A key insight in Complex Modelling is that the integration of machine learning in the networked model enables adaptive parametrisation and breaks the reductionism inherent to parametric modelling. As our models are able to intersect with information drawn from the world around us through the sensing and simulation of environment and structural and material behaviour, we encounter the dilemma of the information-rich but data-heavy design model. Complex Modelling exposes the nature and sheer volume of the information of future modelling paradigms that seek to design for performance and optimise material deployment. These processes introduce radically new scales and types of data into design modelling.
Complex Modelling examines the idea of information-rich design through a focus on machine learning. By employing machine learning across the networked model, we combine analytical methods for evaluation and classification with creative methods for design generation and evolution. Machine learning is different from parametric strategies in that models are not explicitly defined but rather trained on data sets. This means that the information-rich design environment acts as a source of training data. Data can be brought in as a predefined data set or generated continually, and their associated models can be trained either discretely or continuously.
Here, models exist in multiples, in thousands of models, which are spawned by the generative system to then be analysed by the learning system. Models are no longer singular endpoints, where finding the optimum represents the end of the modelling process. Instead, models learn from other models. They belong to processes of expansion, increasing in number and in complexity at each step of evolution (1).
Machine learning presents new practices for architecture. In Complex Modelling, we examine three central emergent practices (2). Firstly, we investigate new modes of mapping and characterising solution spaces in non-explicit ways. Secondly, by intersecting machine learning with simulation, we develop alternate strategies for performance prediction, which avoids brute-force calculation. And thirdly, we extend the adaptability of design information by applying machine learning onto sense data gathered from the fabrication process.
In Learning to be a Vault, unsupervised machine learning is used to map solution spaces and categorise outputs into observable classifications thereby letting the designer navigate these high-order design spaces more easily (3). In Lace Wall, neural networks are interfaced with simulation in order to optimise structural morphology and enhance performance. Here, machine learning is interfaced with multiscale simulation strategies. At the scale of the element, a genetic algorithm is used to optimise the topology of the cable network for both performance and fabrication requirements, and at the scale of the structure, an artificial neural network ...