SHNITSEL - Surface Hopping Nested Instances Training Set for Excited-state Learning

SHNITSEL The Surface Hopping Nested Instances Training Set for Excited-State Learning (SHNITSEL) is a comprehensive data repository designed to support the development and benchmarking of excited-state dynamics methods. Configuration Space SHNITSEL contains datasets for nine organic molecules that represent a diverse range of photochemical behaviors. The following molecules are included in the dataset: Alkenes: ethene (A01), propene (A02), 2-butene (A03) Ring structures: fulvene (R01), 1,3-cyclohexadiene (R02), tyrosine (R03) Other molecules: methylenimmonium cation (I01), methanethione (T01), diiodomethane (H01) Property Space These datasets provide key electronic properties for singlet and triplet states, including energies, forces, dipole moments, transition dipole moments, nonadiabatic couplings, and spin-orbit couplings, computed at the multi-reference ab initio level. The data is categorized into static and dynamic data, based on its origin and purpose. Static data (#147,169 data points in total) consists of sampled molecular structures without time-dependent information, covering relevant vibrational and conformational spaces. These datasets are provided for eight molecules: A01, A02, A03, R01, R03, I01, T01, and H01 Dynamic data (#444,581 data points in total) originates from surface hopping simulations and captures the evolution of molecular structures and properties over time, as they propagate on potential energy surfaces according to Newton’s equations of motion. These datasets are provided for five molecules: A01, A02, A03, R02, and I01 Data Structure and Workflow The data is stored in xarray format, using xarray.Dataset objects for efficient handling of multidimensional data. Key dimensions include electronic states, couplings, atoms, and time frames for dynamic data. The dataset is scalable and compatible with large datasets, stored in NetCDF4 format within HDF5 for optimal performance. Tools for data processing, visualization, and integration into machine learning workflows are provided by the shnitsel Python package published on Github (shnitsel-tools) .(https://github.com/SHNITSEL/shnitsel-tools). An overview of the molecular structures and visualizations of key properties (from trajectory data) are compiled on the SHNITSEL webpage (https://shnitsel.github.io/).