When artificial intelligence delves into the arrangement of metal atoms and performance prediction, traditional materials science is undergoing a quiet yet profound revolution. On January 15th, Shanghai Jiao Tong University and Xiaomi Group jointly announced the world's first multi-agent AI R&D platform for lightweight alloys, with "DeepLight large model + AgentMat agent" as its core architecture. It has achieved full-chain intelligence from composition design, process optimization to performance prediction for the first time, compressing the R&D cycle of lightweight alloys from months or even years to one-tenth.

This platform directly addresses industry pain points: lightweight alloys, as key structural materials in new energy vehicles, aerospace, and high-end consumer electronics, have long been constrained by challenges such as high-dimensional parameter spaces, nonlinear physical mechanisms, and high experimental trial-and-error costs. Now, the DeepLight large model integrates materials science literature, experimental databases, and first-principles calculation results, building a unified cognitive framework covering thermodynamics, mechanics, corrosion resistance, and other dimensions, effectively overcoming the bottlenecks of traditional methods in performance prediction and mechanism reasoning.

The more critical breakthrough lies in the AgentMat agent framework. Rather than relying on a single AI model working alone, this system deploys multiple specialized agents—such as "composition design agent," "process optimization agent," and "failure analysis agent"—which can autonomously negotiate, divide tasks, and provide iterative feedback, simulating the collaborative R&D process of human expert teams. For example, when a user proposes "developing a high-strength heat-resistant magnesium alloy for electric vehicle motor housings," the system can automatically decompose the task, parallelly call different agents, generate candidate formulas, recommend heat treatment paths, and predict service life within a few hours, all without manual intervention throughout the process.

To measure technological progress, both parties also launched the world's first dedicated large model evaluation benchmark for lightweight alloys—LightAlloy-Bench—which includes 12 core tasks such as phase diagram prediction, mechanical property regression, and process window optimization, providing the industry with a standardized capability benchmark.

This collaboration deeply integrates Shanghai Jiao Tong University's decades of experience in fundamental research and engineering applications of lightweight alloys, as well as Xiaomi's technical advantages in large model training, agent architecture, and high-performance computing. As Xiaomi accelerates its layout in the intelligent electric vehicle field, this platform is expected to first empower its next-generation body and powertrain system lightweight design, while also being open to the supply chain, promoting China's independent innovation in the strategic emerging industry of high-end new materials.

When AI can not only write code, draw pictures, and order takeout, but also "design metals," the paradigm shift in materials science has already begun—perhaps future new materials will no longer be born in crucibles in laboratories, but first take shape in the dialogues of intelligent agents.