The origin of the herringbone reconstruction on Au(111) surface has never been explained properly at the atomic level because the large periodic length (~30 nm) does not allow ab initio simulations of the system and because of the lack of highly accurate empirical force field. We trained a machine learning force field with high accuracy to explore this reconstruction. Our study shows that the lattice deformation in Au deeper layers, which allows the effective relaxation of the densified and anisotropic top layer lattice, is critical for the herringbone reconstruction. The herringbone reconstruction is energetically more favorable than the stripe reconstruction only if the slab thickness exceeds 12 atomic layers. Furthermore, we reveal the high stability of herringbone reconstruction at high temperatures and that a slight strain of about ±0.2% can induce a transition from the herringbone pattern to the stripe pattern, and both agree well with the experimental observations.