Robust system design
Multi-objective optimization of tightly coupled turbocompressor systems under manufacturing variation.
Methods and resultsPostdoctoral Researcher · ETH Zurich
Soheyl Massoudi
I develop AI methods that generate and evaluate executable engineering models, tools, and optimization code. My work builds on robust design, simulation, and CAD for coupled physical systems.
Physical systemsrobust optimization of coupled turbocompressor systems
Executable AIengineering models, tool-using agents, and reusable optimization code
Open researchpublic software, datasets, checkpoints, and experimental traces
Research
Research areas
My work has developed through three related areas: robust mechanical design, software for simulation and CAD, and AI methods for engineering workflows.
Learning-enabled tools
Ensemble surrogates, real-time simulation, and parametric CAD that connect optimization results to inspectable geometry.
Tools and artifactsAI for engineering design
Executable models, tool-using agents, engineering benchmarks, and reinforcement learning for reusable solvers.
Current researchProjects
Selected projects
Reinforcement learning for reusable solver synthesis
A code model trained to generate standalone optimization programs rather than solve each instance through repeated sampling. The study includes compile-once evaluation, semantic analysis of generated code, and public training and evaluation artifacts.
Project details
Agentic LLMs for conceptual systems engineering
A structured comparison of nine-role and two-agent systems for requirements extraction, functional decomposition, and executable physics-model generation. The study reports both gains in design granularity and unresolved limitations in requirement coverage and code fidelity.
Project details
Robust, all-at-once system optimization
A family of methods for optimizing interdependent compressor, rotor, and gas-bearing subsystems while treating efficiency, operating range, constraints, and manufacturing robustness together.
Project detailsContact
Contact and academic profiles
For research collaborations, technical discussions, invited talks, or student projects, email is the best way to reach me.