ML + AI in Design, Engineering + Architecture
Overview of Presentations from CDFAM Computational Design Events
As computational design continues to evolve, machine learning and AI are increasingly embedded into workflows across design, engineering and architecture.
At CDFAM Computational Design Events, these technologies are not discussed in isolation—they are being (slowly) integrated, and augmenting existing engineering systems to deliver reliable, repeatable, measurable impact.
But before these tools can be applied effectively, they must be fed the right data—and in engineering, that’s a significant challenge.
While the world breathlessly celebrated breakthroughs in image and language models, those systems were trained on billions of publicly available data points of various quality scraped from the internet, sometimes legally, sometimes not.
Engineering, manufacturing, and architectural design data, by contrast, are fragmented, fiercely proprietary, and rarely centralized, even within a single department of an organization.
For machine learning to deliver value in these domains, significant groundwork must be laid: cleaning, labeling, and structuring datasets that were never intended for algorithmic use.
This includes everything from simulation outputs and CAD files to process metadata and real-world performance feedback.
The difficulty isn’t (only) in building the machine learning model, it’s in preparing the ground beneath it.
The past and future presentations at CDFAM reflect the reality that we are still in the early stages of integrating AI meaningfully into engineering practice.
Again and again, presenters return to the challenge of data readiness—highlighting that effective application of ML depends on careful preprocessing, curation, and structuring of domain-specific data.
Rather than pursuing a general ‘text-to-CAD’ fantasy that promises to solve all engineering problems with a single prompt, we see work that trains models on narrow, well-bounded applications. These targeted tools offer immediate value—provided the data is there.
NASA’s “Text to Spaceship” project exemplifies this tension.
While the goal is ambitious, automating the translation of mission intent into design through natural language interfaces, NASA researchers were careful to stress that this is very early work.
Turning a paragraph of text into a CAD model and simulation result is a vision that will require years of data collection, structuring, and feedback integration to become reliable. Their presentation showed what is possible, but also what it will take to make it practical.
“If you don’t think data is a problem, you haven’t really approached your data problem.”
Alexander Lavin of Pasteur Labs
So before we can get to our text-to-spaceship, or bracket, or reactor, or cat hammock, or whatever, the presentations from previous CDFAM events archive show the way forward.
Emerging Themes Across Presentations
The use of AI is increasingly domain-specific, with customized models for structural steel, lattice mechanics, and organ generation.
Presenters emphasized human-in-the-loop systems, where AI augments rather than replaces expert decision-making.
Many tools aim for real-time or near-real-time feedback, often by blending lower-fidelity models with intelligent prioritization of high-fidelity validation.
Integration is key—successful systems are end-to-end, bridging the gap between design intent and fabrication or clinical use.
The following highlights from CDFAM 2023 through 2025 show how designers, engineers, and researchers are tackling this challenge, and how AI and ML are reshaping design, optimization, simulation, and fabrication.
Data Capture, Preparation, and Synthesis
Lexset (Francis Bitonti) – Focused on synthetic data generation for computer vision and machine learning training in manufacturing. Their work uses domain-randomized, physics-based rendering to simulate edge-case conditions for AI model training, reducing the burden of collecting and labeling real-world data.
Pasteur Labs (Alexander Lavin) – Shared a framework for simulation intelligence using differentiable programming and scientific machine learning to improve CAD-to-manufacturing workflows. Their AutoPhysics platform enables AI-native simulation across digital engineering toolchains, emphasizing the need for structured, machine-learnable data.
Carnegie Mellon (Chris McComb) – Introduced the “Design for Artificial Intelligence” (DfAI) framework, including the role of design data curation as a distinct discipline. Highlighted the need for data pipelines that include validation, versioning, and traceability to make engineering data usable for machine learning systems.
Autodesk AI Lab (Karl DD Willis) – Highlighted the difficulty of applying ML to CAD workflows due to limited, complex, and high-dimensional data. Their work involved building models trained on structured B-rep geometry to create editable design outputs, demonstrating how valuable curated and labeled design data can unlock generative capabilities in professional design tools.
Simulation Acceleration and Field Prediction
SimScale – Showcased the integration of cloud-based simulation and AI to provide real-time design feedback in engineering applications. Their platform enables scalable, collaborative simulation workflows where physics solvers are increasingly augmented by ML surrogates, reducing time to insight in the early stages of design.
Neural Concept – Presented the use of deep learning models for rapid prediction of simulation outcomes, trained on domain-specific engineering datasets to enable real-time feedback in product development, particularly in the automotive and aerospace sectors.
Navasto – Demonstrated how their ML platform enables accurate, fast prediction of aerodynamic performance in automotive design, reducing reliance on full CFD simulations. Their workflow accelerates early-phase iteration and integrates seamlessly with existing design and engineering environments.
Design Generation and Optimization
NASA (Ryan McClelland) – Showcased an AI-driven pipeline that transforms mission requirements, written in plain language, into CAD-ready geometry and simulation-ready models, reducing iteration time from months to days.
Generative Engineering – Demonstrated decision-centric engineering using LLMs and generative models to guide design exploration and summarize tradeoffs in mechanical and aerospace contexts.
ARUP (Rick Titulaer) – Presented a case study-driven overview of how informed, data-driven computational design is being implemented in real architectural projects. His presentation highlighted the integration of contextual data into parametric workflows to improve environmental performance, decision-making, and stakeholder engagement.
Philosophical Reflections on Machine Learning in Design
New Balance (Onur Yuce Gun) – Reflected on the limitations of current ML systems in creative design, noting that while algorithms can rapidly generate variations, they often reinforce known patterns rather than introduce meaningful novelty. His talk emphasized the importance of maintaining human authorship in brand-defining design processes.
Massachusetts Institute of Technology’s DeCoDE Lab (Kristen M. Edwards) – Evaluated the capabilities and limitations of vision-language models in early-stage design workflows. Her research focused on how these tools interpret design prompts and the implications for engineering creativity, authorship, and system reliability.
Technical University of Denmark (Ole Sigmund) – Delivered a pointed critique of AI in engineering, arguing that while ML tools may excel at interpolation within known data ranges, they fail at extrapolation—where innovation often lives. He warned that blind reliance on AI may inhibit, rather than enable, breakthrough discoveries, and whether AI for topology optimisation is a dead end?
Upcoming AI + ML Presentations at CDFAM Amsterdam
Looking ahead, upcoming presentations at CDFAM Amsterdam 2025 will extend these explorations—ranging from specific industrial applications to foundational advances in simulation and optimization.
CustoMED.ai will present TEMP: Enabling Surgeons With Patient-Specific Guides, showing how AI can streamline medical workflows from scan to print.
SimScale will continue its exploration of cloud-based workflows in Physics & AI Engineering Simulation in the Cloud.
nTop will present new developments in Accelerating Design Optimization Using Implicit Geometry and ML enabled by their recent acquisition of CloudFluid.
Pasteur Labs will return with a philosophical and technical follow-up titled The Unreasonable Effectiveness of Simulation Intelligence.
Datameister will expand on its work in constrained design workflows in Constrained Creativity In AI-Accelerated Automotive Design.
KeyWard will introduce a no-code approach to ML with Building Surrogate Models for Physics Simulation.
PhysicsX will present Building Beyond Imagination: Engineering the Unthinkable with AI, exploring large physics models and generative workflows.
Together, these talks will challenge assumptions about AI’s creative and technical limits, and further clarify where machine learning offers meaningful acceleration—and where it may still fall short.
While the technology continues to advance and data continues to be gathered, it is ultimately the people building, using, and challenging these tools who will shape their impact.
At CDFAM, it’s not just about seeing what’s possible—it’s about meeting those doing the work, asking the hard questions, and collaborating across disciplines.
Whether you’re deep in the technical trenches or exploring broader implications, you’ll find people at CDFAM who share that commitment to thoughtful, rigorous design.
Register to attend CDFAM Amsterdam or CDFAM NYC to connect with the experts shaping these tools and ideas. Whether you’re building ML pipelines for simulation, exploring data-driven design systems, or questioning the role of AI in engineering creativity—CDFAM is where these conversations are happening.
