Boson quantum error correction is an important means to realize quantum error correction information processing. In this paper, we consider the connection of a single-mode Gottesman-Kitaev-Preskill (GKP) code with a two-dimensional (2D) surface (surface-GKP code) on a triangular quadrilateral lattice. On the one hand, we use a Steane-type scheme with maximum likelihood estimation for surface-GKP code error correction. On the other hand, the minimum-weight perfect matching (MWPM) algorithm is used to decode surface-GKP codes. In the case where only the data GKP qubits are noisy, the threshold reaches σ ≈ 0.5 ($$\bar{p}\approx 12.3 \%$$
p
¯
≈
12.3
%
). If the measurement is also noisy, the threshold is reached σ ≈ 0.25 ($$\bar{p}\approx 10.02 \%$$
p
¯
≈
10.02
%
). More importantly, we introduce a neural network decoder. When the measurements in GKP error correction are noise-free, the threshold reaches σ ≈ 0.78 ($$\bar{p}\approx 15.12 \%$$
p
¯
≈
15.12
%
). The threshold reaches σ ≈ 0.34 ($$\bar{p}\approx 11.37 \%$$
p
¯
≈
11.37
%
) when all measurements are noisy. Through the above optimization method, multi-party quantum error correction will achieve a better guarantee effect in fault-tolerant quantum computing.