A neural network that picks the right steel connection.

Choosing a steel connection is a judgment a structural engineer makes from a handful of design parameters, guided by code provisions and experience. The research question behind this project was whether that judgment is learnable from historical design decisions, and if so, whether a model could make it fast enough to be useful early in design, when engineers are still exploring options.

The deliverable was not a paper alone. It was a trained model, an honest evaluation of it, and a deployed interface the panel could operate during the defense.

Machine learning theses tend to fail in one of two places, and this project had to survive both. The first is data: what gets collected is often not in a shape a model can learn from. The second is justification. A model that works in a notebook but has no story behind why that architecture, and no comparison against anything simpler, does not survive a panel.

There is a third problem specific to defenses. A model demonstrated by scrolling through a notebook on a laptop asks the panel to take the student's word for it. A model the panel can type inputs into and watch respond is a categorically different kind of evidence.

The design inputs were structured into a clean tabular dataset with the connection type as the target, and the problem framed as multi-class classification. Categorical fields were encoded and numeric ones scaled, with the scaler fitted on training data only. Fitting it across the whole dataset leaks test information into training and quietly inflates every number you report afterwards.

An artificial neural network was chosen over a simpler classifier because the decision boundaries between connection types are not cleanly separable on the raw inputs. That choice was made against baselines rather than assumed: simpler models were trained first, precisely so the thesis could answer the question a panel always asks, which is why not something simpler.

The model was developed in Python and trained on the prepared dataset, with the architecture and training regime tuned against validation performance rather than test performance. The trained model was then wrapped in a web interface and deployed to a public URL.

The interface is deliberately plain. It takes the design inputs, returns the predicted connection type, and does nothing else. During a defense, anything more is a distraction from the model, and the model is the thesis.

The tool is live and publicly accessible, which is unusual for a thesis artifact; most end their life as a notebook on a student's laptop. The student defended with a working system the panel could operate, backed by evaluation numbers computed properly.

The more useful outcome for anyone reading this as a prospective client: this project is what a research engagement with Kleio actually produces. Not a report about a model, but a model, the evidence for it, and something you can click.