The Argonne Training Program on Extreme-Scale Computing (ATPESC) brings researchers together for two weeks of intensive training in the software, algorithms and computing systems used for large-scale scientific applications. This year, 75 researchers were selected from a worldwide pool of applicants, including 4 from USC Viterbi School of Engineering. The curriculum covered GPU architectures, parallel programming, performance profiling, numerical methods and AI-enabled scientific workflows.
Calculating material properties #
Benran Zhang, a doctoral student at USC Viterbi’s Mork Family Department of Chemical Engineering and Materials Science, uses first-principles methods to study excited-state properties of materials, including interactions between strongly bounded electrons and holes, and lattice vibrations. His calculations solve many-body quantum-mechanical equations without empirical parameters.
Led by Zhenglu Li, assistant professor of chemical engineering and materials science, Zhang is part of a research group that relies on extreme-scale computing codes to calculate material properties. ATPESC gave him a structured introduction to areas that he had encountered but had yet to investigate in depth, including GPU architectures, programming models, numerical software and scalable solvers.
“Routine calculations in my research can use up to tens of thousands of GPUs,” Zhang said. “The largest computations I do in my research involve running on the entire Frontier machine at the Oak Ridge Leadership Computing Facility and the Aurora machine at the Argonne Leadership Computing Facility.”
Measurement for optimization #
Nitish Baradwaj, a postdoctoral researcher at USC Mork, studies how atomic motion determines the properties of complex materials. As a member of the USC Collaboratory for Advanced Computing and Simulations, he uses machine-learned interatomic potentials to approach quantum-mechanical accuracy while extending simulations to larger systems and longer timescales.
Direct quantum-mechanical calculations provide accurate reference data but are generally limited to relatively small systems and short simulation times. Machine-learned potentials trained on that data can be deployed across many accelerators to simulate millions of atoms. These simulations can capture rare chemical reactions, defect formation, transport and evolving interfaces.
“The exciting part of working with supercomputers isn’t only about being able to run larger simulations,” he said “It’s about being able to ask a scientific question that was previously computationally inaccessible,” Baradwaj said.
At ATPESC, Baradwaj learned to profile a program before deciding what steps to take next. Memory access, communication between GPUs, load imbalance or input/output can constrain a calculation that initially appears limited by arithmetic. “One practical lesson that has stayed with me is to measure before optimizing,” he said.
Computation for hypersonics #
Aishwarya Krishnan, a doctoral student in USC Viterbi’s Department of Aerospace and Mechanical Engineering (AME), uses computational fluid dynamics to study how turbulence develops as a space capsule’s heat shield erodes during hypersonic reentry to Earth’s atmosphere. This process can significantly increase surface heating, making accurate prediction important for space vehicle thermal protection systems.
As part of the USC Computational Aerospace Lab led by Ivan Bermejo-Moreno, associate professor of aerospace and mechanical engineering, Krishnan develops models that capture the relevant physics without the computational expense of resolving every scale of turbulence. Her simulations can involve hundreds of millions of computational cells.
“Some of my biggest technical takeaways at ATPESC involved mixed-precision computing, performance profiling, and GPU programming,” she said. “The GPU sessions gave me a clearer understanding of how scientific software must adapt as processor architectures evolve.”
Ryan Zapp, a doctoral student working alongside Krishnan at the Computational Aerospace Lab, studies how the shape of a space capsule affects aerodynamic heating during hypersonic reentry. Using the lab’s uPDE software, a single simulation may require up to one month of continuous supercomputer time.
At ATPESC, Zapp focused on how AI demand is influencing computer hardware. “Next generation hardware is increasingly being designed with AI workloads in mind rather than traditional extreme-scale, high-precision scientific computing,” Zapp said. “We have access to unprecedented computing power, but we need to find effective ways to adapt our computer algorithms to hardware that was not necessarily designed for them.”
More knowledge means more questions #
Researchers are also using AI to generate and refactor scientific code and automate parts of computational workflows. Krishnan emphasized that researchers still need to verify the numerical and physical validity of the resulting software.
“AI can generate code or suggest an approach very quickly, but the researcher still needs enough scientific and computational understanding to define the problem correctly, recognize when an answer is wrong, and verify that the result is physically and numerically sound,” she said.
Machine-learned interatomic potentials introduce another verification problem. Baradwaj noted that a model may perform well on chemistry and structures represented in its training data, but fail when it encounters unfamiliar conditions. He identifies uncertainty quantification, physical validation and reproducibility as necessary for developing scalable, reliable models.
Back at USC, the four researchers can apply that training to problems already demanding extreme-scale computation, whether first-principles materials calculations, or machine-learned molecular dynamics, or hypersonic-flow simulations. As the hardware changes, they will be in a position to take advantage of new computing systems, while determining which methods remain accurate, efficient and scientifically defensible.
Published on September 3rd, 2026
Last updated on September 3rd, 2026