Agenda

Data-Centric Engineering for Decarbonisation

Authored & Presented By
Adam Sobey (University Southampton)

The Evolving AI Ecosystem and its Effect on Business Adoption

Authored & Presented By
Jonathan Smith (STFC Hartree Centre)

We will introduce the role of STFC Hartree Centre, part of the UK Research and Innovation family, in helping to unlock the potential of advanced digital technologies to create real-world impact. By equipping UK industry with the necessary knowledge, skills and compute, we help derisk business adoption of AI, big data, supercomputing, quantum computing and cloud technologies. To demonstrate the translation of AI research into business adoption, we will go through a few examples from the work we have done. One of which is the use of generative adversarial networks (GANs), originally designed for marketing and media, for quality engineering and maintenance, and even accelerated fluid dynamics for turbulence control. We contrast with other object detection and classification approaches. Taking a deeper dive into how gen AI and “classical” machine learning are being used in consort, we also explore enhancements in the process of novel material engineering. In this example, we will look at Metal Organic Frameworks (MOFs) for carbon capture. Chemical engineering is also being transformed by automation. We will explain how we use AI to generate phase diagrams, enabling the prediction of phase behavior based on the chemical structures of surfactants, which are widely used in formulations within the fast-moving consumer goods industry. This advancement reduces the time and resources required to develop new products. We will also discuss our fusion computing lab project with UK Atomic Energy Authority (UKAEA) and various partners, which acts as a test bed for the development of emerging AI approaches which will bring new design and control capabilities, which will be introduced. We also discuss briefly the different facets of the AI landscape, including a variety of access programmes, a summary of infrastructure, and recent policy positions.

Opportunities for Generative AI in Design, Simulation and Optimisation of Engineering Systems

Authored & Presented By
Andrew Duncan (Imperial College London)

Generative AI is revolutionizing the design, simulation, and optimization of engineering systems by enabling faster, more efficient, and highly innovative solutions. We describe an increasingly common workflow for the use of conditional and unconditional generative models in the context of challenging engineering tasks, such as inverse design and optimal sensor placement. To be effective within the context of complex engineering challenges, several technical challenges must be overcome, including the ability to accurately characterise physical states of the system; being effective at exploring the design space; and supporting rapid exploration of the design space for optimisation subject to constraints.

Using AI Based Generative Design Systems to Trade off Weight Waste and Complex Performance Criteria

Authored & Presented By
Mark Price (Queen's University of Belfast)

Following an artificial evolution paradigm the Re-Imagining Engineering Design (RIED) Programme is developing new design methods and systems to enable superfast design of complex engineering problems. The solution engine called the “RIED Engine” provides concept solutions which can account for manufacturing constraints as well as complex performance constraints such as buckling. The ethos in RIED is to develop key responses to a given state of stress or temperature, for example, and then evolve large populations of these over many generations. The responses are stored as a neural network which is an analogy for gene regulation in nature. The neural network effectively contains the instructions to generate the final shape and form of the design. An engineering tool chain can then be used to generate simulation or manufacturing models as needed for evaluation or can even in some circumstances provide a detailed model-based definition for manufacture.

Physics-Informed Machine Learning for Wind Energy Applications

Authored & Presented By
Jincheng Zhang (University of Warwick)

Wake interactions in wind farms significantly impact power production and structural loads on wind turbines. Current numerical tools for wake prediction mainly fall into two categories: computational fluid dynamics (CFD) models, which are accurate but computationally expensive, and analytical wake models, which are fast but lack accuracy. Despite their limitations, analytical wake models remain the primary tool in practical applications. To bridge the gap between accuracy and computational efficiency, the first part of this work focuses on developing machine learning (ML)-based wake models capable of real-time evaluation, capturing high-fidelity flow features, generalizing well to unseen flow scenarios, and scaling effectively for large wind farms. The proposed ML-based framework consists of static and dynamic wake modeling approaches. Static wake modeling frameworks developed include the use of dimensionality reduction combined with neural networks, deep convolutional neural networks (CNNs) augmented with generative adversarial networks (GANs), and a multi-fidelity modeling approach based on a novel super-fidelity network. These methods ensure accurate wake representation with reduced computational costs. On the other hand, dynamic wake modelling frameworks developed include the use of dimensionality reduction alongside sequential prediction techniques, as well as a Bilateral Convolutional Neural Network (BiCNN) to capture the temporal evolution of wake structures efficiently. The second part of this work focuses on digital twining of wind farm flows based on physics-informed machine learning approaches. Existing tools in wind energy can only provide wind measurements at sparse locations, while low-fidelity analytical models and high-fidelity CFD models often focus on standalone simulation of wind farms. These tools often lack the capability to integrate both physical principles and real-time measurement data. To address this limitation, this work introduces a digital twin of the wind farm flow system, a data-driven, physics-informed ML model capable of predicting the unsteady flow field in front of a single wind turbine. It is then extended to flow fields across the wind farms capturing wake interactions. The developed digital twin integrates data from LIDAR and turbine sensors with physics-based constraints derived from NS equations and actuator modelling f turbine rotors, forming the first system capable of in situ spatiotemporal wind farm flow field prediction. Case studies are conducted under different operational scenarios and the results show that the developed digital twin can achieve accurate predictions of the dynamic flow fields, bringing brand new opportunities for wind farm control with full awareness of its wind environment. For future works, a Physics-Informed Sequential Deep Operator Network (PI-SDeepONet) is under development, eliminating the need for retraining.

Leveraging Scientific-ML for Additive Manufacturing Process Simulation

Authored & Presented By
Ajit Panesar (Imperial College London)

This paper presents a physics-informed neural network (PINN)-based solution framework that predicts the thermal history during a multi-layer Directed Energy Deposition (DED) process. The meshless nature and the readily available derivative information of PINN solution opens up new opportunities for modelling the thermally induced distortion in metal Additive Manufacturing (MAM). The proposed framework incorporates simple yet effective strategies that enable PINN to overcome the usual shortfall of neural networks (NNs) in dealing with discontinuities, critical for applying PINN to the multi-layer problem due to DED’s layer-by-layer nature. The accuracy of the proposed framework is validated via a benchmark test against ANSYS simulation. Leveraging the possibility of initialisation with prior knowledge, PINN is also demonstrating potential computational time-savings, especially for larger parts. The proposed framework sets the foundation for the subsequent exploration of applying scientific machine learning (SciML) techniques to real-life engineering applications.

Use of Artificial Intelligence Techniques in Turbomachinery Aero-thermodynamics

Authored & Presented By
Shahrokh Shahpar (Rolls-Royce Group PLC)

The use of Artificial Intelligence (AI) in engineering is expanding rapidly, with both in-house and commercial solutions now offering AI-assisted simulation (CFD and FE). However, its technology readiness level (TRL) is currently low and its application across aerospace industry needs better focus and coordination. Although great strides have been achieved with the use of Large Language Models (LLMs), the focus of this presentation is on the following areas: 1) Improving the accuracy and speed of the CFD simulations 2) Improving the performance, accelerating the design cycle (by analysis) & reducing lead-time 3) Addressing the manufacturing variation issues related to the engine (e.g. test bed) performance 4) More-time on the wing – end of life, residual life prediction and prognosis

Harnessing Machine Learning Techniques for Multi Modal Damage Detection

Authored & Presented By
Rob Fleming (Airbus Operations Ltd)

This study focuses on Damage Characterisation using Acoustic Emission (AE) data and Unsupervised Machine Learning. AE signals, generated by material response, carry valuable information about damage mechanisms. By leveraging unsupervised machine learning, we classify AE events based on signal features, enabling the identification of different damage types and severities. This unsupervised approach provides a data-driven method for structural health monitoring (SHM), enhancing the understanding of failure modes and improving predictive maintenance strategies. Results demonstrate the effectiveness of clustering in distinguishing damage characteristics without prior labelling, making it a robust tool for real-time assessment.

AI Potential for Aerodynamic Wheel Drag Predictions and Optimisation

Authored & Presented By
Santosh Jayaraju (Jaguar Land Rover)

Utilising AI and ML in the Design Process within Aerospace

Authored & Presented By
Tim Newman (National Composites Centre)

One over review is a vital part of the engineering design process, but it is also costly. Senior engineers checking the work of more junior engineers against a complex set of rules is time consuming. The aim of SmartDFM was to use a feature identification and rule application engine to conduct a first pass for mistakes on CAD models. It is built on Graph Neural Networks (GNN) and expert systems. This process included automatically generating an aerostructures inspired data set, training machine learning models, rules capture using expert interviews, and a working add-in for CATIA.