Machine learning (ML) potentials have become a key tool in atomistic simulations, offering a way to accurately represent potential energy surfaces (PESs) using data from electronic structure calculations. This perspective reviews the development and application of ML potentials, highlighting their advantages over traditional empirical potentials. ML potentials use a set of descriptors to represent atomic configurations and then apply ML methods to predict the energy of these configurations. These descriptors must be invariant to translations, rotations, and permutations of atoms to ensure accurate and consistent results. The most common descriptors include atom-centered symmetry functions (ACSFs), bispectra, and smooth overlap of atomic positions (SOAP). These descriptors capture the local environment of each atom and are used to train ML models that can predict energies and forces. ML potentials are particularly useful because they can handle high-dimensional systems with many atoms, which is challenging for traditional methods. They also provide accurate results that are close to those obtained from first-principles calculations, making them suitable for simulations requiring high precision. ML potentials are constructed by combining a descriptor that transforms atomic configurations into a form suitable for ML and a ML method that maps these descriptors to energy values. This approach allows for efficient and accurate simulations of complex systems. However, the development of ML potentials requires large reference datasets from electronic structure calculations, which can be computationally intensive. Despite this challenge, ML potentials offer significant advantages in terms of accuracy and efficiency, enabling simulations of systems that were previously infeasible. The future of ML potentials lies in their integration with physically meaningful energy terms and the development of more efficient descriptors. By combining ML methods with traditional physics-based approaches, it is possible to create potentials that are both accurate and computationally efficient. This integration will help overcome the limitations of current ML potentials and expand their applicability to a wider range of systems. Overall, ML potentials represent a promising advancement in the field of atomistic simulations, offering a powerful tool for studying complex materials and chemical systems.Machine learning (ML) potentials have become a key tool in atomistic simulations, offering a way to accurately represent potential energy surfaces (PESs) using data from electronic structure calculations. This perspective reviews the development and application of ML potentials, highlighting their advantages over traditional empirical potentials. ML potentials use a set of descriptors to represent atomic configurations and then apply ML methods to predict the energy of these configurations. These descriptors must be invariant to translations, rotations, and permutations of atoms to ensure accurate and consistent results. The most common descriptors include atom-centered symmetry functions (ACSFs), bispectra, and smooth overlap of atomic positions (SOAP). These descriptors capture the local environment of each atom and are used to train ML models that can predict energies and forces. ML potentials are particularly useful because they can handle high-dimensional systems with many atoms, which is challenging for traditional methods. They also provide accurate results that are close to those obtained from first-principles calculations, making them suitable for simulations requiring high precision. ML potentials are constructed by combining a descriptor that transforms atomic configurations into a form suitable for ML and a ML method that maps these descriptors to energy values. This approach allows for efficient and accurate simulations of complex systems. However, the development of ML potentials requires large reference datasets from electronic structure calculations, which can be computationally intensive. Despite this challenge, ML potentials offer significant advantages in terms of accuracy and efficiency, enabling simulations of systems that were previously infeasible. The future of ML potentials lies in their integration with physically meaningful energy terms and the development of more efficient descriptors. By combining ML methods with traditional physics-based approaches, it is possible to create potentials that are both accurate and computationally efficient. This integration will help overcome the limitations of current ML potentials and expand their applicability to a wider range of systems. Overall, ML potentials represent a promising advancement in the field of atomistic simulations, offering a powerful tool for studying complex materials and chemical systems.