Scientists and engineers often work with "woven composites"—materials made by weaving together bundles of super-strong fibers, like carbon fiber, and setting them in a polymer matrix. Think of it like high-tech fabric that's incredibly strong but also lightweight.

The problem is, figuring out exactly how these materials will behave under stress is incredibly difficult and slow. The traditional method is a long process of trial and error, physically creating and testing countless samples, which is both expensive and time-consuming.

Another approach is to use powerful computer simulations. These simulations try to predict the material's strength by calculating all the complex interactions between every single fiber at a microscopic level. While these simulations are more accurate than older methods, they are painfully slow. Running just one detailed simulation can take hours or even days, requiring immense computational power. This major bottleneck makes it impractical to design new materials quickly.

A New Idea: What if an AI Could Learn to Predict Material Strength?

This is where the research of Dr. Ehsan Ghane comes in, offering a clever solution to this long-standing challenge. The core idea of this thesis is to use Artificial Intelligence (AI), specifically a type of AI called a neural network, to solve the problem.

A neural network can be trained to recognise patterns in data. The idea is to show the AI many examples from the slow computer simulations—"if you apply this much force, the material stretches by this much"—and let it learn the relationship. Once trained, this AI model can act as a "surrogate," or a super-fast digital shortcut. Instead of running a slow simulation, engineers could just ask the AI, and it would give a near-instant prediction.

However, this approach comes with its own major hurdle: AI models are famously data-hungry. To train them properly, you need massive amounts of data. But if the only way to get that data is by running those incredibly slow simulations, we're back to square one!

Also Read: Hidden Biases in Medical AI

The Breakthrough: A Smarter AI That Learns from Physics

This is the most exciting part of the story. Dr. Ghane didn't just use a standard AI; he developed a much smarter approach to overcome the data problem. His work focused on two groundbreaking strategies.

First, instead of relying only on expensive, high-quality ("high-fidelity") simulation data, he used a technique called transfer learning. He trained his AI models on large amounts of cheaper, less-accurate ("low-fidelity") simulation data first. This gave the AI a general understanding of material behavior. Then, he "fine-tuned" the model with a much smaller, carefully selected set of high-quality data. It’s like learning to paint by first sketching thousands of simple drawings and then taking a masterclass to perfect the final details. This drastically reduces the need for slow, expensive simulations.

Second, and perhaps most ingeniously, he created a "Physics-encoded Neural Network." Instead of treating the AI like a black box that just memorizes data, he embedded the fundamental laws of material physics directly into the AI's architecture. This means the AI wasn't just learning from examples; it was learning the underlying rules that govern how materials behave. This physics-based approach made the AI much more powerful. It could make highly accurate predictions with far less training data and, crucially, it could even extrapolate—make reasonable predictions for situations it had never seen before.

AI's Thinking model hits walls

Why This Matters: Faster Innovation for a Better Future

The outcome of this research is a set of powerful and efficient AI models that can accurately predict the complex behavior of woven composites in a fraction of a second. These frameworks bridge the gap between analytical simplicity and the realism of complex numerical simulations.

This is a game-changer for materials science. It dramatically speeds up the design process for new composite materials, making it cheaper and more efficient to innovate. This could lead to lighter and more fuel-efficient airplanes, safer cars, and more efficient renewable energy technologies. By making these advanced design tools more accessible, this research paves the way for a new era of material discovery.

Read the full paper: https://gupea.ub.gu.se/handle/2077/85666