AI in Design

We understand 'AI in design' to be the research field that uses artificial intelligence methods to facilitate or partially automate the design of technical products. Contributions from artificial intelligence can relate to specific sub-processes, such as system analysis or the optimisation of subsystems or sub-properties. Alternatively, they can support or even partially replace humans in shaping the design itself. If AI takes over certain routine tasks, this is referred to as AI-assisted design. Conversely, if AI takes on a more senior role in developing the design solution, it is referred to as AI-based, automated, or semi-automated design.

Artificial intelligence is highly effective at automating tasks that were previously carried out by humans. Typically, the algorithms that convert input data into desired outputs are not explicitly programmed. Instead, a structure with a large number of variable parameters (weights) is provided and trained using an equally large amount of data. During this training process, the weights are optimised to achieve the best possible agreement with the data. Image recognition is an impressive example of this. If the task is to detect whether an image shows a cat, for example, the algorithm accepts a bitmap file as input and outputs a statement indicating whether the image shows a cat. Explicitly programming this task to cover all relevant breeds, sizes, viewing angles and postures would be very time-consuming. However, we know that humans can perform such a task quickly, which proves that constructing a suitable algorithm is fundamentally possible. According to the aforementioned principle (parameterised computational architecture and training), this task can be solved.
In engineering, design or synthesis is a task that is traditionally performed by humans. If the characteristics of a system are defined and its behaviour must be determined, this is an analysis task. If the reverse process is intended, i.e. determining the characteristics (e.g. dimensions and materials) of a system based on the desired behaviour, a synthesis task arises. Analysis tasks can usually be solved effectively using explicitly programmed algorithms (e.g. analysing the deformation of a component using the finite element method). Design or synthesis tasks, in principle, require the analysis and evaluation of a vast number of potential solutions, a process that can never be completed in full (by either humans or machines). Until now, design tasks have been carried out by humans by greatly reducing the variety of solutions to be analysed and evaluated using experience, intuition and subjective considerations (such as the choice of style).
When design tasks are formalised, the search for possible solutions is transformed into the exploration of high-dimensional spaces through suitable parameterisation. Previously, attempts to automate design tasks mainly converted them into optimisation problems. This meant that only the solutions lying along a one-dimensional search path were analysed and evaluated. This path depended on the problem formulation (the choice of objective function and the constraints of the optimisation) as well as the choice of search algorithm. This type of focus on a strongly reduced subset of possible solutions differs fundamentally from how focus occurs in human design processes. Using AI methods enables design automation to connect to proven human design methodology.However, compared to image recognition, for example, the design task 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 interact with humans. This is probably the best context in which to develop the concept of "human–machine teaming".
It makes no sense to completely abandon classical engineering, which is based on intuition, creativity, subjective considerations and human experience, because its complete algorithmic transcription is not feasible in the foreseeable future. Therefore, it is promising to rely on a synergy between humans and AI, in which classical computer-aided methods can also be integrated.
The first step in developing such teaming is to transform classical, explicitly programmed methods into AI architectures. This makes them better able to work in a synergistic way with humans or AI modules that map human capabilities. Unlike the classical approach of supervised machine learning, the required training datasets are generated internally (e.g. via random generation) as part of the training process, rather than being produced externally. One successful method developed at the institute in this context is the PEN method for structural optimisation.

In the PEN (Predictor–Evaluator Network) method, the optimisation problem involves searching for the geometry with the greatest stiffness within a given design domain, while adhering to a volume constraint. This approach is based on the interaction between a trainable artificial neural network (ANN) called the predictor, which generates optimal geometries based on datasets, and an appropriate number of evaluators, which monitor the predictor's training. Each evaluator assesses the predictor's results with respect to a specific criterion and returns a corresponding scalar value to indicate how well that criterion has been met. These outputs are then combined to form a single value that serves as an error function. During training, the predictor is fed randomly generated datasets, and its parameters are optimised using the error function as the objective. Compared to conventional topology optimisation, the PEN method is much faster, as the computationally intensive part is shifted to training. Once trained, the predictor can deliver geometries that are almost identical to those produced by conventional optimisers.

The publications on the research area are available here.

 

The Online Tool allows you to test the first development stage yourself.