Extreme-Scale Training for USC Researchers
Four researchers from the USC Viterbi School of Engineering recently completed the Argonne Training Program on Extreme-Scale Computing. This program selects 75 participants from a global applicant pool to undergo two weeks of intensive instruction regarding software, algorithms, and computing hardware. The curriculum focuses on preparing scientists for the demands of modern supercomputers, which now handle vast workloads involving hundreds of millions of cells or entire machine-wide simulations.
Participants from USC included doctoral students Benran Zhang, Aishwarya Krishnan, and Ryan Zapp, alongside postdoctoral researcher Nitish Baradwaj. They spent the time studying GPU architectures, performance profiling, and numerical methods. As supercomputing machines like Frontier and Aurora continue to push the boundaries of what is possible, the ability to translate scientific questions into efficient code remains a critical skill for the next generation of engineers.
Solving Atomic and Material Challenges
Benran Zhang works within the Mork Family Department of Chemical Engineering and Materials Science under the guidance of assistant professor Zhenglu Li. His research focuses on first-principles methods to identify excited-state properties of materials. By solving quantum-mechanical equations, Zhang avoids empirical parameters, though these computations require immense hardware resources. He often utilizes tens of thousands of GPUs simultaneously to run jobs on national laboratory systems.
Nitish Baradwaj, also a researcher at the USC Mork department, approaches material science through machine-learned interatomic potentials. By training models on high-accuracy reference data, his team can simulate millions of atoms over longer timeframes than traditional quantum mechanics allow. This method captures rare events like defect formation or chemical reactions that were previously out of reach. Baradwaj notes that his training at Argonne emphasized the necessity of measuring code performance before attempting to optimize it.
Managing Hypersonic Reentry Physics
Aishwarya Krishnan studies the physics of space capsules during hypersonic reentry. Working in the USC Computational Aerospace Lab led by associate professor Ivan Bermejo-Moreno, she models turbulence on heat shields. Because resolving every scale of turbulence is computationally prohibitive, Krishnan develops predictive models that capture necessary physics with greater efficiency. Her simulations often involve hundreds of millions of computational cells, requiring careful management of resources and precision.
Ryan Zapp works in the same lab as Krishnan, focusing on how the geometry of a capsule influences aerodynamic heating. His simulations can run for one month of continuous wall-clock time on a supercomputer. During the training program, Zapp observed a shift in hardware design. Much of the new high-performance hardware emphasizes AI workloads rather than the high-precision scientific computing required by his specific field. Adapting these algorithms to AI-optimized hardware remains an open challenge for his cohort.
Navigating the Future of AI and Science
The integration of artificial intelligence into scientific workflows creates new opportunities and risks. Researchers like Krishnan warn that while AI can generate code or suggest workflows, human oversight is mandatory. A researcher must maintain enough domain expertise to recognize when an automated output fails a physical or numerical check. Without that grounding, the risk of deploying incorrect scientific software increases.
Baradwaj echoes this sentiment regarding machine-learned potentials. He explains that a model might perform perfectly on training data but falter when it encounters conditions outside that set. The path forward for these four researchers involves maintaining high standards for uncertainty quantification and validation as they return to USC. They are tasked with navigating a landscape where hardware changes rapidly and the pressure to produce reliable, defensible scientific results grows alongside the available computing power.