AI learns to simulate the growth and shape of trees in response to their environment

A research team from the Department of Computer Science and the Institute for Digital Forestry at Purdue University, with collaborator Sören Pirk from the University of Kiel in Germany, has discovered that artificial intelligence can simulate growth and shape of trees.

The DNA molecule encodes both tree shape and environmental response in a tiny subcellular package. In DNA-inspired work, computer science professor Bedrich Benes and his associates have developed new AI models that compress the information needed to encode the tree shape into a neural model the size of one megabyte.

After training, the AI ​​models encode local tree development that can be used to generate complex tree models of several gigabytes of detailed geometry as output.

In two articles, including one published in ACM Transactions on Charts and the other in IEEE Transactions on Visualizations and InfographicsBenes and his co-authors describe how they created their tree simulation AI models.

“AI models learn from large data sets to mimic discovered intrinsic behavior,” Benes said.

Non-AI-based digital tree models are quite complex and involve simulation algorithms that take into account many non-linear factors that influence each other. Such models are needed in fields such as architecture and urban planning, as well as in the gaming and entertainment industries, in order to make designs more appealing to their potential clients and audiences.

After working with AI models for almost 10 years, Benes expected that they would be able to significantly improve existing methods for digital tree twins. The size of the models was surprising, however. “It’s complex behavior, but it’s been compressed into a small amount of data,” he said.

Co-authors of ACM Transactions on Charts The article was by Jae Joong Lee and Bosheng Li, Purdue graduate students in computer science. Co-authors of IEEE Transactions on Visualization and Computer Graphics the paper was by Li and Xiaochen Zhou, also Purdue graduate students in computer science; Songlin Fei, Dean’s Chair in Remote Sensing and Director of the Institute of Digital Forestry; and Sören Pirk of the University of Kiel, Germany.

Researchers used deep learning, a branch of machine learning within AI, to generate growth models for maple, oak, pine, walnut and other tree species , with or without leaves. Deep learning involves the development of software that trains AI models to perform specified tasks via linked neural networks that attempt to mimic certain functionality of the human brain.

“Although AI has become seemingly ubiquitous, so far it has proven very effective in modeling 3D geometries unrelated to nature,” Benes said. These include efforts related to computer-aided design and improving algorithms for digital manufacturing.

“Obtaining a 3D geometric vegetation model has been an open problem in computer graphics for decades,” Benes and his co-authors said in their ACM Transactions paper. Although some approaches to simulating biological behaviors are improving, they noted that “simple methods that would quickly provide many 3D models of real trees are not readily available.”

Experts with expertise in biology have traditionally developed tree growth simulations. They understand how trees interact with environmental conditions. Understanding these complex interactions depends on the characteristics conferred on the tree by its DNA. These include branching angles, which are much greater for pines than for oaks, for example. The environment, on the other hand, dictates other characteristics that can cause the same type of tree grown in two different conditions to have completely different shapes.

“Dissociating the intrinsic properties of the tree and its environmental response is extremely complicated,” Benes said. “We looked at thousands of trees and thought, ‘Hey, let the AI ​​learn it.’ »And perhaps we can then learn the essence of tree shape through AI. »

Scientists typically build models based on hypotheses and observations of nature. As models created by humans, they have reasoning behind them. The researchers’ models generalize the behavior of several thousand trees of input data encoded in AI. Then, researchers validate that the models behave the same as the input data.

The weakness of AI tree models is that they lack training data describing real-world 3D tree geometry.

“In our methods, we had to generate the data. So our AI models don’t simulate nature. They simulate tree growth algorithms,” Benes said. It aspires to reconstruct 3D geometric data from real trees inside a computer.

“You take your cell phone, take a picture of a tree and you get 3D geometry inside the computer. It can be rotated. Zoom in, zoom out,” he said. “This is the next step. And it fits perfectly with the mission of digital forestry.”