Presented at
SIGGRAPH Asia 2024
https://dl.acm.org/doi/10.1145/3680530.3695440 Read Paper Online
Figure 1: Simulation of fungal spread using our neural network-based fungal automaton, supervised and trained by an AI model chain. It is then synchronized with a laser in reality to implement dynamic light containment, guiding fungal growth to match pre-designed patterns consistent with the simulation.
The simulation and control of fungal morphology has garnered increasing interest in fields such as art, gaming, architecture, design, and human-computer interaction (HCI). In recent years, deeper integration of art with biological research has gradually clarified the concept of Bio-Art. As fungal research becomes more entrenched in Bio-Art, it continues to spark new inquiries into ethics, social issues, and aesthetics, while also providing unique inspiration for other fields such as robotics, sensor research, and bio-manufacturing.
Although practitioners of fungal art are often classified as artists, creating art based on fungal morphology typically requires a deep theoretical understanding of biology. The integration of computer algorithms and simulation technologies has bridged the gap between artists and fungal morphology, significantly reducing the time consumption and validation difficulties of traditional fungal experiments. However, for most artists, operating computer programs remains a significant challenge.
Our research treats the fungal spread pattern as a two-dimensional graphic time-series generation problem. We propose a simulation pipeline based on a neural network-driven fungal cellular automaton. This pipeline first employs the Efficient-ViT (E-ViT) image segmentation model and a customized Temporal Convolutional Network (TCN) time-series model to learn the patterns of fungal spread from sequences of images. The TCN model then provides supervised training to the neural network (NN) cells, enabling each cell to independently respond in a manner consistent with the learned fungal spread patterns, thereby achieving realistic morphology simulations. Furthermore, to achieve pre-designed spread shapes, we experimentally introduce the use of lasers to establish dynamic light boundaries that limit the direction of fungal growth. The automaton is then connected to a real-world laser device, which successfully guides the fungus to spread into various complex shapes during practical tests, in accordance with the simulation predictions, without the need for computer vision (CV) monitoring.
Benefiting from the pipeline of AI models and the design of the NN-based cellular automaton, artists can simply provide a video of fungal spread to achieve high-fidelity simulations of fungal morphology and real-world light control, without needing to delve into programming and algorithms.
Demo Video: