Physics
The interdisciplinary research area of physics at the Lamarr Institute aims to improve our understanding of nature by using advanced mathematical and Machine Learning methods. Research at the Lamarr Institute develops application-oriented Machine Learning (ML) and Artificial Intelligence (AI) algorithms at the highest quality level, with outstanding performance, trustworthiness, and resource efficiency. The three universal dimensions of a problem – data, context, and domain knowledge – are to be considered in every solution approach under the heading of “Triangular AI”.
Understanding the Physical World
For various reasons, applications in physics form an excellent background for further developing Machine Learning and Artificial Intelligence techniques. Perhaps the most important point is that in physics, many years of work are usually invested in setting up the numerical representations of theories and models, the so-called Monte Carlo Simulation algorithms, to describe the measurement processes in an experiment. This involves calculating the probabilities that link the classifications or regressions sought in physics with the measured data, for example, the signals detected by a camera. Therefore, a highly accurate virtual reality exists that can as well be used to optimize the experimental setup as to train and evaluate all Machine Learning algorithms. Thus, simulation techniques and Machine Learning complement each other to help us better understand the physical world.
Optimal Use of Data from Experiments
The adaptation of Machine Learning algorithms to experimental data is a cornerstone of applications in physics. In our research at the Lamarr Institute, we intend to incorporate a priori known domain knowledge into the Machine Learning process so that valuable data is not wasted on finding known correlations again. Vice versa, we aim to extract generalizable knowledge from the machine learner for model building.
The tasks to be solved in the physical experiments involve real-time processing of gigantic amounts of data, typically several petabytes per year. This might be required in dedicated research labs such as CERN, but it is sometimes also required at exotic locations such as the geographic South Pole, which is not only theoretically but also practically extremely resource-limited in every respect.
To make efficient use of the very valuable experimental data, the algorithms developed must meet the highest quality standards in their prediction accuracy while at the same time estimating the uncertainty to which the predictions are subject. In this approach, it is less a question of whether an algorithm is trustworthy for answering a given question – what can be shown – but rather which quality criteria it satisfies to which degree.
Interdisciplinary Collaboration
In the interdisciplinary research area of physics, we aim to further explore mathematical methods for the simulation of physical models and measurement procedures. We will continue to investigate the use of Machine Learning algorithms for optimal, resource-efficient data acquisition and analysis. These exemplary applications shall become crystallization points for cooperation between physicists and data scientists. Moreover, these techniques and procedures also form the foundation for developing Machine Learning algorithms and solutions based on Artificial Intelligence for industrial and business applications.
