In order to unravel the aqueous behavior at the solid liquid interface, we employ ab initio molecular dynamics simulations. To minimize finite-size artifacts of the simulation cell, but retain the chemical accuracy of ab initio simulations, we use on-the-fly machine-learned force fields as employed by the Vienna Ab Initio Simulation Package. We investigate two model systems, first we study the hydrophobic gap on graphene stacks with up to 4 layers, and second the dissolution behavior of different ionic salts. The salts are NaI, AgF, and NaCl, which all crystallize in the same NaCl structure, but have decreasing molar solubilities in water in the listed order. We first use static density functional theory calculations to relax the individual phases separately, which is then followed by training the force fields. We compare different functionals, PBE and RPBE, both with and without Grimme D3 corrections, R2SCAN+rVV10, and vdW-DF-cx, to find the optimal compromise in the description of the two phases. The lattice parameters of graphite are best reproduced by the PBE-D3 functional, while the other functionals with dispersion treatment perform reasonably well. The PBE and RPBE functionals without Grimme corrections on the other hand lead to dramatically overestimated c axes. The lattice parameter of the ionic salts is almost exactly reproduced by the vdW-DF-cx functional, but at more than 150-fold the computational expense compared to the other functionals, while the PBE and RPBE functionals overestimate the lattice parameter slightly. The PBE-D3, RPBE-D3 and R2SCAN+rVV10 functionals on the other hand slightly underestimate the lattice parameter. In bulk water, the bond lengths and angles are overestimated by roughly 3% consistently across all functionals, only the R2SCAN+rVV10 functional performs slightly better at 2% overestimation. Each phase is first trained separately by ab-initio simulations, until the Bayesian error threshold of the forces is undercut consistently. Thus, the interactions within the phase are learned in a smaller simulation cell, saving computation time. After training the individual phases, the solid phase is joined with the liquid phase and trained together to include the interface interactions into the force field. This training regime allows efficient training of the force field that can then be applied to much larger simulation cells than the training data.