Accelerating materials discovery pipelines: Incorporating multimodal approaches from neutron scattering data to deep learning and high-accuracy first principles calculations to improve credible predictions

Materials development is hampered by long timelines for the discovery-to-production lifecycle, which can span decades and may require testing and characterization of millions of samples. Accelerating materials discovery represents an urgent need as we address energy, sustainability, climate, infrastructure and biomedical challenges. Supporting the accelerated discovery of diverse types of novel materials by developing credible, trustworthy, and energy-efficient predictive modeling methods is a grand challenge in the field; new paradigms in materials research incorporating deep learning, high-throughput and automated experiments and data science in iterative feedback loops have been envisioned by experts as the key to acceleration. Here we report on our work incorporating deep learning models, data from neutron scattering experiments including inelastic vibrational spectroscopy, and high-accuracy first-principles theoretical calculations using supercomputers to help tackle the grand challenge problem of predicting the most likely solid crystalline structure to be formed given a set of physical conditions. We analyze sources of uncertainty from physical assumptions to numerical errors that can combine and compound within these complex, multi-component workflows and provide some suggestions and solutions.