Architectural material has traditionally been classified as either structural or functional and viewed as a characteristic of form rather than a generating factor. In the biological systems form results from bottom-up growth-induced material organisation. Wood is a naturally complex material consisting of different types of cells and tissues. To study the material organisation of wood the project uses the large dataset (Azizan, Guillon, Caraglio, Langbour, Paradis, Bonnet, Brohard, Heinz, Boutahar, Brancheriau, 2016) on the botanical characteristics of different wood species which were analysed with different contrasting techniques in 3 anatomical cuts associated with the 3 planes of symmetry at 3 different magnifications of the microscope: x40, x200, x400 (Hallé, Détienne, Corbière 2017).
Arbor proposes the use of Generative Adversarial Networks (GANs) as a method for extracting a material organisation principle from an existing dataset of timber structure for its further implementation in volumetric models. The organic data informs the inorganic computational system with its own evolutionary neural network algorithms.
Conceptually, GAN consists of two neural networks which work together - discriminator and generator. The generator tries to generate images that look like the given example and the discriminator compares the generated images with the originals and updates the generator if it succeeds. Such cooperation not only allows two networks to teach each other but also to understand the main principles and patterns of given data. This leads to the main advantage of GANs, which is the ability to generate new unseen imagery based on provided examples, what allows such networks to interpolate between selected pairs of images by generating intermediate steps. During the training process, GAN architecture forms a latent space, which can be imagined as a space of network's knowledge.
The latent-vector interpolation translation into the Z axis of volumetric structure is used as a method to visualise the internal and external morphology of timber material organisation. In this process, basic geometrical data describing the anatomical properties of timber structure as a point cloud is extracted.