A new algorithm mines material libraries up to four times faster than before. It is based on machine learning.
Researchers are working flat out to find new materials for future technologies on which the energy transition depends, for example as electrocatalysts. Due to their versatile properties, materials consisting of five or more elements are of particular interest. With about 50 usable elements of the periodic table, there is an almost infinite number of possible materials. Felix Thelen from the Chair of New Materials and Interfaces at the Ruhr University Bochum, headed by Alfred Ludwig, has developed an algorithm that can investigate material candidates four times faster than before. This is made possible by the concept of Active Learning, a subfield of machine learning. The research team reports in the journal Digital Discovery on September 19, 2023 .
Days or weeks to measure a sample
Despite highly specialized methods that can produce a range of materials in parallel on a single sample and then measure them automatically, every minute counts in their analysis - because days or weeks can pass before the examination of a sample is complete. The new algorithm can be integrated into existing measuring instruments and can increase their efficiency many times over.
Measuring instrument searches for measuring points itself
By using Active Learning, a measuring instrument is able to independently select the next measuring point on a sample, based on the information already available about the material," explains Felix Thelen, developer of the autonomous measuring program. Point by point, this refines a mathematical model about the measured material property until the accuracy is sufficient. Then the measurement can be stopped - and the results at the remaining measurement points are predicted by the generated model.
The Bochum research team was able to prove the algorithm’s functionality using the example of electrical resistance measurements on ten investigated material libraries. Our real work is just beginning here," says Felix Thelen, "because in materials research there are far more complex measurement methods than resistance measurement that also need to be optimized. In cooperation with the manufacturers of the instruments, solutions must now be developed that enable the integration of such active learning algorithms.
Felix Thelen, Lars Banko, Rico Zehl, Sabrina Baha, Alfred Ludwig: Speeding up high-throughput characterization of materials libraries by active learning: autonomous electrical resistance measurements, in: Digital Discovery, 2023, DOI: 10.1039/D3DD00125C
