How AI-Driven Design Automation Works

Introduction

Design is fundamental to how we perceive, understand, create, and interact with the physical and virtual worlds. In essence, design is a form of communication between designer, product and user and is a discipline we use to create beauty, invent tools to make life better and easier, as well as influences how we think, feel and act. In the context of product or service design, design encompasses ideation, the planning of a new product to solve a specific problem with the intention of improving human experience. Unlike domains such as mathematics or physics, in the realm of design an optimal solution may not exist, as there may be elements of human subjectivity that impedes objective and metric-based evaluation. Therefore, it is the designers’ challenge to explore and present various solutions and constantly make changes based on a fundamental understanding of the product’s purpose, user needs and real-world constraints.

Design is a practice that can be exceedingly complex, often iterative but also captivating when there is strong alignment between form and function. Critically, design is constantly adapting to leverage new technologies and satisfy evolving user needs and requirements.

Design automation: where AI meets design

Artificial Intelligence (AI) has the potential to revolutionise the design space across industries by simplifying complex tasks, overcoming the burden of iterative changes via automation, and discovering optimal design solutions faster. The strength of AI is the ability to learn patterns, trends and characteristics in historical design data and using this information to generate new designs that conform with defined parameters such as goals, constraints, and resources, and can be adapted to a variety of different design contexts, challenges and products.

The two main types of design challenges

Everything in the man-made world is the outcome of planned and deliberate design (whether that makes it good is a different matter). Given that products and services are so diverse and radically different, it may be hard to identify common themes. Instead of focusing on the products themselves, the focus should instead be on the nature of the challenge the product solves. Two main types of design challenges emerge: aesthetic design challenges and engineering or operational design challenges. Depending on the nature of the ‘product’, it may have to face both challenges.

Let’s explore the difference between these two types of design challenges:

Aesthetic challenges

For aesthetic challenges, the designer is tasked with creating the visual, cosmetic and stylistic elements of a product or interaction. Aesthetic design is about ideating compelling new concepts, creating attractive and eye-catching products or sensorially stimulating experiences that attract and delight users. Sometimes but not always, the role of aesthetics is to accentuate form over functionality and practicality. Aesthetic design challenges are everywhere, from the clothes we wear, the programs and advertisements we watch, the brands we connect with and the homes we buy. Architects, graphic designers, UX/UI designers, artists and even copy-writers are all great examples of ‘aesthetic’ designers.

Engineering and Operational Challenges

Engineering and operational challenges are more technical and constraint-based challenges that are tackled by more technical designers, such as engineers or software developers. Engineers are challenged with translating concepts into reality and working within real world constraints, such as competing goals and priorities, time, budget, human resources, materials and even laws of physics. To orient themselves and assist with overcoming such complex challenges, engineers use one or more objective functions – a metric or priority that can be maximized or minimized – such as cost, resource efficiency and even regulatory compliance. For example, Daisy AI is an AI platform that automates the design of timber floor plans, uses two objective functions - cost and resource efficiency – and optimises designs towards maximizing these objective functions. Designs are evaluated and rated against both criteria, with those that perform better considered more optimal designs.

How AI can tackle these design challenges

Even with sophisticated software design tools, traditional design methods often require designers and engineers to map out and develop their solution from scratch by hand, which can be a tedious, time-consuming and frustrating process when client requirements or design specifications change. But AI is disrupting this highly manual and iterative process by combining data and intelligent automation to streamline and simplify the process. Let’s look at two AI techniques that can be used to overcome these design challenges: Generative Adversarial Networks (GANs) and Reinforcement Learning (RL).

Generative Adversarial Networks (GANs)

The use of Generative Adversarial Networks (GANs) is an AI technique that generates synthetic data that possess the traits and characteristics of the training data used to train the GAN; a coupled pair of neural networks. To ensure that the generated data is authentic, the GAN pits these neural networks against each other, one called the ‘generator’ and one called the ‘discriminator’. The generator creates new data instances while the discriminator evaluates them for authenticity, with the generator being successful if the discriminator is unable to differentiate between the generated data and the original data. GANs have a host of potential applications such as for image generation (such as faces, buildings, scenery), 3D modelling and much more. The advantage of GANs is the relatively small amount of data input required to train and generate high quality images. Taking Renault’s AMT as an example again, designers can now enjoy automated AMT model generation. What’s more, any model output to designers would have had been examined by the algorithm, making sure it works on a specific car model.

Reinforcement Learning

Reinforcement Learning is a machine learning technique based on a system of reward. In a basic scenario, a machine would try to solve a problem without any instructions. ‘Good’ behavior – actions that progress it towards its goal or some optimal solution – is rewarded, while bad behavior that negatively impacts performance and achieving its goal would be punished. The machine repeats the task and eventually finds the best approach after trying all possible ways of approaching a problem. In situations such as where an engineer is dealing with many simultaneous constraints and is having trouble working out the perfect trade-off among them, RL can simplify the process. Importantly, RL is a process of trial and error with the learnings accumulating towards achieving some optimal solution: the machine will finally work out a solution after enough times of failure.

Conclusion

While there are many different types of products and services that exist in the real world, all of them can be distilled into two main types of challenges: aesthetic and engineering. Understanding the distinction between the two is critical to identifying opportunities where AI can be used to address inefficiencies and bottlenecks and provide insight into the most effective technique to solve it. AI-driven design is an exciting domain that will see rapid growth in the coming years, as sectors from construction, manufacturing and even the arts begin to learn of its transformative potential.

Written by Clayton Black and Yang Shen