Decision-making in early-stage design often lacks robust methods for evaluating circularity, resulting in outcomes that may not fully realize their potential for efficiency. This research presents the development of a computational tool or “Intelligent Design Assistant” that employs Graph Neural Networks (GNNs) to deliver real-time assessments of life-cycle performance and material usage for modular designs. By utilizing user-defined, simplified early-stage representations, the tool provides actionable insights into both design and environmental performance. A central point of this approach is the adoption of a graph-based framework where each building module is represented as a node, and its interactions with neighboring modules are captured through connecting edges. This framework not only reflects the intrinsic properties of each module, but it also dynamically evaluates how a module’s characteristics evolve based on its spatial and functional relationships. Although the study focuses on laminated veneer lumber (LVL)—selected for its extensive environmental data—the scalable machine learning model is designed to be applicable to a wide range of construction methods and materials. Through experimental validation, the integration of GNNs has been shown to enhance early design decision-making by providing real-time feedback. The model achieves an accuracy of approximately 85% -90% under conditions similar to the training data. This capability enables designers, clients, and other stakeholders to engage in informed discussions about design modifications and circularity measures well before detailed construction planning begins, thereby promoting more sustainable and circular design practices across the industry.

Project Information

• Title: Echo | Real-Time Evaluation Of The Life-Cycle Performance And Material Usage In Modular Design
• Author(s): Symeon Maniatis
• Year: 2025
• Link: https://repository.tudelft.nl/record/uuid:d574a330-f402-4bb0-8708-3867c17b5cf9
• Type: Master Thesis, Building Technology
• ML Tags: Graph Neural Network, Supervised Learning
• Topic Tags: Artificial Intelligence, Neural Networks, Circularity, LCA

Thesis: Echo | Real-Time LCA & Material Usage Prediction— Symeon Maniatis 1/1

Technical Aspectslink copied

Software and Plugins

Rhinoceros 8, Grasshopper, Visual Studio Code(python),Sustainer Toolkit For Rhino

Digital Workflow

Echo moves from defining what matters, to building clean datasets, to training a predictive model, to plugging it into design tools, and finally validating it for scale-up.

• WP1 — Define scope & requirements: distill circular metrics, early-stage inputs/outputs, UI principles, and GNN needs.
  WP2 — Build datasets: harmonize LVL EPDs; convert designs to JSON graphs (nodes/edges/features); QC + Latin Hypercube Sampling.
  WP3 — Train & evaluate GNN: predict per-module LVL volume/counts; track accuracy; use feature attributions (Integrated Gradients).
  WP4 — Integrate in Grasshopper: real-time components (Import Project, Material Calculator, Holistic LCA, Design Summary, Carbon Performance, Feature Contribution).
  WP5 — Validate & roadmap: case studies, usability/partner feedback, performance roll-up, enablers/limits, plan for other materials/typologies/regions.

GNN Architecture: Why This Architecture

We needed a model that “thinks like the structure”: modules connected in a graph, where not every neighbor matters equally and some inputs have a fixed physical direction. So each building module in a design is translate into a node, where all the participated nodes/building modules comprise a graph. That’s why we chose a graph-attention network with built-in monotonicity, instead of a generic MLP or uniform message-passing GNN.

How the Pieces Fit

We encode concept-stage inputs with a small MLP to standardize. Self-loops keep each module’s own properties in the conversation during message passing. Stacked GAT layers (×3) then learn which neighbors matter without averaging everyone the same.

Regularization & Stability

Between attention blocks we add dropout to prevent co-adaptation and overfitting. A skip connection from the encoder to the readout preserves the individuality of each node (building module).Together with early stopping and light weight decay, this kept training “clean” and generalizable.

Monotonic Physics Priors

For features with known directionality (e.g., increasing window openings or module intersections should reduce required structural material volume), we enforce sign-constrained weights in the encoder/readout so the partial derivative keeps a fixed sign (non-negative or non-positive, chosen per feature). This turns the network into a partly monotone model that respects basic mechanics rather than chasing spurious correlations.

What was Observed

The stack delivered accurate, real-time per-module volume (and counts), held up on harder, held-out designs, and ran comfortably inside Grasshopper. More importantly, predictions moved in the right direction when inputs changed, and explanations matched engineering intuition.

Wy These Layers (one-liners):

• MLP encoder: clean scales + separate monotone channels.
• Self-loops: preserve node identity during aggregation.
• GAT ×3: weight the right neighbors across 1–3 hops (no blunt averaging).
• Dropout: curb overfitting in high-capacity attention blocks.
• Skip connection: stable optimization and linear fallback when needed.
• Monotone readout: enforce fixed-sign effects for physics-guided features.

Bottom Line & Limitations

Echo’s GNN behaves like the structure it predicts for: nodes are building modules, edges are their physical interfaces, attention learns which neighbors actually matter, and monotonic constraints keep predictions moving in the same direction as the physics. That combination is why it works in real time inside Grasshopper without collapsing into a black box. But it works because we kept the physics and topology simple: limited module counts, straightforward wall-to-wall connections, and no hidden interior load paths. Inside that envelope, accuracy, stability, and interpretability line up. Step outside it, and the architectural/structural complexity can invalidate the very assumptions that make the model reliable.