The short answer is that MCD is a novel framework for counterfactual optimization in design problems. For the long answer, please refer to our public access paper or our open-source implementation of MCD .
The chosen rider-bike combination is passed to MCD, which has been equipped with a method that evaluates ergonomic score/aerodynamic drag, depending on the type of optimization being performed. Riders are represented by 8 bodily dimensions, whereas bikes are represented by 14. MCD will attempt to produce as many unique, novel designs that meet specific performance targets, while also trying to keep generated designs as similar to the original bike as possible.
All sorts of design/optimization problems. MCD places no restrictions on parametric design representation, with integer, continuous, and categorical variables supported. MCD also places no restriction on the type of model used for performance predictions. This means MCD will work just as well with differentiable and non-differentiable machine learning models, as well as simulation or pure mathematical models.
Of course!
Equipped with a machine learning model capable of predicting structural properties from 39 design parameters, MCD has been used to optimize bikes for properties such as weight and safety factor (i.e. the ratio of the maximum stress a bike can withstand to that of a typical loading scenario).
Making use of a Contrastive Language-Image Pretraining (CLIP) model, MCD was used to generate novel bike designs that resembled prompts such as 'A futuristic black cyberpunk-style road racing bicycle' and 'A sturdy compact bright blue mountain bike with thick tires'.
We would like to thank Noah Wiley for his contributions to the image-to-rider-dimensions pipeline, as well as the ergonomic and aerodynamic performance prediction functions. We would also like to thank Amin Heyrani Nobari for his work on the bike rendering pipeline, and Brent Curry for allowing us to make use of the BikeCAD bicycle design software for this demo.