The mechanical performance of a metallic alloy is dictated by its microstructure and texture—the grain size, shape, and crystal orientation. These structural attributes stem from the alloy composition together with processing parameters. Conventional alloy development follows the forward chain composition → structure → property, while the reverse problem—inferring a recipe from an optimized structure—still relies on expert intuition. We test whether this reverse step can be learned from data. Our in‑house dataset comprises 107 magnesium alloy extrusion cases across 14 alloys, each with an optical micrograph and an X‑ray texture measurement. Three descriptors are compared: conventional grain and texture statistics, a vision embedding from a pretrained image encoder, and a graph neural network built on the grain network. Each descriptor feeds two prediction heads: Task A predicts the alloy composition given the process, and Task B predicts the process parameters given the composition. Under 5‑fold cross‑validation, conventional statistics identify the correct alloy for 65% of held‑out cases, far above the 17% baseline of always guessing the most common alloy, while learned embeddings stay below 30%. Process parameters are recoverable but noisier; adding microstructure roughly halves the temperature error compared with using composition alone. Because only a few casting alloys and press settings were used, the outputs are essentially discrete. Heads that select from these known ordered options outperform heads that regress free values.
Review