A Scalable Distributed Workflow for Quantum Chemical Prediction of UV/Vis Absorption Spectra for Organic Molecules
1
Scientific Achievement
An MPI-based HPC workflow that can efficiently manage large number of files was developed to generate two open-source datasets for training and predicting properties of organic molecules. A coordinator-worker design was developed for dynamic load balancing of molecule calculations and efficiently using the hierarchical storage on the HPC system for managing over 90 million files generated by over 10 million molecules.
Significance and Impact
A highly scalable MPI-based workflow was developed to generate two open-source datasets - GDB-9-Ex and ORNL_AISD-Ex - that provide calculations of electronic excitation energies and their associated oscillator strengths for organic molecules. These datasets will be used to train a graph convolutional neural network for predicting properties of organic molecules.
Two open-source datasets – GDB-9-Ex and ORNL_AISD-Ex – that contain over 96,000 and 10 million molecules respectively were generated using first principles time-dependent density functional tight-binding (TD-DFTB) calculations for predicting properties of organic molecules. The figure shows the strong correlation between the HOMO-LUMO gap and the minimum absorption energy for the large number of organic molecules. These datasets will be used to train surrogate models that can be used in the design of compounds. The workflow implements a coordinator-worker pattern for dynamically assigning molecules to processes. It uses an ephemeral in-memory file system along with persistent storage for scaling the workflow for over 90 million files.
Technical Approach
GDB-9-ex: https://www.osti.gov/biblio/1890227, ORNL_AISD-Ex: https://www.osti.gov/biblio/1907919