Research
Learning reduced models under extreme data conditions for design and rapid decision-making in complex systems.
Extreme Data
Energy Earthshot applications involve complex systems with rich physics phenomena which require large-scale, distributed computing to simulate. This creates extreme data conditions where the data must be stored across multiple locations, can only be accessed once during training, or may fail to capture rare and extreme events.
Challenges
1. Data are computed in a distributed manner across multiple nodes because high-fidelity numerical solvers are typically parallelized via domain decomposition techniques.
2. Full state data vectors of high-fidelity physics solvers are rarely written to disk but instead must be processed in-situ as they are computed during time stepping; both because of storage limitations and because writing to disk can increase the run time per time step tenfold.
3. Tail and rare events are typically not seen during simulation runs with standard inputs, and thus data generated without active steering contain little information for learning reduced models that are predictive in such critical regions of interest
Objective and research thrusts
The overall objective of this project is learning reduced models when we do not have access to complete, batch training data, which is especially important for several of the Earthshots and other DOE/ASCR applications that rely on large-scale, distributed computing
Reduced models based on deep networks
We propose to go far beyond the current state of the art by developing mathematical foundations and computational methods that enable (i) actively collecting data that are informative about rare events and (ii) learning reduced models based on deep networks and other nonlinear parametrizations from distributed and streamed data.
Active data acquisition and rare events
We are developing simulations techniques to actively collect data for learning reduced models that can predict well rare and extreme events.
Energy Earthshots
The proposed methodologies will be integrated in computational processes for simulation, design, and decision-making of the Floating Offshore Wind and Carbon Negative Shot, where we will leverage reduced models for speeding up parts of the overall computational process while relying occasionally on the expensive, high-fidelity physics solvers in a principled way to quantify uncertainties and maintain accuracy guarantees.
