Architectural Layout Solver

Architectural Layout Solver is a neural-network-assisted rectangular layout solver for architectural space planning.

It takes a space program table plus a relationship matrix, selects a working subset of spaces, and fits them into a long rectangular planning envelope while balancing adjacency strength, overlap avoidance, bounds, and corridor access. The repository currently focuses on the layout engine only.

This project is source-available for non-commercial use only. Commercial use requires a separate license from the author.

What It Does

• Loads a space program from CSV with geometry, area, access level, and public-role metadata
• Loads a directed relationship matrix between spaces
• Uses a lightweight message-passing neural network to initialize placement
• Optimizes rectangular room positions inside a fixed planning envelope
• Updates corridor paths with grid-based routing
• Optionally drops the most problematic rooms until a feasible layout is found
• Exports both machine-readable and visual outputs

Repository Structure

• src/architectural_layout_solver/layout_solver.py - main solver and CLI
• data/sample/spaces.csv - sample space program dataset
• data/sample/relationships.csv - sample relationship matrix
• outputs/layouts/ - generated outputs
• LICENSE - non-commercial source license

Installation

python -m venv .venv
.venv\Scripts\activate
pip install -e .

Quick Start

Run with the bundled sample dataset:

architectural-layout-solver --rect-length 400

Or run the module directly:

python -m architectural_layout_solver.layout_solver --rect-length 400

Example with a custom output directory and feasibility pruning:

python -m architectural_layout_solver.layout_solver ^
  --rect-length 400 ^
  --n_spaces 60 ^
  --drop_until_fit ^
  --outdir outputs/layouts/run_01

Inputs

The solver expects two CSV files:

spaces.csv

• One row per space
• Must include geometric fields such as Length_m, Depth_m, and NetArea_m2
• Also uses metadata such as SpaceName, AccessLevel, and PublicInterfaceRole

relationships.csv

• Square matrix of directed relationship weights
• First column must be SpaceName
• Remaining columns must match the space names used by the solver

The repository includes working sample files in data/sample/.

Outputs

Each run writes these files into outputs/layouts/ by default:

• layout_result.csv - final room coordinates, sizes, access metadata, and blocker flags
• corridor_network.json - routed corridor polylines
• layout.png - rendered rectangular layout preview
• dropped_spaces.csv - spaces removed during fit attempts, if any

Main CLI Options

• --rect-length - length of the rectangular planning envelope
• --spaces-csv - path to the space program table
• --relationships-csv - path to the relationship matrix
• --outdir - output folder for generated files
• --n_spaces - number of spaces to sample from the dataset
• --drop_until_fit - keep pruning spaces until the layout becomes feasible
• --corridor_w - corridor width
• --no_access_legend - disable the access legend in the output image

Technical Notes

• Implemented in Python with numpy, pandas, matplotlib, and torch
• Uses a compact message-passing neural network for initialization rather than full model training infrastructure
• Designed around rectangular spaces only
• The current sample data comes from a generated research-building program and is included as an example dataset

License

Non-commercial use is permitted under the license in LICENSE.

• Commercial licensing contact: zhorliari@gmail.com
• This is not an OSI open-source license because commercial use is restricted

Suggested GitHub Metadata

Description:

Neural-network-assisted rectangular layout solver for architectural space planning from program and relationship-matrix data.

Topics:

architecture, space-planning, layout-optimization, computational-design, generative-design, rectangular-layout, floorplan-generation, pytorch, graph-neural-network, message-passing-neural-network