Atomic-scale simulations of reactive processes have been stymied by two factors: the lack of a suitable semi-empirical force field on the one hand, and the impractically large computational burden of using ab initio molecular dynamics on the other. In this paper, we use an "on-the-fly" active learning technique to develop a non-parameterized force field that, in essence, exhibits the accuracy of density functional theory and the speed of a classical molecular dynamics simulation. We developed a force field capable of capturing the crystallization of gallium nitride (GaN) during a novel additive manufacturing process featuring the reaction of liquid Ga and gaseous nitrogen precursors to grow crystalline GaN thin films. We show that this machine learning model is capable of producing a single force field that can model solid, liquid and gas phases involved in the process. We verified our computational predictions against a range of experimental measurements relevant to each phase and against ab initio calculations, showing that this non-parametric force field produces properties with excellent accuracy as well as exhibiting a computationally tractable efficiency. The force field is capable of allowing us to simulate the solid/liquid coexistence interface and the crystallization of GaN from the melt. The development of this transferable force field opens the opportunity to simulate liquid phase epitaxial growth more accurately than before, to analyze reaction and diffusion processes, and ultimately establish a growth model of the additive manufacturing process to create gallium nitride thin films. In this archive, we included the LAMMPS compatible force field parameters of gallium nitride developed with FLARE++. Users can download these force field parameters to test and recreate similar Molecular Dynamic simulation discussed in the paper.