Advances in Machine Learning to Improve Scientific Discovery at Exascale and Beyond

Project Status: Active

Deep learning’s powerful ability to capture rich features, directly from the raw scientific data, makes it an attractive choice for working with complex scientific data sets; however, applying deep learning to complex scientific data presents many challenges.  In the past, deep learning research has been focused mainly on classification of relatively simple data, such as photographs, resulting in models that are highly tuned for that specific task, and not well-suited to complex, scientific data that often contain spatiotemporal features that can be very different in their scale and magnitude.  Addressing this issue, we propose using deep learning to both find and classify features within raw sensor readings from scientific Physics and Biology experiments.  This approach requires a deep learning network capable of ingesting tens to hundreds of thousands of inputs, an automated way of constructing the network layers and parameters, and the ability to learn the features over hundreds of thousands of matrices of sensor readings.  This will necessitate leveraging the Oak Ridge National Laboratory’s (ORNL's) Titan and follow-on Summit high performance computing (HPC).

Last Updated: September 1, 2020 - 8:20 pm