Designing alloy from 100% scrap with multi-modal Artificial Intelligence

Abstract Producing alloys from recycled metals can largely reduce the energy consumption, greenhouse gas emissions and environmental damage caused by traditional synthesis starting from ore mining. However, the accumulation of impurities when using a high proportion of recycled metal often leads to the formation of unwanted harmful intermetallic phases and thus deteriorate the performance. Conventional synthesis requires resource-intensive virgin metals to control the overall impurity levels. However, this empirical-driven approach with a low tolerance for impurities often fails to tailor the microstructure and produces alloys with reduced mechanical properties and high cost. Here, we propose a pipeline of Designing Alloys using 100% Scraps by multi-modal Artificial Intelligence. Such design tolerates high impurity levels while enhancing mechanical performance using 100\% recycled metals. The design principles are explored by learning from a multi-modal dataset consisting of the composition, processing, and microstructure of ductile irons, a showcase material with a high recycling rate. Specifically, we integrate neural networks, tree optimization to optimize the mechanical properties of ductile irons containing multiple types of impurity elements in high content and introduces a new method to handle the extremely imbalanced data distribution induced by human bias. Compared to the conventional design, our pipeline promotes the recyclability to 100%, reduces the manufacturing cost by 8.1%, improves the tensile strength up to 73.2% and total elongation up to 15.2%, and addresses the strength-ductility trade-off. The systematic experimental characterizations show that the superior mechanical performance derives from synergistic strengthening and damage-tolerant of the well-tailored complex microstructure across multiple length scales. Such counterintuitive strategy leveraging high impurity content as additional design freedom, yet demands precise tuning of chemistry and process parameters beyond human expertise but well-suited to machine learning. This sustainable alloy design strategy provides a universal research approach toward high-value utilization of recycled metallic materials.