Conventional methods for determining and monitoring the viscosity of oils are time-consuming, expensive, and in some instances, technically unfeasible. These limitations can be avoided using low-field nuclear magnetic resonance (LF-NMR) relaxometry. However, due to the chemical dissimilarity of oils and various temperatures these oils are exposed to, as well as LF-NMR equipment limitations, the commonly used models fail to perform at a satisfactory level, making them impractical for use in heavy oil and bitumen reservoirs and in environments with large temperature oscillations (e.g., mechanical systems). We present a framework that combines supervised learning algorithms with domain knowledge for synthesizing new features to improve model forecasts using only one NMR parameter-T2 geometric mean. Two principal methods were considered, support vector regression (SVR) and gradient boosted trees (GBRT). Models were trained using the experimental data from our previous studies and literature data combining conventional oils, heavy oils, and bitumens from various reservoirs in Canada and United States. The models’ performance was compared against four other intelligent algorithms and four well-known empirical NMR models against which the SVR- and GBRT-based models achieved the highest statistical scores. These two models can be used for oil viscosity prediction in conventional and heavy oil reservoirs with a wide range of oil viscosities and in situations where high precision is needed, such as in the determination of viscosity of petroleum distillates or for monitoring of oil viscosity in mechanical systems. The proposed framework can also be applied to determine other physicochemical properties of oils by LF-NMR, where the application of supervised learning is usually impractical due to the limited volume of experimental data.