The majority of current force field architectures employ simple potentials to model the behaviour of intra- and inter-molecular interactions. Consequently, the ability to simulate systems of increasing size, for increasing timescales has been observed as access to ever improving computational resources has developed. However, there is a trade-off, in terms of a loss of accuracy, as complex electronic effects such as polarization, hydrogen bonding and short-range repulsion are often poorly described. On the other hand, fully quantum-mechanical simulations that are able to better describe these complex effects remain computationally intractable for biologically relevant molecules. A solution is required that allows for accurate simulations of large systems over significant timescales. To this end, a novel, conformationally-dependent force field (FFLUX) has been developed that uses machine learning to interpolate pre-computed quantum mechanical data. The central theme of this thesis is the development and application of FFLUX. Specifically, FFLUX machine learning models are tested for a number of chemically relevant small molecules, as well as the amino acid glycine. These models are used as inputs into a molecular simulation engine (DL_POLY), modified accordingly, and simulations are performed which show the ability of FFLUX to reproduce various quantum mechanical values with excellent accuracy. Factors affecting the quality of machine-learned models, such as the number and distribution of training points, are explored in detail. Aside from developing and applying FFLUX directly, a number of auxiliary themes are also explored that pertain to individual FFLUX components. For example, development of a replacement for the short-range repulsive component of the inter- molecular Lennard-Jones potential, and development of improved sampling methods for training point generation.