The magnets with TM14RE2B phase (TM = transition metal/s, RE= rare earth metal/s, B= Boron also called 14:2:1) are popularly used for energy conversion applications because of their high energy product (430 kJ/m3). Some of the applications involve high power motors, generators, wind turbines and electric automobiles. The excellent magnetic properties are due to the intrinsic properties which are based on the chemical composition of 14:2:1 phases. Intrinsic magnetic properties like magnetic saturation, Curie temperature, anisotropy field and others are responsible for operability of magnets at the desired operating temperatures without degradation in their performance. Currently the rare earth metals have become the most sought-after elements in the world, due to their numerous applications in modern technologies like semiconductors, LEDs, magnets, batteries, etc. Exploration of various chemical combinations of rare earth metals along with transition metals can further lead to better magnetic properties. Synthesizing magnets in the laboratories based on all theoretical chemical compositions and then measuring magnetic properties would not be a feasible alternative. Along with the properties, the chemical composition also determines manufacturing cost and corresponding carbon footprint (sustainability) of magnets. An efficient approach to explore various chemical compositions and optimize it would be utilization of Artificial Intelligence techniques. There are data-driven approaches utilized for predicting different intrinsic magnetic properties from chemical compositions specifically for 14:2:1 phases, such as predicting mass density by Kini et al., Curie temperature and saturation magnetization by Choudhary et al.. The basis of the predictions is the chemical composition data given as an input for training the models, which is available in limited amount in existing literature. These trained models have limited information on the underlying physics as they are based on chemical compositions. Due to small dataset size, the trained models do not generalize well on all chemical compositions of 14:2:1 phases. In such cases, the model struggles with uncertainty and additional information in terms of physics laws can act as useful information. The proposed research explores utilization of various physics laws available in the magnets literature to improve the predictions of the data driven models. With this approach, various theoretical chemical combinations can be explored and different magnetic properties can be predicted. The optimized magnetic compositions can then be synthesized in the laboratories for final validation.