Biologically inspired design holds great value for product design. Inspired by geometric features from natural creations, designers can create organic nature-inspired shapes. Traditionally, even for experienced designers, biologically inspired design is related with time-consuming or trial-and-error effort.
We propose DeepBionicSyn, an approach for automatic creative 3D biologically inspired design. We view the biologically inspired design problem as an out-of-distribution synthesis problem based on a dataset of natural creations and man-made artifacts. DeepBionicSyn is trained on AC-BIONIC, a large-scale dataset containing 28k reference shapes. We mine the required design candidates in a shared neural parametric representation space of the input dataset using a creativity synthetic solver. Finally, a human-in-the-loop local manifold subspace exploration technique allows designers to explore design variants effectively.
Compared with previous three-dimensional generative methods for creativity, our framework can meet organic biologically inspired design requirements and does not require pre-analysis for shape collections. We evaluate the effectiveness of our method on a chair product biologically inspired design task based on the chair-animal hybrid dataset. We introduce a metric for evaluating creative biologically inspired design tasks to quantitatively assess our approach and other potential alternatives.
Overview of our framework for 3D biologically inspired design, which combines deep implicit generative modeling and user-in-the-loop interactive. (a) Dataset consist of natural creations as inspiration and man-made artifacts as product target. (b) After training an IMAE, there exist some biologically inspired synthetics in the shape perception boundary in latent space. The trained encoder and decoder form a parametric representation of 3D shape. In latent space, a problem is how to find out creative biologically inspired design. Creativity Synthetic Solver obtains final results via two steps. (c) A hybrid function containing weak-shot classifiers(W-Cls) and quality term is deployed to get high score candidates. Through multimodal optimization sampling, candidates filtered out are recommended to users as if cold start in recommendation. (d) Candidates with high scores are then inputted into a human-in-loop system. The Projector view of the web interface allows the user to explore the entire design space. After selecting the desired synthetic, a local manifold subspace search algorithm is employed to explore potential design variants in a slide-bar manner.