Research focus

Pillar 1: Machine Learning for Modelling Technology  

The study and optimisation of technical processes is often based on differential equations which represent our mechanistic understanding of the system at hand. In most cases analytical solutions are out of reach so that numerical methods have to be applied for creating a digital twin as a faithful representation of the actual processes. This requires modelling highly complex processes which is typically very time and energy consuming. An alternative path can be taken with modern machine learning techniques. This kind of modelling is data driven and replaces the mechanistic description entirely or in parts by a black-box description of the technical processes. It typically leads to a computational simplification but creates challenges due to the loss of transparency and interpretability of the processes. Physics-informed machine learning or explainable AI address some of these challenges.

Pillar 2: Machine Learning for Controlling Technology  

Technical systems often consist of physical as well as digital software components. The field of cyber-physical systems appreciates this fact and explores its implications for the development of technology. The components of a cyber-physical system typically interact through sensors and actuators and thereby generate highly complex processes. Equipped with machine learning algorithms, these processes can be  shaped towards desired functions, which potentially incorporate synergistic interactions with the user. This kind of control of technical processes highlights a close connection between cyber-physical systems, robotics, and embodied cognitive systems. In order to make use of this connection, however, the notion of a body has to be extended from being connected and spatially restricted to a disconnected and spatially distributed system. Learning then relates to the field of federated machine learning and edge AI.

Pillar 3: Machine Learning for Analysing Technology   

Ultimately, only the use of technical systems reveals their functions in the real world which ideally coincide with those they were engineered for. Even the best possible digital twin or the deepest neural network provides only an approximation of the actual processes, which can lead to unforeseen negative  outcomes. Here, machine learning methods can be applied to evaluate the use of technical systems in the real world, potentially leading to required adjustments. This evaluation can range from the detection of anomalies in products to the quantification of the impact of technology on the human society. The analysis of processes that involve the interaction of humans with machines can reveal ethical challenges that have to be addressed via regulation or adjustment of technology. The lack of transparency and interpretability of machine learning systems constitutes one of these challenges, which could cause difficulties in assigning or determining responsibility for AI-mediated outcomes.

Mathematical Foundations

The MLE initiative develops the conceptual and mathematical foundations for the application of machine learning methods along the three outlined conceptual pillars. One main target is to study the benefit of machine learning methods in comparison with classical methods. Under which conditions, for instance, does a black-box model of a process provide better predictions than a corresponding detailed mechanistic model? Which one of the two ways of modelling is energetically favourable?

Applications

The MLE initiative uses machine learning methods within application fields along the three outlined conceptual pillars. The applications are, on the one hand, motivated and ideally supported by theoretical work. On the other hand, their results give rise to new theoretical questions. This interplay of basic research and applications represents an exploration-exploitation loop, ultimately leading to the efficient use of machine learning in engineering.