From Simulation to Synthesis: Architectural Modeling with Context-Based Encoding Using Data-Driven Computational Machines

Digital architecture has been a fast-evolving domain since the first computer program was written to imitate the technique of drawings. During the last few decades, various computational modeling tools have been developed to equip architects with the ability to computationally solve complex design problems and produce novel building geometries. However, the rapid propagation of programmable design tools has raised a critical situation where researchers are celebrating the productivity of computing power while losing a proper understanding of what computation is about. This brings digital architecture to a deadlock where most of our state-of-the-art computational models in architecture nowadays are not fundamentally advanced compared with their early beginnings decades ago. This situation motivates this work to build a comprehensive theoretical understanding regarding the capacity and limitations of our modeling techniques, and to explore a way out of the current dilemma. Towards this goal, this work emphasizes that computational modeling in architecture is not about geometric modeling which evaluates the capacity of models by the geometrical complexity of the designs, rather, it has always been a branch of artificial intelligence which helps architects to develop new ideas and concepts by the means of intelligent customized computer programs. Following this aspect, this work first reviews the state-of-the-art computational models as well as their counterparts in artificial intelligence. The review shows that, on a technical level, computational modeling in architecture faces two fundamental obstacles – the combinatorial explosion and the frame problem – resulting a phenomenon that most problem-solving programs in architecture succeeded in small-scale research had failed in larger-scale applications. These obstacles are credited for the conceptual assumption that having a faithful representation of the target is a necessity for dealing with the target. The modeling paradigm that follows this assumption is named as simulation. This work then reveals that the contemporary artificial intelligence challenges the paradigm of simulation by giving up the explicit representation of an object and shifting towards the object’s contextual relations to other objects. For contemporary artificial intelligence, any object can be the context of any object. Hence, the computational process no longer represents specific problem content nor the problem-solving process. This allows contemporary artificial intelligence to bypass the mentioned obstacles and use the same technique to solve problems in different fields. This new data-driven context-based modeling paradigm is named as synthesis. Finally, this work applies the idea of synthesis to architectural modeling, using a series of experiments as demonstrations. These experiments cover a wide range of topics, including urban-scale physics, urban dynamics, and ontologies for representations. The experiments show that by giving up reasoning we outperform the conventional modeling techniques. This derives an important implication that, as now all problems can be solved by the same principle, the new challenge ahead of us is no longer how to solve problems, but what questions to ask. This corresponds to the changing role of computers from the tools for computation, to the machines for automation, and finally to the infrastructures for communication.