Cracking the Code of Composite Manufacturing: How AI Predicts Polymer Flow in Fibrous Materials
Fiber-reinforced composites are essential for high-performance products such as aircraft components and wind turbine blades. However, ensuring that polymer resin properly permeates the intricate network of reinforcing fibers during manufacturing is a complex and costly process. Jimmy G. Jean, a PhD candidate in aerospace engineering at TU Delft, is pioneering an innovative AI-driven approach to overcome this challenge. By developing high-fidelity machine learning surrogate models, Jimmy’s research delivers predictions of resin flow up to 2,700 times faster than traditional simulations, without sacrificing accuracy. In this Q&A, Jimmy shares insights into his cutting-edge work, its practical implications, and the breakthroughs reshaping composite manufacturing.
PhD Candidate at TU Delft, Jimmy G. Jean
The CAELESTIS Consortium visiting TU Delft in 2024
How would you describe the main goal of this research to someone without a technical background?
Imagine a bird carefully weaving a nest—an intricate tangle of twigs, grass, and small branches creating a complex maze of tiny gaps and channels. Now imagine trying to predict exactly how water would flow through that nest’s winding pathways, but instead of twigs, you’re working with carbon/glass fibers that go into airplanes, wind turbines, and sports cars. Welcome to the world of fiber-reinforced composites manufacturing, where getting the polymer resin (a viscous liquid) to flow right can mean the difference between a flawless part and one that fails. This research tackles a deceptively simple question: How does polymer resin move through tangles of fibers?
What practical problems in manufacturing or daily life could benefit from your approach to predicting flow in fibrous materials?
Manufacturers of advanced composite parts—like carbon fiber bicycle frames, wind turbine blades, and aircraft components—inject resin into mats of densely packed fibers using a process called resin transfer molding. The resin must saturate every small space. If it doesn’t, dry spots or air pockets can form, creating weak points that jeopardize the entire structure. Manufacturers face several costly problems when this happens: defective parts that fail quality tests, wasted materials, long production times, and products that might not withstand real-world stress.
Mitigating these problems involves running complex computer simulations that can take days, weeks, or even months to complete depending on the level of complexity. For engineers trying to design new products or optimize manufacturing processes, this is painfully slow and computationally expensive. Our machine learning (ML)-based approach resulted in a speedup of up to 2700 times faster on the specific problem of simulating flow in fibrous microstructures. This acceleration means that engineers can iterate faster through design cycles and better produce high-quality composites.
How does using deep learning and image-based methods improve on traditional simulation techniques? Can you share a concrete example?
Our research contribution is the following. Our team trained two ML models on thousands of numerical simulations. Then, after training, one can feed the model a microscopic image of the fiber arrangement, and it quickly predicts both how fast the polymer resin will flow (velocity field) and what the pressure will be throughout the material (pressure field). Moreover, one can use a simple equation (Darcy’s law) to compute the permeability of the sample from the predicted flow fields. The permeability value basically represents how resistant the material is to fluid flow. The most innovative part of our research is the introduction of a sliding window technique. The limitation of convolutional neural networks (which our ML models are based on) is that they work on fixed-size images. The sliding window technique allowed us to extrapolate predictions to images with variable sizes, without the need to generate new training data or change the ML model architecture.
A Concrete Example
In an international benchmark by Syerko et al. (2023), traditional simulation methods required up to 136 hours to analyze and predict the permeability of a real microstructure sample gather via μ-CT scans. Our ML approach accomplished the same task in just 3 minutes—up to 2,700 times faster—while producing results that matched the accuracy of traditional simulations.
Were there any surprising findings or challenges during the project that might interest non-experts?
We trained our ML models on microstructure data generated synthetically, meaning that we algorithmically determined the distribution of fibers within them. We then implemented a sliding window technique (well-known in computer science) based on our intuitive understanding of the underlying physical process. When we tested the entire procedure on a synthetic as well as a real microstructure, we were surprised by its remarkable performance in both cases. The main takeaway there is to cultivate the habit of combining bits from different fields; it might result in surprising breakthroughs.
How could open-sourcing your methods and data help other researchers or industries? Could everyday products eventually be affected?
The most impactful aspect of this work is the commitment to open science. We released our entire dataset, alongside the ML model code and training procedures. This transparency means other researchers and engineers can directly apply or expand on these tools. For industry, such accessibility could dramatically speed up product development.
If you could visualize the research as an image or analogy, what would it look like?
Coming back to the introduction, imagine a bird’s nest made of thousands of twigs, tangled in a beautiful mesh. Now, picture water flowing through the mesh, watch as it wends its way through the gaps—sometimes fast, sometimes slowing, following hidden paths carved by the twists and turns of the nest. With our research, it is as if you could glance at a photo of the nest and know instantly how that water would move around. It’s seeing the invisible journey of fluids, revealed through the power of images and AI.
Find out more about TU Delft's with CAELESTIS
Technische Universiteit Delft is leading CAELESTIS’ efforts to create a generative structural design based on process-property relationships (WP5), including the delivery of open-source code and probabilistic material models.
Delft University of Technology ranks among the top 10 engineering and technology universities in the world. Based in the Netherlands, it contributes to solving global challenges by educating socially responsible engineers and performing world-class research.
