Integrated Computational and Data Infrastructure (ICDI) for Scientific Discovery

The DOE SC program in Advanced Scientific Computing Research (ASCR) hereby announces its interest in funding research and development projects to create an advanced Integrated Computational and Data Infrastructure (ICDI) program. This FOA is composed of two topics.

Topic A “Experimental/Computational/Computer Science collaborations” addresses the challenge of creating collaborative teams of scientists to accelerate science discoveries supported by the SC programs. Applications to this topic must be submitted by multi-investigator teams.

Topic B: “Intelligent Distributed Infrastructure Simulation Capabilities” addresses the challenge of modeling, simulating, and validating the performance of geographically distributed science infrastructures. Both single and multiple investigator applications may be submitted.

The Integrated Computational and Data Infrastructure (ICDI) program will accelerate research activities across the entire SC complex. Over the past four decades experimental scientists (domain scientists working with a physical device, at a user facility, or in the field to understand scientific interactions), computational scientists (domain scientists developing and executing simulation codes to explore scientific phenomenon on leadership class computers), and computer scientists (computer/data scientists and applied mathematicians developing scientific algorithms and codes) have increasingly used a wide array of computers and experiments to generate, analyze, and manage vast amounts of data. This data may be transferred out of the facility and stored at PI’s home institution, a facility storage repository, or in some large central or distributed repository which individuals and teams access to conduct the detailed analysis tasks. Combining, in real time, a variety of data, facilities and computers resources coupled with machine learning and artificial intelligence techniques will greatly accelerate scientific discovery.

The growth of computational science as a peer method for scientific discovery has been nothing short of phenomenal. In less than four decades, computational science has grown from a small supporting activity where computers were used simply to analyze experimental data to a recognized scientific discipline where simulations explain and predict scientific phenomena and generate scientific data. During the same time Scientific User facilities experiments have been taking advantage of Complementary Metal Oxide Semiconductor (CMOS) technologies to build better, faster, more capable instruments and detectors, thus increasing the volume and rates of data that is being generated. More recently, Artificial Intelligence/Machine Learning/Deep Learning (AI/ML/DL) technologies are being incorporated in both computational and experimental science efforts.

Moving forward it is increasingly necessary to leverage the symbiotic relationships that exist between the experimental and computational sides of numerous DOE science communities (i.e., biology, chemistry, cosmology, Earth system science, environmental science, geosciences, material science and physics). Coupling AI/ML/DL with this continuum of computational and experimental resources (edge to supercomputer), each specialized for a specific task will greatly accelerate the analysis power of the entire system, at a huge cost in complexity. AI/ML/DL technologies rely on massive amounts of data to train, and continuously retrain, models which require a robust data storage and management environment that may also be distributed. Work is just beginning to perform training on large heterogeneous supercomputers with inference performed by specialized computational cores near the edge or at other locations along the endto-end data path. This requires vast amounts of data to be sent, or generated via simulation, at the supercomputer facility and continuous model updates be returned to the inference site. Managing this complexity effectively and efficiently will be a major challenge for all concerned.

Awards under this FOA will develop new software workflows and tools to accelerate the scientific discovery process through the convergence of experimental/simulation data, computational/experimental facilities and a broad community of scientists to both generate high fidelity simulations and steer experiments. It will also develop the modeling and simulation capabilities needed to predict and debug workflow performance in distributed computational and data infrastructures.

Applicant institutions are limited to no more than three (3) letters of intent or applications. 

Deadlines:

• Duke Internal: Interested applicants from within Duke should contact dukeiln@duke.edu as early as possible.
• Required LOI: April 5, 2021
• Applications: May 14, 2021