Active Matter in the Training Camp: Scientists Shed Light on the Current State of the Application of Machine Learning in Research on Active Materials

Active Matter in the Training Camp: Scientists Shed Light on the Current State of the Application of Machine Learning in Research on Active Materials

Machine learning processes have experienced a huge increase in applications in many areas in recent years due to the availability of enormous amounts of data: from classifying objects and analyzing time series to controlling computer games and vehicles. In a current review in the journal "Nature Machine Intelligence", authors from the universities in Leipzig and Gothenburg shed light on the current state of application and application possibilities of machine learning in the field of research on active materials.

Computer-generated graphics of a particle as it is also used for the machine learning experiments. You can see a polymer particle with many gold nanoparticles on the surface. Some of the gold nanoparticles are irradiated with a green laser. Photo: Prof. Dr. Frank Cichos
Prof. Dr. Frank Cichos, Photo: Swen Reichhold/Universität Leipzig

Systems that are driven by the conversion of energy are referred to as active materials. The best example of active materials are biological systems from the individual molecular motor to bacteria and cells to entire organisms and swarms of animals. Active materials also include artificial systems made of nano- and microparticles that “imitate” the function of biological systems.

To build artificial intelligent systems for future technologies, one first has to understand natural intelligent systems, some of which have gone through millions of years of evolution. "There are amazing examples in nature," says Prof. Dr. Frank Cichos from the Peter Debye Institute for Soft Matter Physics at Leipzig University. Birds like the albatross have learned to use atmospheric currents to glide. Plankton navigates in turbulent ocean currents and sperm control their motion based on noisy chemical signals that determine the microscopic world of active matter. In many cases, animals show collective behavior, they form swarms and use communication channels that allow them to change direction blazingly fast. To recognize and understand such complex processes, scientists worldwide are increasingly using machine learning methods, which are among the tools in the field of artificial intelligence.

What different approaches are currently available here and for which fields of application are they suitable? Frank Cichos with colleagues Kristian Gustavsson, Bernhard Mehlig and Giovanni Volpe from the University of Gothenburg in Sweden. The authors also show which pitfalls to watch out for. Prof. Dr. Frank Cichos has together with his colleagues Kristian Gustavsson, Bernhard Mehlig and Giovanni Volpe from the University of Gothenburg in Sweden assembled what different approaches are currently available here and for which fields of application are they suitable. The authors also show which pitfalls have to be taken into account.

Obvious applications of machine learning can be found in image analysis, i.e., in the detection of objects, for example, in microscopic images, their tracking over time and the characterization of their motion. Often, neural networks are trained with training data, which are also generated artificially or can be obtained from numerous experiments. "However, the variety of different processes that are already used in research on active materials is much greater," says Cichos. Reinforcement learning – learning through rewards – is used to explore navigation strategies in complex currents, and deep learning methods help in the search for simpler physical models for pattern formation in complex currents. While these applications are all implemented on computers, there are also attempts to use artificial active matter as neural networks.

In addition to the fundamental knowledge that can be gained about active matter with the help of machine learning, technological applications also open up. The efficient gliding in air currents by means of sensory information, as examined for birds, can serve to optimize aircraft. Understanding collective behavior in swarms could be of interest for autonomous driving and navigation strategies can help in the active transport of medication in the human body.

Prof. Dr. Frank Cichos and his workgroup Molecular Nanophotonics at Leipzig University deal with artificial active matter in the micro- and nano-range. They create artificial particles that are driven by light and study their behavior. "Among other things, we want to investigate microscopic particles that show adaptive collective behavior and learn on the smallest length scales," explains Cichos. For this purpose, the workgroup uses procedures of reinforcement learning so that active microparticles learn to explore their environment. They are also assisted by neural networks for the detection of their active particles in optical microscopy, which is superior to algorithmic methods, especially for many different particles.

From the authors' perspective, the research of active matter can also help to improve the methodology of machine learning. “Research on active matter can easily generate large, high quality data sets across many length and time scales through experiments and physical models. Based on this data, new models for machine learning can be developed,” says Frank Cichos. But not all questions of active matter have to be solved with machine learning, Cichos points out. "With all the hype about machine learning, you also have to realistically assess whether you really need such a method if you can also tackle the problem with classic methods.“

Reference:
F. Cichos, K. Gustavsson, B. Mehlig, G. Volpe
Machine Learning for Active Matter
Nat. Mach. Intell. 2, 94–103 (2020)

Source:
Translated from the Press Release of Leipzig University

Further Information:
Prof. Dr. Frank Cichos
Peter Debye Institute for Soft Matter Physics
Tel.: + 49 (0) 341 97-32 571
Email | Website

letzte Änderung: 19.05.2020