Barium titanate (\ce{BaTiO3}, BTO), a well-known perovskite oxide, undergoes intricate ferroelectric phase transitions. These transitions are characterized by a shift from a paraelectric cubic phase at high temperatures to a sequence of low-temperature phases (Cubic → Tetragonal → Orthorhombic → Rhombohedral), predominantly driven by antiferrodistortive modes. Although \textit{ab-initio} molecular dynamics is used to explore the finite-temperature properties, the high computational cost and scaling to only a few hundred atoms restricts the study for a longer time and length scale. Conversely, classical molecular dynamics simulations, though efficient in understanding atomic-scale dynamics, often lack accuracy compared to first-principles-based methods. To overcome these limitations, we develop a machine learning interatomic potential (MLIP) for \ce{BaTiO3}, based on Atomic Cluster Expansion (ACE) formalism using data from density functional theory (DFT) calculations. The ML potential achieves DFT-level accuracy while facilitating simulations over significantly extended time and length scales. Using the trained potential, our research investigates the temperature-driven cubic-to-Rhombohedral phase transition in BTO and examines the influence of pressure on the transition temperature. Moreover, we highlight the capability of the potential to describe BTO grain boundaries and domain wall structures. We tested our BTO machine learning potential to understand the different grain boundary structures in BTO and compared it with the classical shell-type model.