We propose to continue exploring a recently developed neural network potential, So3krates [Frank2022], for the simulation of temperature-dependent properties of complex materials. These properties range from thermodynamic stability to thermal transport and spectroscopy, relevant both to a fundamental understanding of solid state systems as well as technological applications such as waste heat management and optoelectronics [Knoop2023, Ye2024]. Qualitative and quantitative understanding requires a microscopic analysis of the atomic motion (= heat) which can be gained from molecular dynamics (MD) simulations or phonon theory. These simulations require an accurate description of the interatomic interactions which can be achieved with density functional theory (DFT), or machine learning potentials (MLPs) trained on DFT reference data [Langer2023, Langer2023b]. A particular challenge are cases where important dynamical processes occur only seldomly and on large time scales -- so-called rare events. An example from our own research is the spontaneous formation of Frenkel pairs in copper iodide (CuI) which occurs on the time scale of dozens of picoseconds and can therefore be overlooked when creating training data [Knoop2023]. In this project, we plan to detect rare events and other dynamical phenomena that pose a challenge to the accurate parametrization of machine learning potentials by equipping the model with the ability to assess the uncertainty of its predictions. If a configuration with high uncertainty is encountered, the simulation can be halted and additional training data can be generated, creating an active learning loop. Uncertainty quantification for machine-learning potentials is an active field of research, in particular for neural network potentials. We propose to evaluate a recently developed method, direct propagation of shallow ensembles [Kellner2024], for the So3krates model and the task of vibrational dynamics simulations. Shallow ensembles have the advantage of introducing only a small performance overhead compared to traditional committee approaches, and can provide statistically meaningful uncertainties for aggregated properties like the potential energy or forces. Insights gained from this project will not be limited to the So3krates potential, but will be helpful for any neural network MLP. This work continues the developments started in the projects Berzelius-2023-28 and NAISS 2023/22-478. Training data is either already available [Knoop2023] or will be obtained through other projects like NAISS 2024/1-37. The work is done in collaboration with Marcel Langer and Matthias Kellner from EPFL Lausanne, Switzerland. The scientific part of the project is supported by the Swedish Research Council (VR) program 2020-04630, and the Swedish e-Science Research Centre (SeRC). ## References [Frank2022] JT Frank, OT Unke, KR Müller, in *Advances in Neural Information Processing Systems*, Vol. **35** (2022), pp. 29400–29413 [Knoop2023] F Knoop, et al., Phys Rev Lett **130**, 236301 (2023) [Ye2024] K Ye, M Menahem, T Salzillo, F Knoop, *et al.*, https://arxiv.org/abs/2402.18957 [Langer2023] MF Langer *et al.*, Phys. Rev. B **108**, L100302 (2023). [Langer2023b] MF Langer, JT Frank, and F Knoop, J. Chem. Phys. **159**, 174105 (2023) [Kellner2024] M Kellner, and M Ceriotti, https://arxiv.org/abs/2402.16621