Today’s societal challenges require rapid response and smart materials solutions in almost all technical areas. One major challenge is driven by electrical vehicles replacing combustion engines: It is expected that the amount of aluminium scrap alloys from conventional automotive sector increases tremendously in the near-future. Such alloys are very complex (often > 10 alloying elements) and are highly specialised for their current application. Consequently, to keep the metals in the materials life cycle, fast design of new alloys with different materials properties is mandatory.

In our work, we demonstrate how machine learning can be applied to boost the exploration of the whole alloy design space, the prediction of new designs, and the discovery and optimisation of alloys. First, we generated a dataset of $\approx$ 1000 compositions and calculated > 60 materials properties such as phase fractions, solidification temperatures, and (thermo-) mechanical properties using the CALculation of PHAse Diagrams (CALPHAD) method and thermo-physical modelling. Secondly, we trained random forest regression models to predict these properties as a function of their composition. After training and validation of the models, we integrated them within a multi-objective evolutionary optimisation. The optimisation gives a number of possible candidate compositions fulfilling specific design criteria. Finally, these optimal designs were tested again using CALPHAD and thermo-physical modelling.

We show that the total time needed to find promising candidate compositions for a set of target properties can be reduced to a minimum using the combination of CALPHAD, thermo-physical modelling, and machine learning. Moreover, we want to show our lessons learned about how a typical data science workflow, i.e. OSEMN can be applied on a materials science specific scenario for rapid alloy development.