
Affordable curved glass façades in sight with AI tool
Image credit: Ruslan Guseinov / IST Austria
Computer scientists at Austria’s Institute of Science and Technology (IST) have developed a tool, based on a deep neural network, which allows for the free-form design of affordable curved glass façades.
Curved glass panels can be used to create beautiful and unusual façades, but traditional construction methods are extremely expensive. The panes are typically created through 'hot bending' - an energy-intensive and expensive process in which the glass is heated and formed using a mould.
The alternative is cold bending – developed around a decade ago – in which flat panels are bent in and fixed to frames on site. Cold bending is cheaper, as well as making the glass easier to transport and of higher geometric and visual quality.
The IST scientists, working with an international team of researchers, set out to create an interactive tool to help designers develop cold-bent glass forms which are both beautiful and possible to manufacture. This is a challenge because designing full cold-bent glass façades is such a complex problem that using traditional computational methods to obtain stresses and shapes is impractical.
“While it is possible to calculate when an individual panel will break, or provide a safety margin for additional loads, working with the full façade – which often comprises thousands of panels – is simply too complex for the conventional designer tools,” said Dr Ruslan Guseinov, co-first author of the study.
The researchers set out to create software which would allow a non-expert user to interactively edit a surface while receiving real-time feedback on the shape and associated stresses for each panel.
They decided on a data-driven approach and ran more than a million physical simulations to build a database of possible curved glass shapes, represented in standard CAD format. They then trained a deep neural network on these data. This neural network precisely predicts possible glass panel shapes for a given boundary frame, which can then be used in a façade designed by an architect.
“We knew that a given boundary does not uniquely define the panel, but we didn’t anticipate that the [neural network] would be able to find multiple solutions, even though it had never seen two alternative panels for a single boundary,” said Dr Konstantinos Gavriil, co-first author.
From the set of possible solutions, the tool selects the pane geometry which best fits the façade design, taking into account factors such as smoothness of frames and reflections. The user can then adapt their model to reduce stress and improve the aesthetic quality, or automatically optimise the design to minimise the number of “infeasible” panels (which must be created through hot bending).
Once the user is satisfied with the form, the software exports the flat panel shapes and frame geometries necessary to construct the façade.
The computer scientists put the tool to the test by manufacturing frames and glass panels, including panels under very high stress. In the worst case, they observed only miniscule deviation from the predicted shapes and all panels were fabricable.
“We believe we have created a novel, practical system that couples geometric and fabrication-aware design and allows designers to efficiently find a balance between economic, aesthetic and engineering criteria,” said Professor Bernd Bickel of IST.
Sign up to the E&T News e-mail to get great stories like this delivered to your inbox every day.