Plenary Lecture 1
Title
Overcoming the AI Data Bottleneck in Turbomachinery: Physics-Enhanced Machine Learning via 3D Inverse Design
Lecturer
Mehrdad Zangeneh,
Department of Mechanical Engineering,
University college London
Mehrdad Zangeneh is Professor of Thermofluids at University College London and Founding Director of Advanced Design Technology, Ltd. For the past 30 years he has been involved in development of advanced turbomachinery design codes based on 3D inverse design approach and automatic optimization. His research has resulted in important breakthroughs in radial turbomachinery, such as the suppression of secondary flows in radial and mixed flow impellers and the suppression of corner separation in vaned-diffusers. He is recipient of Japan’s Turbomachinery Society’s Gold Medal, the IMechE Donald Julius Grone Prize and the ASME Henry Worthington medal.
Summary
The integration of machine learning (ML) and data-driven surrogate models into turbomachinery design optimization promises near-instantaneous performance evaluation. However, traditional workflows face a critical "data bottleneck": training high-accuracy deep learning models typically requires thousands of expensive CFD simulations. This computational penalty is severely worsened by conventional CAD-based or geometric parameterizations, which move vertices blindly and generate physically impractical shapes, wasting massive compute resources. Furthermore, pure "black-box" ML models frequently fail to generalize when predicting highly non-linear flow phenomena, such as transonic shocks, boundary layer separation, and complex secondary flows.
This lecture presents a paradigm shift that overcomes this data bottleneck by utilizing 3D Inverse Design as a foundational, physics-guaranteed filter for machine learning. Rather than manipulating raw geometric coordinates, the inverse method uses aerodynamic design parameters—specifically, blade loading and circulation distributions—to directly generate the 3D geometry. Because the underlying mathematical framework ensures that every generated shape satisfies specified work and mass flow targets, the design space is constrained entirely to physically valid configurations from the outset.
This "physics-clean" data generation approach yields a 10x to 100x increase in data efficiency, enabling the training of high-accuracy machine learning surrogates . The efficiency of the process can be further enhanced through improvements in machine learning algorithm, such as the Reactive Response Surface (RRS). The lecture will demonstrate the efficacy of this method through practical multi-point optimization cases of high-performance centrifugal compressors, Francis turbine rotor and axial fan. Finally, we will explore how this physics-grounded approach can be used to create high fidelity machine learning expert systems that can be used under industrial conditions by using an example of a mixed flow pump stage.
Plenary Lecture 2
Title
Challenge of Multi-Physics CFD Simulation in Jet Engines
Lecturer
Makoto Yamamoto,
Physics Laboratory for Mechanical Engineering
Waseda University
Dr. Makoto Yamamoto completed the doctoral coursework at the University of Tokyo, Graduate School of Engineering in 1987 and received his Doctor of Engineering degree in 1988. From 1987 to 1990, he worked in the aerodynamic design and development of jet engines at Ishikawajima-Harima Heavy Industries Co., Ltd (now IHI). From 1990 to 2026, he worked at the Faculty of Engineering, Tokyo University of Science. He became a professor in 2004 and served as vice president from 2014 to 2018. He has been a professor at Waseda University since April 2026. During this time, he served as president of the Japan Society of Mechanical Engineers, the Gas Turbine Society of Japan, the Japan Society of Fluid Mechanics, and the International Association for Exchange of Students for Technical Experience. In April 2026, he became an Honorary Member of the Japan Society of Mechanical Engineers.
Summary
Jet engines operate while ingesting large amounts of tiny solid particles, liquid droplets, and ice chunks suspended in the atmosphere. These tiny particles cause various multi-physics phenomena inside the jet engines, such as erosion, deposition, and icing, posing a serious risk to the aerodynamic performance, safety, maintenance, and lifespan of the jet engine. The present speech will introduce the numerical methods and representative computational results of multi-physics CFD simulations, using deposition and icing as examples.
Jet engines are frequently exposed to harsh environments containing dust particles such as sand or volcanic ash, which can be solved in the combustion chamber due to the high temperature, and adhere to turbine vane surfaces or internal cooling channels. This phenomenon is referred to as “deposition”. The deposition phenomenon can lead to adverse effects on the engine performance, including aerodynamic degradation and blockage of cooling holes.
Numerous supercooled droplets and/or ice crystals exist in a cloud. When an aircraft passes through a cloud, they impinge on the aircraft wing and fuselage, and also they enter into the jet engines. Such impinging droplets and ice crystals can form ice layers on the surfaces. This phenomenon is referred to as “icing”. Apparently, the icing adversely affects the performance of an aircraft by reducing the lift and thrust, and it may cause a crash. Four types of icing are important in engineering: rime icing, glaze icing, supercooled large droplet (SLD) icing, and ice crystal icing.
In our simulation code, the Euler-Lagrange method is employed, assuming one-way coupling. The temporal progression of the phenomenon is simulated by performing multi-shot computations, which involve repeating a series of computational processes: flow field computation, trajectory computation of minute particles, thermodynamic computation, and grid regeneration. In the flow field computation, the turbulent flow is modeled by the Reynolds-Averaged Navier-Stokes equations (RANS) because of the short computational time.
I hope this speech will provide useful insights for future research and development of the participants.