Abstract:
Computational Fluid Dynamics (CFD) is a powerful tool for understanding complex engineering systems, but its computational cost often limits real-time prediction, optimization, and control. At the same time, modern facilities generate large amounts of data through sensors and monitoring systems. The challenge is therefore not to choose between physics-based modelling and artificial intelligence, but to combine them effectively. This lecture presents a hybrid approach in which high-fidelity CFD provides detailed physical knowledge, while AI-based surrogate models reproduce the most computationally expensive parts of the solution at a fraction of the cost. Physical transport, conservation, and process equations are retained to ensure consistency and robustness. Examples from wastewater treatment and environmental fluid mechanics will show how these models can be connected to real-time data, transforming CFD from an offline analysis tool into a dynamic digital twin for prediction, optimization, and engineering decision-making. This enables faster, safer, and more sustainable industrial operation.

Abstract:
Turning kinematic theory into a mechanism that moves as intended is a hard step for students and early-career engineers. Graphical methods analyze a mechanism one position at a time, and analytical methods need lengthy vector-loop equations to cover the whole cycle. Dynamic effects and forces also decide the choice of joints, bearings and supports. This talk shows how interactive design tools can support these decisions for the kinematics and dynamics of linkages and cams: where a mechanism binds, how large its joint forces are, and how small a cam can be. The examples use LinkLab and CamLab (mechsim.dev), free, open-source tools I developed with AI coding assistants, which support the mechanical design workflow and serve as teaching aids for Mechanics of Machinery.

Abstract:
Many problems in engineering science share a difficult feature: the governing physics is only partially known, and experimental data is scarce or noisy. Classical simulation methods such as CFD are mature and accurate but require a complete description of the physics and are computationally demanding, while purely data-driven machine learning is data-hungry and yields non-physical predictions. Physics-Informed Neural Networks (PINNs) bridge this gap by embedding physical laws – expressed as partial differential equations – directly into the training of a neural network, making them well suited to the ubiquitous regime of limited physics and limited data.
This talk gives a self-contained introduction to PINNs, from the underlying principles to their practical implementation, and illustrates their power on two engineering applications: recovering three-dimensional gravity currents from sparse experimental data and identifying rheological parameters in granular flows. Together these examples show how PINNs offer a robust, GPU-efficient alternative to classical data assimilation.