Fantastical Training Data and How To Generate It

In an ideal world we would train machine learning models for organic materials on training data generated using coupled-cluster singles-doubles-triples-quadruples or something similar and fully take into account the environment of the molecules. In practice this is impossible considering the number of training points required for most models. As a result, we have to make choices. Even more so when we want our model not to predict ground state energies and/or geometries but excited state properties, such as optical excitation energies or electron affinities/ionisation potentials, as well as to be transferable. Models that could accurately predict excited state properties for a wide range of organic materials would be very useful in virtual high-throughput screening of materials for organic LEDs, photovoltaics and photocatalysts.<br/><br/>In my contribution I’ll discuss our work and that of others on training data generation for excited-state properties of organic materials. I’ll consider which properties to aim for, what the best possible training data might look like and how hybrid schemes based on tight-binding methods calibrated to higher-level methods can be exploited.