A team of computational scientists at the Department of Energy’s Oak Ridge National Laboratory has generated and released datasets of unprecedented scale that provide the ultraviolet visible spectral properties of over 10 million organic molecules. Understanding how a molecule interacts with light is essential to uncovering its electronic and optical properties, which in turn have potential photoactive applications in products such as solar cells or medical imaging systems. 

Using high-performance computing resources at the Oak Ridge Leadership Computing Facility, the ORNL team ran quantum chemistry calculations to generate the vast datasets. For each of these organic molecules, the team ran atomistic material modeling calculations with various approximations to compute different excited-state properties of interest. The team’s findings were published in Nature Scientific Data. 

The ultimate intended use for the open-source datasets is to train a deep learning model to identify molecules with tailored optoelectronic and photoreactivity properties, an approach that is much faster and easier to conduct than current methods. 

“The use of DL models for molecular design is essential because the chemical space that must be explored for the search of these molecules is extremely large,” said lead author Massimiliano Lupo Pasini, a data scientist in ORNL’s Computational Sciences and Engineering Division. 

“Both experiments and existing first-principles calculations, which are based on the physical laws that determine how matter and energy interact at the subatomic level, are simply unaffordable for different reasons. Experiments are labor intensive, and first-principles calculations can easily slam supercomputing facilities. But DL models provide very promising tools to overcome these barriers,” Lupo Pasini said. 

The project got off the ground when Stephan Irle, leader of ORNL’s Computational Chemistry and Nanomaterials Sciences group, identified the ultraviolet-visible spectrums of molecules as a useful property to predict with DL models. Building a DL model sufficiently complex to identify desirable molecular properties requires training it with huge volumes of data that explore all different regions of chemical space. The more data collected, the more the DL model trained on it can achieve the necessary robustness and generalizability to function effectively. However, collecting such large volumes of scientific data for scalable DL may present data-flow issues, especially at facilities with multiple users like the OLCF, a DOE Office of Science user facility located at ORNL.