Rendered model showing a complex 3D structure made up of thousands of repeating microscopic patterns. By teaching an AI model to predict the behavior of these patterns, the Drexel team can design stronger, lighter parts much faster than with traditional simulations.
Designing better products means getting two decisions right: the shape you can see and the material structure you cannot. The hidden pattern inside a part acts like its internal skeleton, and small changes there can dramatically transform how it performs. A Drexel University team trained artificial intelligence to stand in for slow computer simulations of these patterns, allowing engineers to try many more ideas in far less time.
The study, published in Structural and Multidisciplinary Optimization, was led by Ahmad Najafi, PhD, associate professor of mechanical engineering and mechanics, with contributions from Nolan Black, PhD, a recent doctoral graduate. The team built a neural network that takes over one of the most time-consuming steps in computer-aided design by quickly estimating how an internal pattern carries force during use.
“Designing both the material and the overall part at once is powerful because the internal pattern can make something lighter, stronger, or more flexible,” Najafi said. “The tradeoff has been speed. Our goal was to keep the accuracy of physics-based modeling while cutting the cost so engineers can explore many more options.”
This advance builds on Najafi’s earlier AI-assisted multiscale design framework, which enables engineers to optimize materials and structures at several scales simultaneously—from microscopic patterns to full-size components. That framework was introduced in two prior studies:
Together, these works establish a foundation for integrating artificial intelligence with solid mechanics to accelerate the design of multifunctional materials and structures.
In typical workflows, engineers rely on detailed finite element analyses to predict how microscopic patterns such as lattices or honeycombs behave under load. Those simulations must be repeated thousands of times as the design changes, consuming enormous computational time. Drexel’s neural network was trained on data from those high-fidelity simulations so it could predict the same outcomes and how they shift when a design is adjusted but in a fraction of a second.
“The key was teaching the model to mirror how the results change when you nudge the design,” Najafi said. “That is what lets the optimization move confidently toward a better solution.”
To test the approach, the researchers set up classic tests that mechanical designers know well. One was a cantilever beam fixed at one end and loaded at the other. Another was a plate pulled in tension. Both used materials with repeating micro patterns that deform in complex, nonlinear ways. Across multiple loading steps, the neural network’s answers tracked closely with conventional analysis on displacements and stresses. Typical errors were in the low single digits when the forces stayed inside the model’s training range. The payoff came in speed. The surrogate ran more than one hundred times faster than the calculation it replaced, which turned weeks of iteration into hours.
“The performance was remarkable,” Najafi said. “That efficiency means engineers can iterate, refine, and test ideas that used to be out of reach.”
Looking ahead, Najafi’s group, with contributions from PhD students Jonathan Gorman, Alireza Ashkpour and Sobhan Honarvar, is extending the framework to handle more realistic materials that can stretch, yield, and even experience damage over time, capturing behaviors critical for advanced engineering applications. They are also working to integrate the model with large-scale design systems used in aerospace, automotive, and biomedical industries, where weight, strength, and cost must be balanced simultaneously. Another goal is to expand the neural network’s training to three-dimensional designs and to use adaptive learning, allowing the model to improve as it encounters new data.
“This is about building smarter design tools,” Najafi said. “By combining artificial intelligence with solid mechanics, we can accelerate discovery and bring advanced materials and structures to practical use much faster.”
Read the full paper: https://link.springer.com/article/10.1007/s00158-025-04133-5
Funding Source: NSF CMMI Program, Grant Number: 2143422