Author: erikswahn
Machine learning, translation and architecture
Exploring cGAN-based image-to-image translation in an architectural context.
Paths
Forking paths
Here, each plan is a single, continuous forking space generated using a depth-first search algorithm until all cells are visited and all walls are removed in the path of the search – then sorted by wall layout.
The number of cells in the space limits the number of unique alternatives generatable (removing mirrored and rotated duplicates).
2 x 2 cells 1 unique maze
2 x 3 cells 5 unique mazes
2 x 4 cells 12 unique mazes
2 x 5 cells 31 unique mazes
3 x 3 cells 12 unique mazes
3 x 4 cells 112 unique mazes
3 x 5 cells 509 unique mazes
3 x 6 cells 2133 unique mazes
4 x 4 cells 481 unique mazes
4 x 5 cells 5395 unique mazes
4 x 6 cells 20132 unique mazes
5 x 5 cells 9054 unique mazes
Here, two openings are created on side walls, and the size of cell columns and rows are randomly expanded in 1:1 / 2:1 proportions.
From bits to matter, from plan to space
An alphabet: laser cutter operations
Plan generation with nesting shape grammars in the Forsythia editor, written by John Greene:
https://github.com/johnalexandergreene/Forsythia/tree/master/app/grammarEditor
Going from plans to laser cut sheet:
From plans to cavities and elevations:
From grammar to building:
Modular x Reciprocal: Diagrams
MODULAR x RECIPROCAL
Stochastic Reconstruction
Iterative substitutions in two- and three-dimensional matrices
using Markov random field models of von Neumann neighborhoods
Tiling
Building a probabilistic model
Iterative substitutions in two dimensions
Iterative substitutions in three dimensions, starting with source data
Iterative substitutions in three dimensions, starting with Perlin noise
Isolated bodies
Machine Learning for Artists
In my current project, I’m working on generative processes with some connections to machine learning techniques. I just found a great resource, Machine Learning for Artists, with lots of video lectures and code related to principle component analysis, neural networks, deep learning and much more.
Check out the “classes” tab for the video lectures.
Digital Grotesque
For anyone interested in the techniques behind the Digital Grotesque project and others by Michael Hansmeyer / Benjamin Dillenburger, check out these course pages from ETH Zürich with Processing scripts:
http://www.caad.arch.ethz.ch/blog/high-resolution-2-materialise/
http://www.caad.arch.ethz.ch/blog/high-resolution-architecture-sacred-spaces/
More of the theory behind it is described in the article “Mesh Grammars –
Procedural Articulation of Form”:
https://cumincad.architexturez.net/system/files/pdf/caadria2013_259.content.pdf