<p>With ever-expanding heterogeneous catalyst data generated from experiments and theoretical calculations, machine learning (ML) has emerged as a powerful tool to predict key catalyst features from large databases and identify stable, active, and selective catalysts across diverse applications. However, to account for surface complexities such as surface defects, alloying, and support interactions, advanced approaches are essential to extract relevant insights while utilizing minimal computational effort. In this review, recent advancements in catalyst discovery addressing these complexities, with a focus on the prediction of DFT-derived properties through physics-inspired, data-driven, and interpretable ML approaches, are highlighted, along with recent applications in high-throughput screening of alloys, metal oxides, and single-atom catalysts. Further, current challenges and future perspectives on the application of ML-based approaches in heterogeneous catalyst discoveries are presented.</p>