From September 16 to 18, the 2026 Beijing International Microelectronics Symposium & IC WORLD Conference, held for twelve consecutive sessions, took place at the Beiren Yichuang International Convention and Exhibition Center in Beijing Economic and Technological Development Area. Centered on the theme Harness Intelligence to Forge the Chip Road, Strengthen Chips to Write a New Chapter, the conference consists of an academic forum and an exposition. Nearly 200 organizations across the integrated circuit industrial chain participated in the exhibition. Centered on wafer manufacturing and packaging & testing, the two core industrial segments, the event gathered upstream and downstream resources including IC design, equipment, materials, components and applications in Beijing. Meanwhile, it showcased the coordinated industrial ecosystem across the Beijing-Tianjin-Hebei region.
Shanghai Glorysoft Co., Ltd. (hereinafter referred to as “Glorysoft”) was invited to participate in the conference. Wang Hui, Solution Director of Glorysoft, delivered a keynote speech titled Cutting-edge AI Application Practices – Vision and Data for Semiconductor CIM at the Special Forum on New Ecosystem of Equipment and Processes in the Semiconductor Industrial Chain. Focusing on two major pillars — AI vision and data applications, the speech systematically presented Glorysoft’s technical roadmap and implementation results for intelligent upgrading of semiconductor CIM.
With years of experience in pan-semiconductor informatization, Wang Hui has led overall IT planning and full-stack CIM implementation for leading domestic enterprises specializing in 12-inch large silicon wafers and high-end electronic materials, boasting abundant hands-on expertise. His sharing covers not only trend analysis but also addresses real pain points on production lines.
Wang Hui, Solution Director of Glorysoft
01
AI Vision: From Point Identification to End-to-End Closed Loop
Wang Hui noted that as advanced processes and heterogeneous integration raise manufacturing complexity, traditional CIM systems face growing limitations including data silos, blind spots in visual inspection and delayed response. Especially under three high-growth tracks — AI computing chips, advanced packaging, and optoelectronic interconnection & CPO, there is an urgent need to restructure CIM capabilities.
Take optoelectronic interconnection and CPO as an example. Efficient coupling of optical chips, optical fibers and lasers constitutes the core process, whose manufacturing logic differs greatly from logic and memory chips. Optoelectronic manufacturing is extremely sensitive to compound semiconductor materials and requires ultra-high precision for coupling alignment. Even micron-scale defects can directly undermine the stability of AI computing interconnection.
Confronted with such special processes, traditional CIM often struggles with process parameter fluctuation and interception of micron-level defects. Through tight integration between LOFA AI Vision and CIM systems, Glorysoft can automatically trigger defect interception once micron-scale defects are detected, and feed the results back to the process side for parameter tuning. This drives the evolution of CIM from a “recording-oriented” system toward an “intelligent control-oriented” system, supporting the manufacturing of high-speed optical communication modules such as CPO.
For full-link AI vision inspection of silicon wafers, algorithms powered by LOFA AI Vision have been deployed across core processes including crystal pulling, forming, cleaning, epitaxy and polishing. The solution achieves accurate identification of morphologically similar defects in EBFIS and grinding AOI, and connects directly to the MES system to automatically block NG products and conduct root cause traceability.
For automation retrofitting of core tools, Glorysoft leverages LOFA remote control, Bots and AI vision co-automation to realize automatic recipe setup, defect localization and automatic metrology. Among these, OM tools complete full automation via a combined hardware-software solution, enabling intelligent classification of macro and micro defects and automatic linkage with shipping systems, effectively removing the final manual check before shipment. In addition, AOI re-inspection has replaced manual work, and pan-industrial inspection covers digital meters, analog gauges, scene recognition and other scenarios.
Traditional Algorithm Solution vs. AI Model Vision Solution
02
Data Application: From “Being Recorded” to “Enabling Decision-Making”
On data applications, Wang Hui highlighted the closed-loop workflow of Glorysoft’s data management solution: ingestion, governance, analysis and application, built upon nearly 20 years of accumulated semiconductor data models. He cited the crystal pulling scenario as an example. In the past, furnace tuning relied on senior technicians’ experience, and process fluctuations were hard to quantify. Today, data converts empirical know-how into executable strategies. AI models lift the single crystal yield from 57.8% with manual operation to 85%. Furthermore, for yield analysis and other scenarios, engineers only need to ask questions in natural language; the system can automatically complete root cause attribution and output conclusions and countermeasures.
Wang Hui summarized Glorysoft’s evolution path for semiconductor data applications into the DTT framework: Data (Trusted Data), Training (Trusted Models), Thinking (Decision Closed Loop). From trusted data to trusted models and then to decision closed loops, the framework ultimately builds an integrated dark factory assistant for semiconductor fabs.
The Evolution of AI Data Application: DTT
From full-link silicon wafer inspection and core tool automation, to data-driven yield analysis and intelligent decision-making, Glorysoft’s presentation at IC WORLD 2026 demonstrated real-world practices where AI moves beyond technical demos and into production lines. When AI is no longer merely a concept but embedded into daily workflows of inspection, tool operation and yield analysis, the next stage of semiconductor smart manufacturing will shift from “seeing” to “predicting”.
