Expertise, playfulness, analogical reasoning: three strategies to train AI for design applications


Since the birth of computer science, computational tools have been developed to automate such design tasks as drafting, analysis and optimisation. In terms of human-machine interaction, these tools passively respond to the instructions provided by the designer. The commands of the tool are known, and therefore the output is largely predictable. For this reason, computational tools have a limited capacity to inform conceptual design and idea generation.

Recently, a new form of computation based on AI has been proposed. AI models can learn instructions independently and only require input in the form of a dataset and a learning objective. The functioning of these models is mostly hidden, and therefore the output of human-machine interaction is less predictable.

This PhD leverages this feature to develop computational tools that can autonomously learn design strategies and interact with the designer to suggest design options that are unbiased by formal descriptions of the design problem (or set of instructions).

The aims of this PhD are (1) implementing different strategies to train AI models in architectural and structural design, and (2) develop AI-based computational tools that allow new forms of human-machine interaction in Computer-Aided Design.

Learning how to design is a daunting task for humans, let alone for machines. Designing involves the mastery of technical skills and the proficient use of imagination and creativity, which requires the designer to make good use of both specialist and general knowledge.

This research assumes that the acquisition of such knowledge is due to three fundamental cognitive mechanisms: expertise, playfulness and analogical reasoning. Training AI in design should therefore attempt to simulate these mechanisms.

The results of the PhD are a set of design tools that illustrate how the simulation of each mechanism can be used to train AI models in design, and how the AI model can interact with humans to support conceptual design and idea generation.


AI models that acquire knowledge by simulating expertise are trained using datasets of design precedents. The tools developed employ AI to augment the natural process of translation and reinterpretation occurring during the development of design ideas by means of sketches. The dialogue with the tool takes place via a visual interface whereby the designer communicates by sketching, while the tool responds by showing the results of the interpretation process. Interaction with the tool does not end after a first iteration. Instead, the designer is encouraged to adjust the initial sketch – or even make new sketches – for several times to explore, with the aid of the machine’s feedback, different elaborations of an idea.

‘Sketches Of Thought’ is a design tool that allows exploring multiple instantiations of a design idea. At the core of the tool, there is an AI model which acquires knowledge from a dataset of pictures of architectural and structural designs and learns to reuse such knowledge to translate hand-drawn sketches into new photorealistic architectural/structural images. Sketches Of Thought was installed at the Future Prototyping Exhibition (Melbourne School of Design, March 2020).

Read more at gabrielemirra.com and on the Future Prototyping Exhibition website.

Here you can download the exhibition catalogue.


Image by Gabriele Mirra


Image by Gabriele Mirra

An analogous tool was developed to explore AI-generated forms that can stimulate the conceptual design of a large span architecture. The AI model was trained on a dataset of 40 well-known design precedents of shell and spatial structures to construct a design space. Unlike the current approaches to parametric design and optimisation, the exploration of the design space does not take place via design variables but through a visual input, which corresponds to a potential design footprint. A design suggestion can be inferred from the AI-generated architectural forms by analysing such features as the location and extension of the support edges, the shape of the openings, or the curvature inversions.


Image by Gabriele Mirra


AI models that acquire knowledge by simulating playfulness do not rely on existing datasets. Instead, these models learn to develop ‘design strategies’ from the exploration of an environment. Similarly to a kid, the AI models explore the environment via trial and error by performing an action and observing the effects of such an action. The driver of an action is the achievement of a design goal. These models learn by reinforcement, that is, by reproducing more frequently those actions that lead to a design outcome that satisfies the goal.


Image by Gabriele Mirra

The figure shows how an AI model can be trained via Reinforcement Learning to generate a structural form. The agent can place 3D blocks in any position, and receives feedback from the environment about the structural feasibility of a designed structure. Once trained, the model can be used to interpret any partially defined input configuration of blocks and complete it by adding extra blocks. The resulting form will be the best attempt made by the model to achieve the design goal starting from the input configuration defined by the designer, thus providing valuable suggestions on how to further develop an idea.

AI models that acquire knowledge by simulating analogical reasoning also engage in a process of free exploration, but are conditioned by a dataset. Unlike expertise, however, the dataset does not contain examples of design precedents, but rather examples of forms from a different domain, such as nature.

Analogies are constructed by mapping information from the source domain (nature) to the target domain (design). This happens via a process of visual abstraction, i.e. by first extracting features from the source domain, and then attempting to reproduce such features to generate an architectural/structural form.

The AI-based computational design tools that are developed in this thesis are meant to support design exploration during conceptual design. The designer can decide to use the outputs obtained through the interaction with these tools to inform the next stages of the design process, including problem-framing and decision making. Although no tool is guaranteed to expand the designer’s creativity or automatically lead to outstanding design solutions, AI models reveal a certain degree of intentionality and thus have a higher potential, compared to other computational techniques, to support conceptual design at a deep level.

DigitalFUTURES Doctoral Consortium - Gabriele Mirra