Sandia Labs FY22 Laboratory Directed Research & Development Annual Report

ACCOMPLISHING MORE WITH LESS COMPUTATION USING MONTE CARLO SAMPLING-BASED PARTICLE SIMULATIONS. Particle models for Brownian dynamics, radiation

State University, provided a proof-of-concept demonstrating a gradient-based optimization capability for Monte Carlo sampling-based particle simulations is possible. The PI for this two-year project gave two invited talks and two contributed talks on the research. (PI: Richard Lehoucq)

transport, low-density fluids, and plasmas lack a gradient-based optimization capability for various problems of interest, e.g., inverse problems. Such a capability enables the particle simulation communities to go beyond forward simulation by accomplishing more with less computation. Gradient based methods crucially depend upon sensitivities, which are synonymous with the calculation of a

derivative that measures the change in a quantity with respect to a change in another quantity. The team, who partnered with faculty at Tulane University, University of Delaware, and Sandia National/Regional partner North Carolina An example problem involves inferring the location of the orange material along the horizontal axis given radiation leakage measurements at the left and right. The example, showing four different gradient descent iterations computed via reuse of Monte Carlo samples for the desired values of 2 cm and 6 cm overlaying the surface, is proof of-concept that a gradient-based optimization capability for Monte Carlo sampling-based particle simulations is possible.

69

LABORATORY DIRECTED RESEARCH & DEVELOPMENT