Stephen James Krol

Overview

The project aims to develop a unified computational design framework that guides the emergent, bottom-up behaviour of generative systems, such as agent-based and multi-agent models, through machine learning–driven, top-down control. It addresses the core challenge of managing the complexity and unpredictability of generative algorithms by training deep learning models (e.g., ViTs) to map high-level, design-centred criteria to low-level generative parameters, thereby conditioning and directing emergent outcomes. This approach enables designers to intuitively navigate and control expansive generative design spaces by restructuring latent representations so that specific dimensions correspond to meaningful design features, allowing top-down intention to inform bottom-up processes. Through this integration, machine learning becomes a mechanism for translating design intent into controlled emergent behaviour, redefining AI’s role from stylistic automation to the active guidance and modulation of complex generative systems.

Early Explorations

An initial proof-of-concept is being developed using a simple parametirc Harmongraph system, where a ViT is used to rank generations from the system based on specific criteria. This is then used to evolve the designs to find generations that better suit the designer's goals. Below are two example of early results. More work, including publications and open-sourced code will be posted here when available.