The article by Jörg Behler reviews the potential of machine learning (ML) in representing potential energy surfaces (PES) for atomistic simulations in chemistry, physics, and materials science. ML methods, which have traditionally been used for classification tasks, are now being applied to the more complex task of fitting PESs from large datasets of electronic structure calculations. The central ideas, challenges, and current applicability of ML potentials are discussed, emphasizing the importance of suitable structural descriptors and the choice of ML methods. Key descriptors include atom-centered symmetry functions (ACSFs), bispectrum of neighbor density, smooth overlap of atomic positions (SOAP), and Coulomb matrices. ML potentials offer significant advantages such as fast computation, high accuracy, and the ability to handle large systems, but they also face challenges like the need for large reference datasets and the complexity of configuration space. The article highlights the potential of combining ML with physically meaningful energy terms to overcome these limitations and discusses the future trends and limitations of ML potentials in atomistic simulations.The article by Jörg Behler reviews the potential of machine learning (ML) in representing potential energy surfaces (PES) for atomistic simulations in chemistry, physics, and materials science. ML methods, which have traditionally been used for classification tasks, are now being applied to the more complex task of fitting PESs from large datasets of electronic structure calculations. The central ideas, challenges, and current applicability of ML potentials are discussed, emphasizing the importance of suitable structural descriptors and the choice of ML methods. Key descriptors include atom-centered symmetry functions (ACSFs), bispectrum of neighbor density, smooth overlap of atomic positions (SOAP), and Coulomb matrices. ML potentials offer significant advantages such as fast computation, high accuracy, and the ability to handle large systems, but they also face challenges like the need for large reference datasets and the complexity of configuration space. The article highlights the potential of combining ML with physically meaningful energy terms to overcome these limitations and discusses the future trends and limitations of ML potentials in atomistic simulations.