The National Science Foundation renewed the Artificial Intelligence Institute for Advances in Optimization (AI4OPT) with a $20 million grant, extending the institute's work through 2031. The funding supports a second phase of research that fuses AI with mathematical optimization to tackle real-world problems in energy systems, supply chains, and manufacturing at societal scale.
Georgia Tech remains the lead institution, with professor Pascal Van Hentenryck continuing as lead principal investigator. The University of Southern California and the University of California, Berkeley serve as partner institutions. USC's role expanded significantly in this phase - its funding more than doubled to $5.6 million, and its research team grew to five faculty members.
"This renewal allows us to dig deeper into AI and optimization, moving basic science toward real-world applications that make a tangible difference in society," said Bistra Dilkina, co-principal investigator and associate professor of computer science and industrial and systems engineering at USC. Dilkina leads the USC team and directs the institute's AI for Optimization research thrust.
Four research thrusts target applied and foundational work
The institute's second phase organizes work into four integrated areas. Two thrusts advance the underlying science of AI and optimization. Two thrusts apply those advances to specific domains.
Dilkina's AI for Optimization thrust develops "learning to optimize" approaches and foundational models for discrete optimization. The goal: decision-making systems that respond to changing conditions in milliseconds, whether adapting power grids or managing complex logistics. The Optimization for AI thrust, led by Berkeley's Alper Atamtürk, creates new optimization techniques to make AI training more scalable, trustworthy, and energy efficient.
The application-focused thrusts sit at Georgia Tech. The Energy Systems thrust, led by Chelsea White and Daniel Molzahn, addresses growing complexity in electrical grids as renewable energy and data center demand increase. The Supply Chains and Manufacturing thrust develops agile systems for logistics, production, and semiconductor manufacturing that withstand disruptions from disasters, pandemics, and other global shocks.
USC's interdisciplinary team grows
The expanded USC team spans the Viterbi School of Engineering, the Stevens School of Computing and AI, and the Marshall School of Business. Computer science professor Willie Neiswanger focuses on uncertainty modeling within the AI for Optimization thrust. Industrial and systems engineering professor Andres Gomez serves as USC lead for Optimization for AI. Karmel Shehadeh, also in industrial and systems engineering, leads research on agile logistics and drone deployment within supply chain applications. Vishal Gupta, a data sciences and operations professor, works on methodologies at the intersection of AI and optimization and their supply chain applications.
USC will also host the Seth Bonder summer camp in computational and data science for engineering, extending the institute's educational reach to high school students in Los Angeles.
Building on the first phase
During its initial five years, AI4OPT produced more than 250 publications, completed 12 technology transfers, and released new datasets including Distributional MIPLIB. At USC, the institute helped launch the first PhD specialization and certificate in AI plus Optimization, supported by a National Science Foundation Research Traineeship grant.
For researchers working at the boundary of AI and operations research, the renewal signals sustained federal investment in optimization as a complement to machine learning - not a field left behind by the deep learning boom.
Why this matters for science and research professionals
The institute's explicit focus on "learning to optimize" - teaching machines to solve complex optimization problems faster than traditional methods allow - represents a shift in how AI and operations research intersect. Instead of treating optimization as a separate step after prediction, the research integrates the two. For scientists and engineers working on resource allocation, scheduling, or network design, the techniques emerging from this work could reduce solution times from hours to milliseconds, enabling real-time decision systems that current methods cannot support.
Your membership also unlocks:
