MATAI: A Unified Interactive Platform for AI-Driven Alloy Discovery

Yixuan Li, Yanchen Deng, Chendong Zhao, Penghui Yang, Jianguo Huang, Yuhao Lu, yushan xiao, Dong Xuyu, Zhonghan Zhang, Cuntai Guan, Bijun Tang, Xinrun Wang, Zheng Liu, Bo An

Track: Track 1: Original Research/Position/Education/Attention Track

Abstract: AI-driven materials discovery is increasingly bottlenecked not by individual models but by the fragmentation of the surrounding tooling: composition–property databases, property predictors, and constrained optimisers typically live in disjoint projects, forcing researchers to hand-stitch scripts, file formats, and provenance. We present MATAI, a unified, browser-based platform that integrates visual data analytics, composition-aware property prediction, and simulated-annealing-based constrained alloy design into a single closed-loop workflow. A conversational layer exposes the same prediction and design engines through natural-language commands, and a document-ingestion pipeline converts unstructured literature into structured composition–property records. Every prediction is paired with dual-provenance neighbour views and screened by an inline LLM feasibility gate that checks classical metallurgical heuristics before inference. By coupling domain-aware visualisation with interactive inference and design, MATAI lowers the engineering burden of applying AI to materials science and provides a reusable blueprint for closed-loop AI-for-Science systems.

Keywords: AI for materials science, alloy design, interactive visualization, constrained optimization, property prediction, closed-loop discovery, AI-for-Science platform

Submission Number: 45