Autonomous X-ray experiments using artificial intelligence

Artificial intelligence (AI) is revolutionizing our everyday lives, for example in speech recognition or in the future in autonomous, self-driving cars. However, AI for autonomous measurements does not yet play a major role in everyday laboratory work. Today, mostly standardized measurements are carried out according to a given pattern or experts have to think about special measurement procedures before the measurement in order to measure special samples. In this project, an artificial neural network should act as an AI in order to optimally adapt measurements to a given sample. In contrast to today`s measurements, in which the conditions such as measurement duration and measurement accuracy are usually defined before the measurement begins, the new, autonomous measurement should be controlled dynamically `live` while the data is still being received. As a result, the first partial results from the beginning of the measurement can influence the further course of the measurement, and the AI can decide faster than a human which next measurement point is the most valuable. As an example, in this project, X-ray measurements for materials research are automated by an AI to be developed. The X-ray reflectivity is measured, i.e., the X-ray beam is `specularly` reflected at the sample, whereby the angle of incidence is the same as the angle of reflection. This results in interference effects similar to a thin oil film on water that shimmers in the light. Just as with the oil and water layers in visible light, the sequence of material layers in a solar cell or battery can be measured using X-ray reflectivity, with the measurement accuracy being about the size of an atom due to the high energy of the X-ray radiation. Thanks to the autonomous AI-controlled measurement, the measurement time in an optimized X-ray measurement is firstly reduced, secondly, radiation damage in the sample is avoided and thirdly, greater accuracy is achieved by scanning the regions of interest particularly precisely. X-ray measurements are particularly suitable for control by an AI, since X-rays are dangerous for people and the measurements are therefore not carried out manually but automatically today. Therefore, this project can built on existing X-ray hardware such as motorized adjustment and scanning devices as well as sample changing robots and these can be made more powerful with AI software. In addition, X-ray measuring devices are expensive and autonomous AI control in industry and research will save valuable measuring time through faster AI tests. Due to the high speed of artificial neural networks, the information of each individual X-ray photon coming from the sample will be processed in real time at a speed that is impossible for humans. This optimizes the measurement at lightning speed, so that every X-ray photon counts and provides valuable information.

X-rays are well known from medical applications, but in research they can do much more than take images of bones. When scientists direct X-rays onto materials, they can reveal structures on the atomic scale-showing how atoms are arranged and how layers of material are built up, for example inside a solar cell. One important method is X-ray reflectometry (XRR), which is used to study ultra-thin films only a few nanometers thick. Such films are essential in semiconductors, solar cells, sensors, and even superconducting layers in quantum computers. However, these measurements are often slow and costly. At large research facilities such as synchrotron X-ray light sources, each day of measurement time can cost more than 10,000. In traditional XRR experiments, scientists record hundreds of measurement points to build a complete picture of how X-rays are reflected from a surface. This takes time, can cause beam damage to sensitive materials, and limits how quickly changing samples can be observed. This project developed a new method based on reinforcement learning, a form of artificial intelligence in which an agent learns through trial and error to make the best decisions. The agent behaves a bit like a human scientist who adjusts an experiment while watching the results come in. After each new X-ray measurement, it evaluates the information and decides in real time where to measure next. Unlike humans, it can think and react within milliseconds. The agent was trained on simulated data representing many possible thin-film structures. This allows it to start experiments with prior knowledge about good measurement strategies and then adapt its choice of measurement positions to each specific sample as real data are collected. The result is a measurement process that becomes smarter and more efficient with every data point. Using this approach, the same film thickness and structural information can be obtained with only about one third of the usual data points. For the same number of measurements, the accuracy improves by up to a factor of two. This not only makes experiments faster but also reduces beam damage, energy use at light sources, and costs for valuable beamtime. It also opens the door to real-time studies of fast, time-dependent processes, such as how thin films grow atom by atom. The project's data are published openly on Zenodo, following the Open Reflectometry Standards Organisation (ORSO) format, ensuring that others can reuse them. The success of this research has already led to a follow-up project in collaboration with the European Synchrotron Radiation Facility (ESRF)-the world's first fourth-generation light source-to implement and further develop these AI-based measurement strategies in real experiments.