Skip to main content
Prof Ryno Laubscher delivering his inaugural lecture
Image by: Ignus Dreyer

Prof Ryno Laubscher from the Department of Mechanical and Mechatronic Engineering delivers his inaugural lecture.

Engineering and technology

Prof Ryno Laubscher explores deep learning for thermal-fluid systems

Alec Basson
Corporate Communication and Marketing
06 October 2025
  • Prof Ryno Laubscher from the Department of Mechanical and Mechatronic Engineering delivered his inaugural lecture on 2 October 2025
  • He harnesses deep learning to help optimise energy system components
  • His research is relevant to coal-fired power plant reheaters and air-cooled condensers

Prof Ryno Laubscher from the Department of Mechanical and Mechatronic Engineering in the Faculty of Engineering at Stellenbosch University delivered his inaugural lecture on Thursday 2 October 2025. The title of his lecture was “Harnessing deep learning for thermal-fluid systems”.

Laubscher spoke to the Corporate Communication and Marketing Division about how he harnesses deep learning in his research to help address the challenge of optimising energy system components.

Tell us more about your research and why you became interested in this specific field.

In a subset of my research, I combine deep learning (DL) with thermal-fluid systems (combustion, power cycles, turbomachinery) to create faster, more accurate predictive models. Thermal-fluid systems are systems where fluids carry and transfer heat, such as in heating, cooling, or energy processes. 

I became interested because traditional computational fluid dynamics (CFD) simulations (using computers to model and predict how liquids and gases move and behave) are computationally expensive, while machine learning offers near-instant predictions once ‘trained’These can be leveraged in areas such as design space exploration (the process of testing different design options to find the ones that best meet the requirements, covering both hardware and software) and optimisation problems.

How would you describe the relevance of your work?

Artificial intelligence (AI) surrogate models can replace computationally expensive CFD simulations with near-instantaneous predictions, significantly accelerating design optimisation processes. Through generative learning approaches, these models enable comprehensive design space exploration, discovering geometries (the shapes and forms of objects) and configurations that extend beyond conventional engineering approaches.

Can you give examples of how your research is applied in real-world contexts?

Some of our completed deep learning projects are: forecasting how hot the metal in coal-fired power plant reheaters will get to prevent tube failures during load cycling; predicting the pressure in air-cooled condensers while accounting for uncertainty, to optimise power generation; and developing fast surrogate models that replace expensive CFD simulations for devices that control the burning of methane, a greenhouse gas.

What are some of the limitations of data-driven approaches when applied to thermal-fluid problems?

These models require large datasets and don’t function well for situations beyond the data they were trained on. Models can break basic physical rules if the provided simulation or measurement data isincomplete or inaccurate. 

How can the limitations of data-driven approaches be addressed?

Two exciting avenues to address the above limitations are scientific deep learning and universal ordinary differential equations (UODEs). 

Scientific deep learning embeds physical laws (conservation equations) into neural network training, reducing data requirements while ensuring physical consistency. UODE solvers use neural networks as learnable components within established physical models, while keeping the known physics accurate. Although these approaches address some of the limitations of purely data-driven techniques, they themselves are not without challenges.

Looking into your crystal ball, what developments do you see in your field?

I believe geometric deep learning— an approach that allows AI to handle data with more complex shapes and connections like networks or 3D surfaces by making use of the built-in patterns and structures in the data — will have a noticeable effect on thermal-fluid modelling (creating computer models to understand and predict how fluids move and transfer heat within a system). It will do this by processing unstructured CFD mesh data directly, eliminating geometric parameterisation constraints and enabling truly geometry-agnostic surrogate models. 

Additionally, geometric DL can also be combined with scientific DL, further unlocking potential advantages in building surrogate models.

In future, I hope to develop networks of 3D geometric DL component models which interconnect to simulate entire thermal-fluid systems like gas turbines, enabling system-level design optimisation across multiple components simultaneously.

The higher education environment can be challenging. What keeps you motivated when things get tough?

I genuinely love my research and teaching work, which reenergises me. My faith reinforces this by helping me stay focused on making meaningful contributions that serve both society and a higher purpose.

Tell us something exciting about yourself that people would not expect.

I really enjoy competing in Brazilian jiu-jitsu.

How do you spend your free time?

Sleeping.

Related stories