I doubt the errors you see aren’t related to the Lammps implementation of AIREBO. The second generation REBO was not fit to reproduce surface energies. Having done similar tests myself with our research code, I know it gets (100) wrong. You could conceivably do the fitting yourself by adjusting the appropriate Fijk terms. However, I am sure that if you look at a number of DFT papers on diamond surfaces you will find a somewhat wide range of results for surfaces energies making it difficult to actually do this fitting. It is unclear, at least to me, what the correct energies should be, making fitting a moot point.
Last note: Be sure not to confuse the first generation REBO (Brenner 1990) with the second generation REBO (Brenner 2002) and AIREBO (2000). 2nd Generation REBO and AIREBO are equivalent when the LJ and torsion terms are off. The 1991 paper you cite below for your “computational” result would have most certainly been referencing first generation REBO (Brenner 1990) which is an entirely different potential, more closely related to Tersoff in form.
Hope that helps,