Carl Laird, Carnegie Mellon University

"Democratization of High-Performance Computing at the Interface of Optimization and Data Science"

Abstract:

Emerging global challenges in science and engineering are pushing the performance boundaries of off-the-shelf optimization tools. However, historically, the development of tailored, high-performance computing approaches required significant time and expertise in low-level compiled languages. The success of high-level languages like Python and Julia have helped democratize high-performance computing, putting these capabilities in the hands of the broader research community with the promise of increased impact. We have built several advanced optimization tools, and in this presentation, I will focus on three vignettes to illustrate the effectiveness (and ease of use) of these capabilities in Python.

The COVID-19 pandemic has clearly exposed significant challenges in effective mitigation of emerging infectious diseases. I will discuss large-scale optimization techniques and the use of ParaPint to efficiently estimate county-level transmission parameter dynamics using a fully-coupled, national-scale model. We have used these transmission parameter profiles to estimate the impact of different non-pharmaceutical interventions on the spread of COVID-19.

Machine learning (ML) models are being increasingly used as surrogates for complex processes within engineering. The Optimization and Machine Learning Toolkit (OMLT) supports the use of ML models within optimization-based decision-making applications. Here, I will showcase the capabilities of OMLT, discuss performance improvements using disjunctive programming formulations, and demonstrate the capabilities on engineering applications, including a collaboration with the Boukouvala group at GeorgiaTech.

Progress on climate change goals requires broad deployment of novel process technologies covering a wide range of feed and performance requirements. While traditional process design approaches focus on economies of scale and optimize each installation uniquely, here we focus on the simultaneous design of entire families of processes while exploiting opportunities for shared sub-components. This approach allows for significant reduction in manufacturing and engineering costs, with rapid deployment timelines comparable to modularization.

Bio:

Prof. Laird’s research focuses on large-scale nonlinear and discrete optimization with applications in process and energy systems, manufacturing, homeland security, and largescale infectious disease spread. He is the recipient of several research awards, including INFORMS Computing Society Prize, CAST Division Outstanding Young Researcher Award, National Science Foundation Faculty Early Development (CAREER) Award, and the prestigious Wilkinson Prize for Numerical Software for his work on IPOPT, a software library for solving nonlinear, nonconvex, large-scale continuous optimization problems.