Session III — High Performance, Quantum and Scientific Computing

Monday, October 26 · 2:00–3:30 p.m.
Ballroom A
Chair: Harish Bhat

New Methods for Splitting Stochastic Linear Programs

Robert Bassett · Naval Postgraduate School

We introduce a method for splitting Stochastic Linear Programs, derived from the Splitting Conic Solver of O'Donoghue et al. This method exploits problem structure inherent to stochastic programs, and utilizes linear algebra operations which are amenable to implementation on a GPU. In this talk, we will introduce the method, describe its benefits relative to competitors like interior point methods and PDLP, and discuss its computational performance in applications relevant to national security.

Quantized Tensor Trains for Gaussian Process Regression

Michael Kielstra · UC Berkeley

Gaussian process regression (GPR) is a common tool for estimating an underlying function from scattered, noisy observations. However, naively computing the solution to a GPR problem requires forming, storing, and inverting a large matrix. Recently, there has been significant interest in representing large matrices as tensor networks. We specifically use a matrix product operator (MPO) format, coupled with a tensor-train (TT) representation of our inferred function, to achieve efficient storage and computation across a wide range of common GPR scenarios. The resulting algorithm, built largely from standard, proven MPO/TT routines, has a runtime that grows only linearly with the number of observations, while simultaneously using less memory than the naive approach. Preliminary data validate our method for both one- and multi-dimensional GPR problems.

Analysis of Vertical Discretizations for Space Weather Models

Francis X Giraldo · Naval Postgraduate School

We analyze and compare various options for developing vertical discretization schemes for the ground-to-thermosphere variant of the Navy's NEPTUNE global weather prediction model. The goal is to stabilize the numerics at high altitude where maintaining stability becomes challenging. We will report on the analysis of finite difference, finite element (finite element exterior calculus or FEEC), finite volume, and discontinuous Galerkin methods and on additional mechanisms required for maintaining stability. We will report on the conservation properties of these schemes such as mass (absolutely essential), total energy, and entropy. We are also interested in the time to solution of the methods. To compare the various methods, we use acoustic wave test cases.

Electromagnetic Modeling of Superconducting Quantum Chips

Andy Nonaka · Lawrence Berkeley National Laboratory

Effective design of superconducting quantum chips requires advanced simulation tools for modeling electromagnetic wave propagation in complex circuits. We present recent efforts using ARTEMIS, a framework for solving the time-dependent Maxwell equations in microelectronic systems, to predict and optimize signal propagation in these chips. We describe a hierarchy of classical models for superconducting physics in intricately patterned device geometries, together with nonlinear circuit elements that capture relevant aspects of qubit behavior. We demonstrate the scalability of ARTEMIS through full-machine simulations on Perlmutter of dual-layer “flip-chip” processor designs. Finally, we show how LLM-based scientific agents can invoke simulation and analysis tools and make context-dependent decisions within automated geometric design-optimization workflows.

Scalable Characterization of Quantum Computers Using Quantum Physics-Aware Neural Networks

Bradley Yount · UC Merced

For quantum computing to be successful, we must be able to learn and reduce error within the system. State-of-the-art systems now contain hundreds of qubits, and many characterization methods such as process tomography or gate set tomography (GST) cannot scale. More complex forms of error such as qubit-qubit interactions must be understood and managed as scale and density increase. We propose characterizing a quantum computer's errors with neural networks whose parameters correspond to the rates of different kinds of errors in a physically interpretable, sparse basis. These parameters can then be used to predict noisy circuit outcomes using efficient approximations. In addition to learning error rate information, the network also allows for scaling into intermediate ensembles of tens to hundreds of qubits. We demonstrate a physics-aware network’s ability to accurately predict noisy circuit outcomes and learn context-dependent and coherent crosstalk errors in a simulated 4-qubit system.