Unlocking Large-Scale Structural Synthesis: High-Performance GPU Topology Optimization for Architectural and Civil Engineering Applications
Interview with Raul C. Llamas-Sandin – Universidad Europea de Madrid & Airbus Operations SL
Raul Llamas-Sandin is a PhD researcher at Universidad Europea de Madrid and aerospace engineer at Airbus Operations SL, whose work sits at the intersection of large-scale structural optimization and high-performance computing. His research addresses stress-constrained topology optimization at scales that are genuinely relevant to architectural and civil engineering, running models of up to hundreds of millions of elements on a single consumer-grade GPU. The computational approach he has developed avoids the need for structural sensitivities, improving both speed and robustness, while an outer control loop ensures constraint fulfillment throughout the process.
At CDFAM, Raul will present the method’s application to real-world structural synthesis problems, including multi-load configurations, mixed materials, self-weight, and tension-compression asymmetry.
He will also speak to the role of AI in accelerating code development and explore broader possibilities in computational design beyond his core research area.
Can you introduce your work and explain what you’ll be presenting at CDFAM regarding high-performance GPU-based topology optimization for large-scale structural synthesis?
My PhD research focuses on stress-constrained, large-scale structural topology optimization. I work on methods that don’t require structural sensitivities, making the algorithms faster and more robust, though they do require an outer control loop to guarantee constraint fulfillment.
For large models, this would be unfeasible without utilizing the GPU for the finite element solve. Consequently, a significant part of my research involves developing an extremely efficient algorithm capable of tackling up to hundreds of millions of elements on a single consumer-grade GPU.
What are the main computational and modeling challenges in scaling topology optimization to architectural and civil engineering domains, and how does your solver address them?
Moving beyond elementary form-finding in these fields requires the ability to model multiple load introduction points, the self-weight of both designable and fixed domains, mixed materials, and their corresponding stress allowables.
To obtain crisp, unambiguously interpretable results, the mesh needs to be exceptionally fine, which leads to extremely large models.
For example, filling the bounding box volume of the Sagrada Familia in Barcelona would require about 1.2 billion bricks, with the final building requiring roughly 130 million. Therefore, large-scale optimization is the only way to tackle truly interesting problems, with GPU technology and modern programming languages serving as key enablers.
How does the solver manage real-world constraints such as tension-compression asymmetry, thermal loading, or the integration of fixed non-designable zones into the optimization process?
The finite element solver handles all of these features in the standard manner. The structural topology optimizer then directly uses the internal element stresses to assign a material density -a common approach- based on local stress allowables. Next, a variable-radius diffusion process followed by a thresholding operation, both controlled by the outer loop, redistributes the material and removes volume at each iteration. Before the subsequent iteration, the elements corresponding to voids or non-designable domains are restored to their nominal properties.
Making this process efficient and stable is non-trivial, but our results demonstrate the method’s feasibility.
Structural topology optimization with 110 million solid elements in the ground structure ran on a 10Gb gaming GPU in 8 hours
What does the workflow look like from initial problem definition to geometry generation, and how do you structure solver configurations to remain flexible across different design scenarios?
Although the software is currently still in the research phase, it features a functional user interface. Users can define point or distributed loads and supports, as well as voids, regions with different material properties (like stiffness, weight, and thermal expansion coefficients), and external models to outline designable volumes.
The numerical computations are handled automatically whenever possible; the code performs its own diagnostics and selects the appropriate solution parameters. Post-processing currently includes a dedicated results visualizer, STL export capabilities, and outputs in ParaView format encompassing stress, density, displacement, and force fields.
Can you speak to your recent exploration of using the .3mf file format to capture and communicate both geometry and metadata, and how that compares to more traditional file exchange methods in your workflow?
The .3mf format is highly appealing because it enforces model quality and eliminates the ill-conditioning that frequently occurs with other formats. It is also much more compact.
It is still early days, however, as the .3mf parser for Julia—the programming language I use—is very new and is actually being developed alongside this project.
Structural topology optimization with 140 million solid elements in the ground structure ran on a 10Gb gaming GPU in 4 hours (using symmetries)
What do you hope to share with and learn from the CDFAM community through your participation this year?
I believe my research will interest structural design practitioners across various disciplines. However, as an active aerospace engineer working in conceptual aircraft design, I am especially looking forward to cross-pollination with ideas from fields entirely different from my own.
During my presentation, I will highlight the key enablers of my research—particularly the extensive use of AI in code development—and explore possibilities in computational design beyond my core topic, which I will illustrate with practical examples.
If this work raises questions you want to pursue, CDFAM Barcelona is the place to do be.
The symposium brings together practitioners and researchers working across computational design, structural engineering, architecture, and AI-assisted development, creating the conditions for exactly the kind of cross-disciplinary exchange Raul describes.
Join us in Barcelona to hear Raul present alongside others advancing the state of the art in computational design and design engineering.
