Advancing Finite Element Analysis with Interpretable AI: Innovator Spotlight on the British University in Egypt

We recently sat down with Sherif Samy, a distinguished PhD researcher from The British University in Egypt (BUE), to delve into his work using Rescale to explore new AI methods. Samy’s work, which was recently published in Nature, focuses on applying interpretable machine learning to solve complex challenges in engineering design, particularly with composite materials. He highlighted BUE’s unique engineering program, industry impact, and Rescale’s role in accelerating his advanced simulations and AI model training. This interview offers a glimpse into the growing trend of using AI in materials engineering and how computational tools are shaping the future of engineering research.

To start, could you tell us about The British University in Egypt (BUE) and what makes its approach to education and research unique?

The British University in Egypt (BUE) engineering school is committed to providing a competitive education aimed at developing globally competent, research-driven graduates. Our core mission is to foster individuals who can significantly contribute to the advancement of both research and industry.

Could you elaborate on your research topics, methodologies, and collaborations?

My work is all about tackling real-world structural design puzzles using advanced computational methods. Specifically, I’ve been focused on “interpretable” machine learning for engineering. While some powerful AI models–for example ensemble models which are increasingly popular–can be highly accurate, they can be a bit of a “black box,” meaning they give us great predictions, but they don’t show us how they got there, making it tough to understand the relationships between different design variables.

Explainability is all about making these intelligent algorithms transparent so we can understand the underlying principles and how they arrive at their conclusions. A big part of what I do is bridging the gap between this theoretical knowledge and what’s happening in industry. We team up with industry leaders, using our applied research and data-driven modeling to directly solve the challenges they face. For example, we collaborated with the Arab Organization for Industrialization (AOI) and Misr Commercial Vehicles (MCV) on a comprehensive simulation and design-for-manufacturing study of a 20 kW GFRP wind turbine blade. Our work included high-fidelity finite element analysis under both operational and extreme wind conditions, material characterization, and optimization of composite layups to meet structural requirements and accommodate changes in tower design. This collaborative approach ensures our research not only remains relevant but also directly contributes to real-world engineering and production outcomes.

Tell us about your focus on applying HPC and AI to composite design and why this area is particularly important in the engineering world?

Solving the interpretability challenge is particularly important in materials and parts engineering applications like composite bolted connections, which are rapidly gaining adoption in industries such as aerospace and automotive due to their superior strength-to-weight ratio. However, these connections are complex, and existing comprehensive design standards are often insufficient or non-existent. High-Performance Computing (HPC) and AI are essential for efficiently running complex, high-fidelity simulations, handling compute-intensive demands, many iterations, and varied design scenarios to get accurate predictions in a reasonable timeframe.

Our work aims to overcome this by using symbolic regression, specifically PySR (Python Symbolic Regression), to derive explicit, interpretable equations from high-fidelity simulation data. These equations reveal the underlying physics and governing mechanics of composite structures. This approach not only provides high predictive accuracy but also offers transparency, allowing engineers to understand why a certain design performs the way it does. This interpretability is vital for trust, validation, and ultimately, for developing more robust and reliable designs, especially in safety-critical applications where understanding failure mechanisms is paramount.

How are these materials advancements applied in various industries?

Composite materials and fiber-reinforced polymers (FRPs) are extensively used across a diverse range of industries due to their superior properties. In aerospace, they are crucial for lightweighting, appearing in structural components of aircraft such as the Airbus A350 XWB and Boeing 787 Dreamliner, and in engine parts like those found in LEAP engines, often secured with composite fasteners. The automotive sector leverages composites for enhanced performance and fuel efficiency, integrating them into high-performance vehicles such as the BMW i-series, BMW 7 series, and Audi R8 e-tron for components like body panels and chassis. In the marine industry, FRPs are employed in propeller designs, as seen in projects like “FabHeli,” and for specialized products such as pipe fittings, flanges, nozzle plates, and backing rings. The construction industry benefits from these materials in infrastructure projects, with examples like bridges that use FRP girders integrated with concrete slabs using bolted connections, showcasing their application in modern bridge construction.

Walk us through your approach to AI, particularly how you generate data and validate your models, and how Rescale fits into this process?

Our methodology is comprehensive, integrating experimental testing, finite element modeling (FEM), and machine learning. We begin by fabricating and experimentally testing hybrid L-joints to determine their damage initiation loads. These physical tests are crucial for generating real-world data and, more importantly, for validating our finite element models, primarily developed in Abaqus. This experimental validation ensures the accuracy and reliability of our FEM simulations. The validation process requires extensive simulations to tune penalty stiffness coefficients to capture the delamination initiation and progression accurately, an effort that was significantly accelerated using Rescale’s scalable HPC platform.

Once validated, these FEA models become a powerful tool for generating large, high-quality datasets. We employ a design of experiments (DoE) approach to systematically explore the design space and create a comprehensive dataset. This is where Rescale becomes indispensable once again. Running complex finite element simulations, especially for generating extensive datasets, demands significant computational power and scalability. Rescale allows us to execute these ABAQUS jobs efficiently, scaling our computational needs on demand. The ability to run these simulations from anywhere, coupled with Rescale’s seamless integration with our academic objectives, has dramatically accelerated our data generation process.

After generating the data, we use feature selection to identify the key parameters influencing joint performance. Machine learning models then assess the quality of the dataset and guide the selection of PySR loss function. Finally, we leverage PySR to derive those interpretable design equations from the FEM-generated datasets. Rescale also supports our AI model training, providing the necessary computational backbone for these advanced analyses.

Before adopting Rescale, what were some of the challenges you faced in your research or typical engineering design workflows?

Before Rescale, our workflow faced several significant bottlenecks, primarily related to computational resources and software management. Running the large-scale finite element simulations required for our extensive datasets was incredibly time-consuming on local machines or even smaller clusters. This limited the number of design variations we could explore and significantly prolonged our research cycles. There were also challenges with managing different software licenses, ensuring compatibility across various tools, and dealing with hardware limitations. The setup and configuration for each simulation environment could be cumbersome, leading to valuable time being spent on IT overhead rather than core research. Essentially, the raw computational power and the ease of access to it were major constraints that slowed down our progress and limited the scope of our investigations.

How has Rescale helped you overcome those challenges and improve your research?

Rescale has been a game-changer for us. Its primary impact has been on scalability and efficiency. The platform’s ability to provide on-demand access to high-performance computing resources means we can now run hundreds or even thousands of simulations in parallel, drastically reducing the time it takes to achieve our research objectives, the number of FEM simulations for penalty stiffness tuning, validation, and to generate the training dataset is not less than 500 simulations and for training ML models on the dataset is around 200 trials . What used to take weeks or months can now be completed in days. This incredible acceleration allows us to explore a much wider design space, leading to more robust and generalized interpretable equations.

Beyond raw computational power, Rescale’s intuitive design and integrations with many of our existing software has been invaluable. We no longer spend excessive time on infrastructure setup or troubleshooting software conflicts. The collaborative support from the Rescale team has also been exceptional, providing guidance when needed. The combination of powerful technical capabilities with academic usability, including considerations for licensing compatibility and cost control, has allowed us to focus more on the scientific challenges and less on computational overhead, directly translating into significant efficiency gains and higher research output.

Looking ahead, what trends do you expect to see in material sciences and engineering design, especially related to AI and simulation?

The future of material sciences and engineering design will undoubtedly be shaped by the deeper integration of AI and advanced simulation. We are moving towards a paradigm where AI is not just a tool for analysis but an integral part of the design process itself, enabling generative design and accelerated material discovery. I foresee an increasing emphasis on “digital twins,” where high-fidelity simulations create virtual replicas of physical systems, continuously optimized and updated with real-world data.

The pursuit of interpretable AI, as demonstrated in our research, will also become more critical. As AI models become more complex, understanding their decisions will be paramount, especially in regulated industries. Furthermore, the combination of advanced experimental techniques with data-driven modeling will lead to unprecedented insights into material behavior at various scales. High-performance computing platforms, like Rescale, will be the backbone supporting these trends, enabling the massive simulations and data processing required for these next-generation design and discovery workflows.

What advice would you offer to other academic and higher education researchers who are considering adopting HPC solutions for their work?

My primary advice would be to look for platforms that strike a balance between technical power and academic flexibility. Raw performance is certainly important, but it’s not the only factor. For academic researchers, aspects like licensing compatibility, cost control, and the ease of onboarding are equally crucial. You’ll want a solution that can scale your computational needs without introducing prohibitive costs or steep learning curves for your students and research teams – this is where Rescale excelled for us. Look for platforms that offer robust support, clear documentation, and examples that cater to common academic software and workflows. Investing in learning programming languages is also an essential skill for simulation customization and machine learning model training, as it empowers researchers to truly leverage the full potential of these HPC environments.

Lastly, what are your overall thoughts and feedback on using the Rescale platform for your research?

Overall, my feedback on the Rescale platform is extremely positive. Its intuitive interface and capabilities are well-suited to our research objectives. I particularly appreciate the ability to run simulations from anywhere, which offers immense flexibility for our team. The cooperative support we’ve received from the Rescale team has also been outstanding, providing quick and helpful assistance whenever needed.

If I were to suggest areas for even further improvement, I’d say more guidance on advanced workflows like design of experiments features and command-line tips for different software within Rescale could be beneficial for new users. However, these are minor points in an otherwise excellent experience. Rescale has proven to be an invaluable asset in accelerating our research, enabling us to push the boundaries of interpretable machine learning in engineering design, and ultimately, to publish our findings more efficiently.

Learn more about Rescale for Higher Education and Industry

If you are interested in getting started with flexible resources for research or engineering projects, contact us to learn how to easily deploy a range of hardware and software for AI and HPC use cases. To learn more about Dr. Sherif’s work published in Nature, visit Integrating Machine Learning and Symbolic Regression for Predicting Damage Initiation in Hybrid FRP Bolted Connections.