Which alloy composition, what process parameters? Inferring the recipe from optimized metallic microstructure and texture

The mechanical properties of a metallic alloy are set by its microstructure and texture: how large its grains are, what shape they take, and how their crystals are oriented. That structure is in turn set by a recipe, meaning the alloy composition together with the processing parameters used to form it. Alloy development runs this chain forwards, tuning the structure until the target property is met. Running it backwards, from an optimized structure to the recipe that would produce it, still leans on expert knowledge. We ask whether that backwards step can be learned. On an in-house experimental magnesium alloy dataset of 107 extrusion conditions across 14 alloy classes, each with an optical micrograph and an X-ray texture measurement, we compare three microstructure and texture descriptors: conventional grain and texture statistics, a vision embedding taken from a large pretrained image model, and a graph neural network trained on the network of grains and their boundaries. Each is paired with a prediction head on two tasks: Task A, the alloy composition, and Task B, the process parameters. Under 5-fold cross-validation, the conventional descriptors predict the correct alloy for 65\% of held-out conditions against a 7\% chance level, while the two learned embeddings stay below 30\%; the processing parameters are recoverable but noisier, with the exact extrusion speed named about three times more often than chance. The practical message is that on small, heterogeneous experimental datasets, matching the predictor to the discrete structure of the answer matters more than scaling up the representation.