Artificial intelligence and FORC analysis in Carbide production

Cemented carbides today find a wide range of applications and are tailor-made according to their tasks. Carbide can only be produced by powder metallurgy. The mechanical properties are determined by the composition and granularity of the carbide material, metal binder and any additives. In addition to hardness testing, density testing, structural analysis and porosity tests, non-destructive measurements of magnetization and coercive force are routinely used for quality control according to DIN-ISO 3326. These magnetic properties provide information on the structure, composition and contamination in the sintered state. Although the evaluation draws on a great deal of experience, the measurement results in binary or ternary systems with complex manufacturing processes are often not clearly interpretable.

This project aims to make the magnetic characterization of cemented carbides and their conclusions on structural and mechanical properties more accurate. In addition to traditional measurements of M(H)-hysteresis, modern methods such as First-Order-Reversal-Curve (FORC) analysis and Artificial Intelligence (AI) for FORC diagrams will be added, which will make the interpretation more quantitative and clearer.

Today's magnetometers allow FORC measurements to be done within a reasonable time frame of a few minutes. These provide additional information about phase formation and impurities in the powder and sintered state. Despite vast experience, FORC diagrams are not easy to interpret. A so far unused approach should help. We want to interpret the FORC diagrams using Artificial Intelligence. After a learning phase, we expect quantitative statements on structure, composition, phase formation and contamination, as well as on mechanical properties such as hardness, tensile strength, etc. The above improvements are supported by (micro)magnetic simulations. On the one hand, these can calculate FORC diagrams of model systems and make them available to the deep-learning algorithm (or similar: Random Forest for example), and on the other hand, experimental data (M(H) and FORC) can be physically interpreted.