The recycling of metal scrap as secondary raw materials is the safest, most ecological and most economical form of raw material supply. Although metals can in principle be recycled an infinite number of times, remelting without loss of quality is only possible if the composition can be accurately recorded before remelting and is melted down by type. Determining the elemental composition of scrap non-destructively in real time would make it possible to optimally control input flows of the recycling process for the first time. As a result, the scrap content can be further increased and high-quality alloys can be produced in a targeted manner. In copper and aluminium production, there is great interest in the elemental analysis of recycling materials in real time in order to classify the metallic secondary raw materials according to existing standards and regulations and to recycle them according to type. A satisfactory metrological solution is currently not available for either copper or aluminium production and is to be developed within the framework of MetalClass.

 
Within MetalClass, AI-based evaluation algorithms are being developed for real-time classification of metallic scrap. The aim of the project is to develop a measurement method in which AI methods are used to evaluate measurement data from PGNAA technology in seconds or even fractions of a second in order to classify measured scrap parts according to their composition.
Artificial intelligence (AI) and machine learning methods are used for this purpose, in which knowledge and techniques from computer science, statistics, mathematics and application knowledge are combined. The aim is to draw practical conclusions from data, in particular by using characteristics of the measurement data as application knowledge in the procedure to be developed. Frequently used models are deep neural networks, which are particularly well suited for classification tasks. The real-time capability of such a method represents a further challenge. The reduction of the measurement time is particularly critical, because shorter measurement times lead to highly noisy data. Therefore, models are being developed within the project that can successfully carry out a classification in the shortest possible time despite highly noisy data.

Experimental measurement data are collected by a demonstrator facility in AiNT's technicl center, using innovative detector systems. AiNT develops and tests measuring systems for elemental analysis based on prompt gamma neutron activation analysis (PGNAA). These measuring systems enable non-destructive elemental analysis of a wide variety of materials and differ from existing methods in that the material to be measured is analysed as a whole and without sample preparation. During a measurement, a gamma spectrum is recorded, the evaluation of which enables the complete elemental composition of the material to be determined. These data are analysed by means of AI methods in order to make a classification.