AI Trends

Wang Jinqiao of CASIA's Zidong Taichu: Multimodal Large Models Accelerate Factory Adoption, Ushering Manufacturing into the “AI-Native” Era

Securities TimesUpdated Nov 19 09:14

AI summary

Wang Jinqiao of CASIA's Zidong Taichu said that multimodal large models are rapidly entering factories, driving manufacturing from “intelligentization” toward “AI-native manufacturing.” Its scientific foundation model can simulate flow fields in 10 seconds, with companies’ internal tests showing errors within 5% and efficiency gains of three to ten times; WISER’s zero-shot detection performance improved by 57%. The related solutions remain at the engineering validation and large-scale rollout stage.

Body · AI translation

“Manufacturing competition is, on the surface, a contest of equipment, but at a deeper level, it is a contest of intelligence density and intelligence sovereignty.” On November 18, at the 2026 Enterprise Management Innovation Conference, Wang Jinqiao, deputy chief engineer at the Institute of Automation, Chinese Academy of Sciences and chairman of Zhongke Zidong Taichu, said that manufacturing is transitioning from “intelligent manufacturing” to “AI-native manufacturing,” with large models gradually becoming factories’ unified cognitive hub. The focus of competition in manufacturing has moved beyond a race for efficiency toward a contest over intelligence density and intelligence autonomy.

Wang first defined the broader context for AI-enabled manufacturing: Having progressed through mechanization, electrification and digitalization, manufacturing is now entering a new stage centered on intelligence. The industry as a whole is currently transitioning from “intelligent manufacturing” to “AI-native manufacturing.” The former can perform perception and recognition, but execution still depends on people; the latter requires systems to autonomously orchestrate processes and schedule equipment, with people providing confirmation only at critical junctures.

The driving force behind this transformation is a structural change in the factors of production. Wang noted that in the intelligent era, intelligence can be supplied, deployed and metered at scale like electricity. In the age of the token economy, the core question for manufacturers is no longer simply whether they can use AI, but whether they can embed intelligence into R&D, production and management processes at an affordable cost.

He said frankly that manufacturing currently faces four practical bottlenecks: multiple isolated knowledge systems on the design side, production equipment that can only passively receive instructions, quality-control decisions that rely heavily on subjective human experience, and a proliferation of maintenance systems with fragmented data.

Across manufacturing sites, multimodal large models are being deployed experimentally throughout design, operations, quality inspection and maintenance workflows, though the relevant solutions remain at the stages of engineering validation and large-scale rollout.

In process design, Zhongke Zidong Taichu is using its independently developed foundational model for science to explore an AI4S approach that replaces computation with learning. According to Wang, optimizing aircraft shapes through conventional CFD simulation often takes several weeks, while the model can predict parameters in milliseconds and simulate a flow field in 10 seconds. The company’s own tests showed prediction errors below 5% and a three- to tenfold increase in solution efficiency. This tiered simulation solution, combining rapid AI-based preliminary screening with precise final CFD verification, is being validated in scenarios involving aircraft, high-speed trains and wind turbines. The model can also interpret engineering drawings, recognize the engineering intent they embody, support automated drawing reviews and work with CAD agents to produce 3D models.

Deployment in production is focused on embodied intelligence. Wang believes that embodied intelligence in industrial settings is not simply a matter of adding vision modules to robots; its essence is to build an autonomous closed loop adapted to the physical environment. In his view, the focus of industry competition is shifting: “The industry’s competitive barrier has moved from algorithmic advantages alone to the ability to maintain a sustainable data loop. The essence of a training ground is to transform the process expertise accumulated by veteran technicians into reproducible data assets.” A weld-seam tracking solution jointly developed with HGTECH addresses the challenges of intense arc light and smoke interference in welding scenarios. The companies reported a recognition rate above 95% and millisecond-level correction. Its core algorithms were developed in-house and could replace the expensive imported solutions used previously.

Defect inspection has long been constrained by the scarcity of industrial defect samples, prompting Zhongke Zidong Taichu to adopt an unsupervised anomaly-detection approach. The team’s algorithms have won competitions at CVPR 2024 and ECCV 2024. According to company test results, its WISER method improved zero-shot detection performance by 57% and does not require retraining for new products. In theory, this feature can reduce data-collection costs when production lines switch products and mitigate losses caused by shutdowns for training. The technology has already given rise to inspection equipment for textile inspection, supercrystalline panels and steel-pipe defect detection.

For maintenance, the team has built an integrated air-ground-site inspection system on the AgentOS task-scheduling platform. Wang noted that the challenge in industrial inspection is not the number of sensors deployed, but matching the right carrier to each inspection objective: drones handle elevated locations, robot dogs inspect equipment at close range, and integrated radar-vision devices cover enclosed areas. The project follows an implementation approach of “first verifying coverage, then configuring rules,” identifying inspection blind spots in digital space before deploying control strategies. Once an anomaly is detected, AgentOS can dispatch multiple types of inspection carriers to verify it, conduct cross-analysis and automatically generate work orders, addressing the shortcomings of conventional inspections that depend heavily on manual on-site handling.

Industrial large models also face practical barriers. Wang said that in real-world engineering, the greatest pain point is often not an insufficient number of systems but the absence of a unified cognitive hub, leaving heterogeneous systems isolated from one another.

Looking ahead, Wang summarized four core judgments: the cost of producing intelligence will determine industrial competitiveness; competitive barriers will shift from individual algorithms to data loops; autonomy and control over core technologies constitute a strategic bottom line; and industrial intelligence must achieve coordination between algorithms and equipment. He emphasized that large models can only deliver genuine value after being deeply adapted to existing industrial-site workflows. The deeper logic of manufacturing competition is shifting from the early contest over efficiency toward a contest over intelligence density and intelligence sovereignty.

Source

Securities Times