with 28.Qf8+, White gets a lot of compensation for its queen and avoids the forced mate (yes, a mate is still apparently forced) for quite a few moves. If you left MPV=1 (the typical engine default) or set MPV=2 when you did your analysis, the engine will not <display> the game continuation of 27...Rxb2 28.Rxb2 Qa2#. It will find this line (it's trivial to find for any engine, including one that's rated as low as we are), but it simply won't display it in preference for the stronger moves in the lines starting with 28.Qf8+ or 28.Qc8+. And these two moves are not easy for a human to find, not even a super-GM like Gelfand.

For example, Stockfish 11 had no problem finding a mate-in-27 at d=42 after 01:13 minutes of calculation on my old and rather slow machine: 28.Qf8+ Rxf8 29.Rxb2 (White has R+B+P vs. Q but Black's pieces are more active and White's pieces are all tangled up) 29...Qa6 30.Nc2 Qf1+ 31.Ne1 Na4 32.Rc2 Rd8 33.Kb1 Qxf2 (33...Rxd2 34.Rxd2 Qxe1+ 35.Kc2 seemed most promising to me but Stockfish prefers either 33...Qxf2 or 33...Qe2 but it doesn't matter, White is lost) 34.Kc1 Nc5 35.Kb2 Rxd2 36.Rxd2 Qxd2+ 37.Nc2 Na4+ 38.Kb3 Nb6 (a queen down White's position is clearly hopeless but 38...Qd1 would really be rubbing it in) 39.Nb4 Qxe3+ 40.Kc2 Qxa3 41.Nd3 Nd5 42.Nc1 Nb4+ 43.Kd1 Qb2 44.g6 Qc2+ 45.Ke1 Qxc1+ 46.Ke2 hxg6 47.h4 Nd5 48.Kd3 Qc3+ 49.Ke2 Qc2+ 50.Kf3 Qd3+ 51.Kf2 Qe3+ 52.Kf1 Nf4 53.h5 Qe2+ 54.Kg1 Qg2#

click for larger view

And it was able to find a mate-in-17 (mate-in-16 at greater search depths) after 28.Qc8+ after only 00:39 secs of calculation: 28.Qc8+ Rxc8 (White also goes down a rook after 28...Qxc8 29.Rxb2 Rxb2 30.Kxb2 but presumably 28...Rxc8 leads to higher evaluations and presumably a shorter mate than 28...Qxc8) 29.Rxb2 Qa6 (same idea as after 28.Qf8+, with the Nc3 protected, Black tries to infiltrate White's position with its queen) 30.Bxc3 Qxa3+ 31.Kb1 Qxc3 32.Nc2 Qd3 (setting up a similar zugzwang as in the 28.Qf8+ line) 33.e4 fxe4 34.g6 hxg6 35.h3 Qd1+ 36.Ka2 Ra8+ (a clearly resignable position after 36...Rxc2 but apparently Stockfish finds a quicker forced mate continuation) 37.Na3 Qd5+ 38.Kb1 Qd3+ (yup, I would have played the simple 38...Rxa3 but now Black has the additional threat of winning the soon-to-be-pinned Rb2) 39.Kc1 Qxa3 40.f3 (I suppose that one move is as bad as another) 40...Qc3+ 41.Kb1 Qd3+ 42.Rc2 Qb3+ (again, I would have played 42...Rc8 winning the rook) 43.Kc1 Ra1+ 44.Kd2 Qd3#

click for larger view