Faculty Research Participant - MCS - Sen Na

Argonne National Laboratory•Lemont, IL

6d•Onsite

About The Position

The main idea for the visit is to gain hands-on experience with GPU-based implementations of modern preconditioning and structured optimization methods, and to understand how far these techniques can be pushed in control and scientific machine learning problems. In particular, recent work by Mihai Anitescu's group on GPU-accelerated optimization and control highlights practical benefits of bringing advanced numerical methods onto GPU platforms (e.g., GPU-enabled solvers for large- scaleoptimalcontrolandnonlinearprogramming). These efforts demonstrate how careful algorithm design (mainly for interior-point method), memory management, and exploitation of hardware parallelism can yield significant performance improvements for problems that are traditionally solved on CPUs. Building upon this context, I hope to extend existing open-source GPU tooling by incor- porating more recent preconditioning-based approaches (such as Muon, SOAP, and Shampoo) and evaluating their effectiveness in structured control settings that are of broad interest in DOE applications (Genesis Mission).

Requirements

GPU-based implementations of modern preconditioning and structured optimization methods
GPU-accelerated optimization and control
GPU-enabled solvers for large-scale optimal control and nonlinear programming
Careful algorithm design (mainly for interior-point method)
Memory management
Exploitation of hardware parallelism
Extend existing open-source GPU tooling by incorporating more recent preconditioning-based approaches (such as Muon, SOAP, and Shampoo)
Evaluating their effectiveness in structured control settings that are of broad interest in DOE applications (Genesis Mission)