CDFAM Computational Design Symposium

CDFAM Computational Design Symposium — Barcelona 2026

Javier Blanco Cordero · Quix

Engineering teams have found AI useful for discrete tasks: SQL queries, graphs, summaries. Higher-value work, root cause analysis, physics-based ML, production deployment, remains out of reach. The standard diagnosis is that the models need to improve. Javier Blanco Cordero's experience points the other way: the models are already sufficient. The infrastructure surrounding them is not.

Using a real rotor balancing case involving 50,000 rotors per year, this talk walks through what it takes to move AI from task-level support to autonomous engineering investigation. With the right infrastructure in place, a model that had underperformed on summary data identified a discriminating signal in 100 kHz vibration data that decades of domain expertise had missed. That same model independently designed a deployment plan that reduced total balancing runs by 41%.

The presentation covers what was built, the results, and what it means for any organization sitting on complex engineering data.

Links

Talk page with full transcript

Watch the talk on YouTube

Cite this talk

Search the full text of every recorded CDFAM presentation

CDFAM Computational Design Symposium — Barcelona 2026

Quercus Hernández · Emmi AI

Engineering is entering a new paradigm where AI is no longer limited to accelerating isolated simulations, but is becoming a foundational layer of the industrial design and manufacturing stack. Across sectors such as automotive, energy, semiconductors, and aerospace, engineering workflows remain constrained by slow simulation feedback loops, fragmented CAD-to-CAE pipelines, and computational bottlenecks that limit design exploration and innovation speed.

In this talk, we present the vision of Large Engineering Models developed at Emmi AI: physics-aware foundation models designed to operate directly on industrial geometry and process inputs while replacing large parts of the traditional numerical simulation workflow. Rather than focusing on narrow surrogate models, this approach consolidates geometry processing, physics prediction, and post-processing into a unified AI-native engineering interface that provides near real-time feedback to engineers and designers.

As a concrete industry showcase, we present NeuralMould, our first commercially deployed Large Engineering Model for injection moulding. The model enables engineers to evaluate design changes such as gate placement, material selection, and process parameters with near-instant feedback, eliminating traditional meshing bottlenecks and dramatically accelerating design iteration cycles. In addition, we introduce Noether, our recently released open-source framework for building and deploying engineering AI models at scale. Noether provides the infrastructure for training, fine-tuning, and operating Large Engineering Models across domains, enabling industrial partners and researchers to develop their own AI-native simulation capabilities.

Links

Talk page with full transcript

Watch the talk on YouTube

Cite this talk

Search the full text of every recorded CDFAM presentation

Beyond Parametric: Explorable Simulation for Real Design Iterations

Presentation Abstract

Design space exploration promises engineers the ability to evaluate thousands of variants in hours. The prerequisite—rarely discussed—is building a parametric geometry model that won’t break. This front-loaded investment assumes you know your design space before exploring it. Real engineering iteration isn’t parametric. Design changes are structural and topological: fundamentally different geometries evaluated against the same performance criteria.

Setup time, not simulation time, is the real bottleneck in enabling simulation-driven exploration for real engineering teams.

At Generative Engineering, this is what we are addressing. We enable explorable simulations that work with both parametric exploration and novel designs. We will demonstrate how this is unlocked by using existing simulation data inside engineering organisations to train geometry preparation models rather than physics surrogates.

This lets the messy, collaborative, real design iterations be simulated with zero setup cost. But the same principle extends further: if a model understands what makes geometry simulation-ready for a given analysis, it can generate that geometry directly. We’ll show work on closing this loop—from a sketch or annotation on a previous design to simulation-ready geometry, collapsing the traditional boundary between design intent and engineering validation.

Speaker Bio

Laurence Cook is Co-Founder and Chief Product Officer at Generative Engineering.

After academic work at MIT, Stanford, and Cambridge, Laurence brought his computational engineering to the commercial industry and is now leading the initial commercial product of Generative Engineering. A 20 strong team of computational and commercial experts who have currently raised €13m.

Deterministic engineering analysis assigns fixed values to loading conditions, geometry, and material properties. The approach is tractable, but it forces a choice between conservative overdesign and exposure to failure modes that fall outside assumed limits. Neither outcome is satisfying.

Greg Grigoriadis of Metisec describes a design toolkit that replaces fixed inputs with statistical distributions and runs Monte Carlo simulations across the resulting parameter space. The output is a probabilistic picture of performance: failure probabilities, sensitivity rankings, and the specific conditions that actually drive risk.

A sensor mounting bracket for a smart wearable serves as the test case. Traditional optimization cut bracket weight by 30%. Probabilistic analysis revealed the design had been tuned to an improbable drop event and still carried unresolved thermal failure risk. Incorporating that information allowed the team to re-optimize and achieve a 50% weight reduction at a demonstrably low failure probability.

Greg is an engineering consultant specializing in consumer electronics, digital twin technologies, and advanced simulation workflows. His practice combines physics-based modeling with data-driven methods across finite element analysis, predictive maintenance, and automated computation pipelines.

Presented at CDFAM Barcelona 2026. Learn more at cdfam.com.

Body-in-White development sits at the intersection of structural performance, manufacturability, and schedule pressure — and traditional workflows are increasingly strained by the pace of platform proliferation and regulatory demands. This episode examines how SEAT/CUPRA is modernizing BIW development through Synera’s Multi-Agent System (MAS), connecting CAD, CAE, and manufacturing processes within a unified automation framework to accelerate iteration and improve engineering traceability.

Tilman Steininger, Customer Success Manager for Agentic Engineering at Synera, and Juan De Dios Escribano Felguera, Team Lead for Body-in-White and Corrosion Development at SEAT/CUPRA, present their collaborative work digitizing structural development workflows across the full BIW development cycle. Recorded at CDFAM Barcelona 2026.

About CDFAM

The CDFAM Computational Design Symposium is a global conference series bringing together researchers, engineers, and software developers working at the intersection of computational design, AI and machine learning for engineering and architecture. The program spans computational geometry, generative design, topology optimization, simulation-driven workflows, and advanced manufacturing across scales from architected materials to architectural systems. CDFAM events are held worldwide.

cdfam.com

The gap between geometric modeling and structural simulation has always been a workflow problem as much as a technical one. Every time a designer changes a shape, the mesh breaks, boundary conditions need redefining, and the simulation has to be rerun from scratch. At the scale of iterative or generative design workflows, that overhead compounds quickly.

In this episode recorded at CDFAM Barcelona 2026, Johannes Mueller-Roemer presents SubSimX, a pipeline that addresses this by treating subdivision surfaces as a native simulation medium rather than a preprocessing step. Designers work on a standard Catmull-Clark control mesh. Meshing and boundary condition setup happen once. From that point, geometry edits propagate to the existing analysis mesh in seconds, structural feedback is continuous, and the design session never has to stop for remeshing.

The technical core is a precomputed mapping between the control mesh and the volumetric interior, combined with a local mesh-repair step that maintains element quality through large deformations without changing mesh topology. Coupled with RISTRA, Fraunhofer’s GPU-accelerated FEM solver, the system delivers real-time or near real-time structural feedback. Johannes demonstrates how this enables dense design sweeps, topology optimization followed by subdivision refinement, and AI-in-the-loop exploration within practical industrial workflows.

Johannes Mueller-Roemer is Deputy Head of Interactive Engineering Technologies at the Fraunhofer Institute for Computer Graphics Research IGD. He received his PhD from TU Darmstadt in 2019, where his dissertation introduced GPU data structures for tightly coupled modeling and simulation — the foundation SubSimX is built on.

Recorded at CDFAM Barcelona 2026 — https://cdfam.com/barcelona-2026/

Topology optimization has long been constrained by the computational cost of working at meaningful scale. For architectural and civil engineering applications — where domains span tens of meters and element counts run into the millions — that constraint has made high-fidelity structural synthesis largely impractical on standard hardware.

In this episode recorded at CDFAM Barcelona 2026, Raul C. Llamas-Sandin presents a GPU-accelerated solver that addresses this directly, reducing solve times for large-scale problems to minutes on a workstation. He covers the key technical features that make it relevant for the built environment: asymmetric tension-compression stress constraints, thermoelastic analysis, and geometric constraint enforcement for fixed architectural elements and fenestration.

Applications demonstrated include bridge design and high-rise structural cores at spans up to 80 meters, as well as free-form architectural geometries exported as watertight STL files ready for digital fabrication. The solver’s modular configuration approach decouples problem definition from the numerical backend, making it adaptable across project types without reengineering the workflow.

Raul is a Future Projects Engineer at Airbus Operations SL and Assistant Professor of Aerospace Engineering at Universidad Europea de Madrid, with over two decades of experience in advanced aircraft design. He holds an MSc in Aerospace Vehicle Design from Cranfield University and 14 patents spanning aerodynamic devices and plasma actuation.

Recorded at CDFAM Barcelona 2026 — https://cdfam.com/barcelona-2026/

Most 3D generative AI has been built to produce things that look right. Hao (Richard) Zhang argues that for the built environment, that is the wrong objective. Buildings must function — and that means AI systems need to reason about spatial relationships, load paths, equipment routing, and regulatory constraints, not just geometry.

In this episode recorded at CDFAM Barcelona 2026, Richard introduces Functional AI as a framework for 3D design automation and walks through Augmenta’s work applying it to real building projects. The technical scope covers agentic AI for MEP path finding, generative modeling of complex structural and mechanical layouts, and progress toward a foundation model for building data. He also discusses what makes non-residential construction — commercial, medical, institutional, and mission critical — a particularly hard problem for AI systems to crack.

The conversation closes on a concrete milestone: two elementary schools in Michigan where the electrical system was fully modeled and delivered using AI-powered generative design, a first for the industry.

Hao (Richard) Zhang is VP of AI and R&D at Augmenta and a Professor in the School of Computing Science at Simon Fraser University. He is a Fellow of the IEEE, a member of the ACM SIGGRAPH Academy, and served as Technical Papers Chair for SIGGRAPH 2025.

Recorded at CDFAM Barcelona 2026 — https://cdfam.com/barcelona-2026/

Mike Frei, of ARENA Physica, joins CDFAM Barcelona 2026 to discuss the company’s AI foundation model for electromagnetism and what it means for the future of electronic hardware design.

Mike brings an unusual combination of deep physics intuition - PhD from Columbia studying geometry and electron behavior at the molecular scale - and operational experience from McKinsey and C-suite roles at Swiss aerospace and defense company RUAG. That background informs Arena’s approach to reimagining electromagnetic simulation across semiconductors, aerospace, and automotive sectors.

The conversation covers how Arena is building what they call Artificial Intuition for EM design, the limitations of current simulation approaches, and why a foundation model purpose-built for electromagnetism represents a fundamentally different path forward.

Watch the full presentation:

Full archive at cdfam.com/barcelona-2026/

Pablo Hermoso Moreno, Senior Solutions Architect at NVIDIA, joins CDFAM Barcelona 2026 to discuss how AI is reshaping Computer-Aided Engineering workflows.

Pablo draws on his background in aerospace engineering, aerodynamics at Mercedes F1, and cloud-scale AI at AWS to frame where CAE bottlenecks actually live - and what NVIDIA is doing about them. The conversation covers AI physics and surrogate models for real-time simulation feedback, the role of digital twins, and how agentic AI built on the NVIDIA NeMo stack is automating end-to-end engineering workflows.

He also walks through the AI-Q Blueprint, a production-ready enterprise research agent, and explains how the NeMo Agent Toolkit manages LLM call prioritization to reduce latency without additional hardware investment.

Full presentation and archive at cdfam.com/barcelona-2026/