<p>In recent years, first principles computational methods such as Density Functional Theory (DFT) have occupied an indispensable position in a catalysis researcher’s toolkit. Together with increasingly detailed experimental characterization techniques, first principles methods have elucidated key relationships between the structure of a catalyst at the atomistic scale and its performance in terms of activity, stability, and selectivity trends. Today, there is a growing need to discover and design new catalysts to address the challenges linked to the sustainable production of fuels, chemicals, and energy. Computational workflows for rapid and automated screening of catalysts can aid in this quest to discover new catalysts; however, DFT calculations remain a significant bottleneck in such workflows, requiring enormous computational resources. Recently, machine learning-based approaches involving neural networks, commonly termed as deep learning, have promised to greatly accelerate molecular and materials property prediction. Many deep learning models, including artificial neural networks, graph neural networks, and large language models, have been employed as efficient surrogates for DFT in catalytic property estimation, molecular dynamics simulations, electronic structure prediction, among other applications. In this review, we discuss the fundamentals and applications of deep learning models in the context of computational heterogeneous catalysis. We examine the architecture of each model and discuss strategies that allow efficient mapping of atomistic structures to their properties. In addition, we review various open, materials datasets that have catalysed the emergence of foundation models that work across the periodic table. Lastly, we discuss specific applications of deep learning models in typical computational catalysis workflows, such as descriptor estimation, kinetic barrier calculations, enhanced sampling approaches, and briefly explore potential future applications.</p>