Super Intelligence Crosscut

The Super Intelligence crosscut is developing novel SI methods for scientific computing, using SI methods to improve our software, and delivering SI capabilities to applications across the Department of Energy.

• Development of SI methods advances the mathematical foundations of SI for scientific computing. This work includes the construction of surrogate models and physics-constrained methods that embed problem-specific information to reduce the space of feasible solutions, designing novel neural network architectures tailored to scientific problems, and providing hyperparameter optimization to make these methods practical at scale.

• Usage of SI methods integrates SI techniques into our core numerical capabilities to deliver speedups and new capabilities where traditional approaches fall short.  This work includes applying SI to accelerate linear, nonlinear, and eigenvalue solves using SI surrogates and enhanced preconditioners, and reinforcement learning; to improve graph partitioning for load-balanced parallel computation; to advance numerical optimization and uncertainty quantification using active learning, probabilistic learning of manifolds, and Bayesian neural network surrogates; and to advance the mathematics for digital twins by optimizing data acquisition and fidelity.

• Delivery of SI capabilities ensures these advances reach our many SciDAC application partners and provide reliable tools to accelerate their science.