Arup needed to evaluate far more structural design options early in projects so engineers could optimize performance, material use, and embodied carbon, but traditional FEM simulations were too slow and queue-bound to support near-real-time iteration.
Reported impact
Arup reduced structural simulation evaluation from hours to minutes, achieved 90-95 percent accuracy against FEM baselines, and ran GNN execution and training up to 50 percent faster, allowing teams to make lower-carbon design decisions earlier and with better quantitative support.
Arup sped up its structural design workflow by replacing repeated full finite element method runs with graph neural network (GNN) surrogate models trained and deployed on NVIDIA DGX Spark. Engineers generate many building-massing or other structural options through parametric modeling, then use the GNN models to predict structural performance much faster than classical simulation so they can compare trade-offs earlier in the design cycle. DGX Spark supplies the local accelerated compute for both model execution and training, which lets Arup keep project data on site while shortening feedback loops from hours to minutes. The result is a practical two-stage architecture in which conventional FEM establishes the reference behavior, GNN surrogates provide fast iterative screening, and engineers use those AI-derived predictions to narrow the design space before making final structural decisions.
Arup generates structural options through parametric modeling, runs those options through GNN surrogate models on DGX Spark to predict performance in minutes, and uses the results to compare cost, carbon, and engineering trade-offs before narrowing designs for deeper analysis.
By training and deploying graph neural network (GNN)-based surrogate models on NVIDIA DGX Spark™, Arup reduced structural simulation evaluation time from hours to minutes while achieving up to 95% accuracy.
DGX Spark delivered up to 50% faster GNN model execution and training compared to alternative platforms.
Running locally, DGX Spark maintains secure data handling while supporting faster project delivery and more sustainable design outcomes.
With GNN-based predictions running on DGX Spark, thousands of structural schemes can now be evaluated efficiently.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
NVIDIA Sample Jun 26, 2026 Zaha Hadid Architects Zaha Hadid Architects needed to shorten weeklong feedback cycles for highly detailed design scenes, protect a proprietary 40-year design archive, and scale AI-assisted visualization without sending sensitive client or… 40-year portfolio DALL-E 2 · NVIDIA NIM large language models · NVIDIA Omniverse NVIDIA Sample Feb 5, 2026 EDAG Production Solutions EDAG needed to help manufacturers launch and optimize factories faster, but fragmented planning tools, disconnected supplier data, and manual status tracking slowed collaboration, increased errors, and delayed readiness… More than 100 workers trained virtually before installation EDAG digital twin platform · NVIDIA Omniverse · OpenUSD NVIDIA Sample Feb 2, 2026 Worley Worley needed to unlock more than 50 years of engineering knowledge across over 100,000 documents, but fragmented tools, isolated repositories, and slow document retrieval caused engineers to waste time and made it hard… 100,000+ documents Microsoft Azure · Milvus · NVIDIA AI Enterprise 