Prof. Matteo Parsani
Associate Professor of Applied Mathematics and Computational Science, and Mechanical Engineering Lead, Advanced Algorithm and Numerical Simulations Laboratory (AANSLab) King Abdullah University of Science and Technology (KAUST)Abstract
From Supercomputing to AI Surrogate Models: Accelerating A2 Engineering with Shaheen III
SonicRiyadh brings Shaheen III’s GPU partition to a challenge central to the future of advanced and urban air mobility: understanding and reducing aircraft noise exposure in cities. In this ongoing grand challenge, our KAUST-developed ASTRA solver has run across all 700 GPU nodes, using 2,800 NVIDIA Grace Hopper GPUs. Accounting for Riyadh’s buildings and changing local weather, the simulations reveal how the urban environment shapes sound propagation and noise exposure. The resulting acoustic maps support comparisons of flight routes, vertiport locations, and approach, departure and overflight scenarios, helping planners assess impacts on residential communities and identify quieter operations. Managing noise is crucial to public acceptance. NASA and ESA initiatives, alongside massive international and in-Kingdom commercial interest, highlight the global significance of this challenge. SonicRiyadh’s simulations provide a foundation for training acoustic prediction models aimed at near-real-time inference for operators, planners and decision-makers.
This achievement builds on years of algorithm development at KAUST’s Advanced Algorithm and Numerical Simulations Laboratory (AANSLab). SSDC, one of the few application codes used to evaluate and select Shaheen III’s CPU system, has now been ported to GPUs. It scales across hundreds of thousands of CPU cores. Its robust, variable-order methods preserve essential physical principles while controlling numerical instabilities. This focus on reliable, scalable simulation aligns with NASA’s CFD Vision 2030 and commercial developments at Ansys, Siemens and Cadence. Applications span subsonic and supersonic aircraft aerodynamics, with predictions compared against measurements, plasma propulsion and hypersonic reentry flows at ten times the speed of sound.
This foundation gives scientific machine learning its engineering value. Detailed simulations provide training data and a reference for assessing predictions of forces, heating and noise. Our novel neural operator framework learns across changing geometries and supplies improved starting conditions for computational fluid dynamics solvers. In tests involving industrial airfoils, wings and aerodynamic shape optimization, it reduced solver cost per simulation by approximately 43–44% on average. Applications include optimization of aircraft sections and automotive parts. Savings accumulate across industrial workflows requiring hundreds or thousands of simulations for design optimization, databases, uncertainty quantification, digital twins, and certification. These capabilities help engineers explore broader design spaces and identify more efficient designs.
Current work includes aerodynamic surrogates under development with Ceer Motors and rapid predictions of very high-speed flow fields. Neural operators support industrial shape optimization and extreme fluid-structure interaction conditions, while reinforcement learning guides where simulations need finer resolution. We also combine fluid-flow and acoustic simulations with machine learning for flying-vehicle detection. Surrogate models for autonomous flying vehicles aim to support robust, safe flight in harsh conditions and during communication loss. These applications address a recurring engineering need: predicting physical behavior accurately and efficiently as geometry, operating conditions or control inputs change.
ASTRA marks the beginning of sustained use of Shaheen III’s GPU partition for large simulation campaigns and future massive multi-GPU training. Building on existing models and software, we aim to scale training for more complex scenarios and operating conditions. The talk connects Shaheen III’s CPU heritage with scientific AI that supports practical decisions in Saudi aerospace and automotive (A2) engineering.
Biography
Matteo Parsani is Associate Professor of Applied Mathematics and Computational Science and Mechanical Engineering at KAUST, where he leads the Advanced Algorithm and Numerical Simulations Laboratory (AANSLab). Previously a postdoctoral fellow and later a research scientist at NASA Langley Research Center, he develops novel algorithms combining mathematical and physical rigor with high-performance computing and, more recently, scientific AI.
A key research focus is developing robust, variable-order, entropy-stable and structure-preserving numerical methods, translating stability and accuracy analyses into scalable software for modern computing systems. This work advances numerical capabilities central to NASA’s CFD Vision 2030 and beyond.
His scientific machine learning (SciML) research spans reinforcement learning for adaptive mesh refinement, neural operators for shape optimization, and surrogate models for high-speed and very high-speed flows. He also combines computational fluid dynamics and aeroacoustics with machine learning for flying-vehicle recognition and detection. Many of these models are trained across a large number of GPUs. With Ceer Motors, his team is developing SciML surrogates for rapid aerodynamic design.
His group’s SonicRiyadh project used the KAUST-developed ASTRA solver on 2,800 NVIDIA Grace Hopper GPUs on Shaheen III to predict advanced air mobility noise propagation through Riyadh under realistic micro-weather conditions. His collaborations include Boeing, Airbus, McLaren and GADD, Ceer Motors, and MIM. He also works to translate advanced simulation and AI into Saudi industrial engineering capabilities.
Beyond academia, Professor Matteo is a former soccer player and activist for people with disabilities. In December 2023, he embarked on an unprecedented coast-to-coast hand-bike journey across the Kingdom of Saudi Arabia to promote physical activities, sports, and awareness about people with disabilities. He covered nearly 150 kilometers daily on a hand-bike, totaling over 3,100 kilometers in about 22 days.
Attend this session
The Saudi HPC/AI Conference runs 26–28 October 2026 at KAUST. Attendance is free, but registration is required.
