DeepLight & AgentMat: Xiaomi and SJTU Launch AI Platform for Light Alloy Design

Lightweight alloys are critical for aerospace, automotive and consumer electronics, yet their development suffers from complex composition‑process‑performance coupling, long experimental cycles, and a lack of unified evaluation metrics.

To address these challenges, Xiaomi and Shanghai Jiao Tong University released the DeepLight large‑model, which is driven by domain‑specific reasoning. The team built a chain‑reasoning data system that integrates material composition, micro‑structure, phase‑transition behavior, micro‑evolution, and mechanical performance. Supervised fine‑tuning combined with reinforcement learning enables the model to construct cross‑scale inference paths and capture highly non‑linear structure‑property relationships, markedly improving inference quality for lightweight alloys.

Using DeepLight’s reasoning capability, the researchers successfully created an ultra‑corrosion‑resistant magnesium alloy whose corrosion resistance matches the best levels reported in public literature, opening new possibilities for high‑demand applications such as automotive and aerospace components.

The accompanying AgentMat multi‑agent framework acts as an efficient hub that links the entire workflow—from knowledge acquisition, data analysis, solution generation, tool invocation, verification, to expert collaboration. By decomposing the process into specialized agents, the system automates hand‑offs and achieves a full‑process design for magnesium alloys in about one hour, delivering a ten‑fold efficiency gain over traditional methods.

AgentMat has already been deployed on an autonomous alloy‑experiment platform, forming an AI‑driven closed‑loop of “model prediction → experimental validation → data feedback” that can be rapidly adapted to automotive, aerospace and other production scenarios.

The LightAlloy‑Bench benchmark, built by the Shanghai Jiao Tong University Light‑Alloy Center, is the first domestic evaluation suite covering the full spectrum of magnesium and aluminum alloys. It aggregates decades of research data, including contributions from the National Engineering Research Center for Precision Forming of Light Alloys led by Academician Ding Wenjiang and Professor Zeng Xiaoqin, and nearly 20 years of publicly available literature, forming the most comprehensive knowledge chain in the field.

On LightAlloy‑Bench, DeepLight achieves the best overall scores, surpassing the general‑purpose large language models DeepSeek‑V3 and GPT‑4o by roughly 20 % in tasks such as alloy knowledge understanding, mechanism inference, and performance prediction.

The initiative reflects Xiaomi’s sustained investment in AI research since 2016, including breakthroughs like the Titan alloy, and the Shanghai Jiao Tong University Light‑Alloy Center’s 60+ national projects, 15 national science‑technology awards, and over 500 publications, positioning the joint lab as a leading force in intelligent material R&D.

Looking ahead, the partners plan to deepen cooperation, extend DeepLight to new alloy families such as titanium and high‑entropy alloys, refine the multi‑agent collaboration mechanisms, and accelerate the transition to an AI‑plus, human‑machine co‑design paradigm that supports strategic industries like aerospace and intelligent vehicles.