Sandia Labs FY22 Laboratory Directed Research & Development Annual Report

REDLY: RESILIENCE ENHANCEMENTS THROUGH DEEP LEARNING YIELDS OFFER BENEFITS FOR POWER SYSTEMS. The planning, operation, and real-time security assessment of power systems require the solution of complex optimal power flow problems

up to 3x improvement in prediction accuracy and 10x improvement in worst-case errors over baseline approaches. The project resulted in the design of the Optimization

that are computationally intractable. Contemporary approaches leverage sub optimal approximations of the power flow physics that limits the ability of operators to protect the grid against threats leading to interrupted service or physical damage. This LDRD addressed these challenges by developing a complete framework for the design, validation, and deployment of physics-informed neural network (PINN) surrogates for power systems. The LDRD

and Machine Learning Toolkit that demonstrated improved verification of PINNs across multiple applications. It also provided an opportunity to collaborate with Imperial College and Carnegie Melon and facilitate growth opportunities for a PhD intern from Alliance partner Georgia Tech and an undergraduate from National/Regional partner

University of Florida. Journal publications from 2022 include Computer Aided Chemical Engineering and Computers & Chemical Engineering . (PI: Michael Shannon Eydenberg)

REDLY PINNS (blue) demonstrate significantly improved worst-case violation error over non-PINN baselines (red).

team applied Lagrangian-based regularizations of power flow physics to design surrogates with

MITIGATING POWER ATTENUATION AND ENABLING MEGASONIC COMMUNICATION IN DEFENSE APPLICATIONS THAT RELY ON HERMETICALLY SEALED FARADAY CAGES.

The electronic shielding necessitated by different mission applications requires new approaches to power transmission and mechanical communication (transduction). This project, which leveraged Sandia’s high performance computing and Sierra’s massively parallel implementation of the adjoint method for partial differential equation-constrained optimization, enabled a new set of tools for the design and optimization of piezoelectric-based ultrasonic mechanical transduction channels. The risk-averse optimization methods developed in the project enable wide bandwidth data rate communication while mitigating power attenuation. This was

accomplished by developing electrical-mechanical coupled physics modeling, simulation and optimization capabilities that optimized electrical circuit parameters, including thermal preloads, material properties, geometries, and signal optimization. In addition, a posteriori error based reduced order modeling techniques were combined with optimization under uncertainty in order to minimize computational expense while achieving optimal designs that are robust to model uncertainties. Several follow-on projects are currently leveraging the new modeling and optimization capabilities for early design studies. (PI: Timothy Walsh)

53

LABORATORY DIRECTED RESEARCH & DEVELOPMENT