LLM-Driven Toolpath Design for Digital Fabrication
Neural Inverse Design of Self-Shaping Composites The research project explores an LLM-driven approach to computational design and additive manufacturing, enabling the generation of fabrication-ready 3D objects directly from natural language and visual inputs. Through a multi-agent workflow, the system generates parametric toolpaths that integrate geometry, material distribution, and fabrication strategies into a unified design representation.…
Building Typology Analysis
Building Typology Analysis This project uses Faster R-CNN (Region-based Convolutional Neural Network), a deep learning object detection model, to detect and classify buildings based on their visual and geometric characteristics in high-resolution satellite imagery. Trained on annotated satellite images, the model locates individual buildings and assigns them to predefined morphological categories, such as rectangular, L-shaped,…
Self-Supervised Urban Analysis
Self-Supervised Urban Analysis This project uses Self-Supervised Learning (SSL) through DINO, a vision transformer model, to extract visual representations of urban environments from high-resolution satellite imagery. Unlike supervised approaches, SSL learns meaningful visual features without requiring predefined categories or annotated training data. The model captures spatial patterns, textures, building arrangements, and other visual characteristics of…
Machine Learning for Architectural Floor Plan Evaluation
Machine Learning for Architectural Floor Plan Evaluation The research focuses on computational methods for assessing architectural floor plan quality, with particular emphasis on machine learning and multimodal models. It investigates different evaluation paradigms, including similarity to reference designs, compliance with architectural requirements, intrinsic design quality, and instruction-following. A central focus is text-floor-plan alignment- evaluating whether…
Neural Inverse Design of Self-Shaping Composites
Neural Inverse Design of Self-Shaping Composites Self-shaping frustrated composites are fiber-reinforced polymer sheets that transform from a flat state into a three-dimensional geometry without molds, driven by internal stresses embedded in the material. The mechanism relies on geometric frustration: during curing, laminated layers with different fiber orientations create incompatible internal strains. These strains drive the…
Machine Learning Models for Analyzing Urban Visual Experience
CNNs for analyzing urban visual experience Urban Spatial Openness Index (USOI), captures the spatial openness of open spaces in urban environments. Open spaces are defined as non-built areas: streets, squares parks, rivers. USOI values range from 0% (no visible space at all) to 100% (complete spatial openness), and work best as a comparative tool between…
Real-time Fabrication Analysis
Real-time Fabrication Analysis of Molds for Concrete Elements In his PhD. Thesis, Dr. Austern developed an analysis method which evaluates the fabrication parameters of complex geometries. The method predicts the feasibility, material use, and machining time required for fabricating moulds. It achieves this by interrogating geometric properties instead of the traditional machining simulations. Using the…
Automated Security Design
Automated Security Design Despite advances in computational design in the building sector, security system design is still based on outdated software and design paradigms. This leads to long design times as well as to inadequate system designs. In his role as Chief Science Officer for BeamUP, Dr. Austern developed a new system design method that…
Reading Legacy 2D plans
Reading Legacy 2D plans with Machine Learning As Chief Science Officer of BeamUP, Dr. Austern invented and helped develop a unique process which converts unstructured, 2D CAD files to intelligent 3-dimensional models which represent building geometry. BeamUP’s process combines geometric and statistical Heuristic algorithms, with state-of-the-art Machine Learning models. The process can automatically identify space…
CGANs for generating architectural interior designs
CGANs for generating architectural interior designs In interior space planning, the furnishing stage usually entails manual iterative processes. Machine learning has the potential to automate and improve these processes while maintaining creativity and quality. The aim of this study is to develop a furnishing method that leverages machine learning as a means for enhancing design…
The Computational Design and Machine Learning lab explores the use of state of the art models and algorithms to solve architectural problems in design, analysis, and fabrication. CD_ML was founded by Dr. Guy Austern with a vision to bridge the gaps between design and coding by developing innovative software tools for the architectural profession.
The candidate will head the lab’s research efforts on Applying Graph Neural Networks on BIM data, which is a part of the new Neuro-Symbolic AI in AEC research arc. In this research arc we join forces with researchers from the Stuttgart ICD as TU delft to envision and develop the next generation of BIM software. Our research will explore the use of Graph based Machine Learning models on Knowledge Graph representations derived from BIM files. We will use these models as the basis for a BIM autocomplete mechanism, which will allow BIM software to actively participate in the design process by making suggestions and filling in design details. More information