AI in design

Under “AI in design” we understand the research field that uses methods of artificial intelligence to facilitate or partially automate the task of designing technical products. The contributions of artificial intelligence can concern closed sub-processes, such as system analysis or the optimization of subsystems or sub-properties. Or they can support humans in shaping the design itself, or even partially replace them. If this is support, i.e., the takeover of certain routine tasks, we speak of AI-assisted design. If, on the other hand, AI takes over a superordinate task in developing the design solution, we speak of AI-based, automated, or semi-automated design.

Artificial intelligence excels at automating tasks that were previously performed by humans. Typically, the algorithms that process certain input data into the desired outputs are not explicitly programmed. Instead, a structure with a large number of variable parameters (weights) is provided, which is trained on an equally large amount of data. During training, the weights are optimized with the goal of achieving the best possible agreement with the data. An impressive example is image recognition. If, for example, the task is to detect whether an image shows a cat or not, the algorithm’s task is to accept a bitmap file as input and output the statement of whether the image shows a cat. Explicitly programming this task—covering all relevant breeds, sizes, viewing angles, and postures—would be very time-consuming. On the other hand, we know that humans can perform such a task quickly, which proves that constructing a suitable algorithm is fundamentally possible. According to the principle mentioned above (parameterized computational architecture and training), this task is solved.
In engineering, a task traditionally performed by humans is design or synthesis. If the characteristics of a system are defined and its behavior must be determined, we are dealing with an analysis task. If the reverse process is intended (determining the characteristics—e.g., dimensions and materials—of a system based on the desired behavior), a synthesis task arises. Analysis tasks can usually be solved well using explicitly programmed algorithms (e.g., analyzing the deformation of a component using the finite-element method). Design or synthesis tasks, in principle, require the analysis and evaluation of an enormous variety of possible solutions, which can never be exhaustively performed (neither by humans nor by machines). Up to now, design tasks have been carried out by humans by strongly reducing the variety of solutions to be analyzed and evaluated, using experience, intuition, and subjective considerations (such as the choice of a particular style).
When formalizing design tasks, the search through the variety of possible solutions is transformed—by suitable parameterization—into the exploration of high-dimensional spaces. Previously, attempts to automate design tasks mainly converted them into optimization problems. Thus, of the variety of conceivable solutions, only those lying along a one-dimensional search path are analyzed and evaluated, which depends on the problem formulation (choice of objective function and constraints of the optimization) as well as on the choice of search algorithm. This type of focusing on a strongly reduced subset of possible solutions differs fundamentally from the way focusing occurs in human design processes. By using AI methods, design automation can connect to human design methodology, which is considered proven.
However, the design task (compared, for example, to image recognition) is of great complexity and variety, so it cannot be expected to be completely transferred to machines. It is more likely that artificial intelligence will act in interaction with humans. This is probably the best context in which to develop the idea of such “human–machine teaming.”
On the one hand, it makes no sense to completely abandon classical engineering based on intuition, creativity, subjective considerations, and human experience; on the other hand, its complete algorithmic transcription is not feasible in the foreseeable future. It is therefore promising to rely on a synergy between humans and AI, in which classical, computer-aided methods can of course also be integrated.
A first step in developing such teaming is to transform classical, explicitly programmed methods into AI architectures. In this way, they are better able to work in synergistic interaction with humans or with AI modules that map human capabilities. In contrast to the classical approach of supervised machine learning, the datasets required for training are not produced externally but are generated in a targeted manner (e.g., by random generation) as part of the training process. A first successful method developed at the institute in this context is the PEN method for structural optimization.