Training workers for the nylon chemical industry has always been a difficult balancing act. The industry chain stretches from raw material supply through polymerization, spinning, quality testing, safety management, and equipment maintenance, and its plants run continuous processes at high temperatures and pressures with hazardous media that make real-world student training costly and risky. A new open-access study published in Discover Artificial Intelligence proposes a data-driven answer: a framework called NylonChem-IEKG that fuses an industry-chain knowledge graph with a digital twin training platform to connect what factories need, what courses teach, and how well students learn, all inside one computable system.
The framework's semantic core is a knowledge graph that converts scattered information from corporate reports, job postings, curriculum standards, training projects, certification systems, and employment feedback into a structured network of entities and relationships. Nodes represent enterprises, industrial chain segments, positions, skills, courses, training tasks, certificates, and equipment, while edges encode relations such as "requires," "supports," "matches," and "feeds back." A polymer process technician who requires process parameter control capability, or a digital twin training module that supports hazardous chemical emergency training, becomes a machine-readable triple. The graph is divided into four subgraphs: IndustrialKG for job competency demands, EduKG for curriculum support, TwinKG for training effects, and GovKG for governance improvement.
Building that graph required substantial data engineering. The experimental dataset comprised 5,420 records drawn from enterprise recruitment information, nylon chemical process documents, curriculum standards and teaching resources, digital twin training logs, and collaborative governance reports. After entity recognition and relation extraction, the researchers generated 18,756 triples, with annotation performed by two chemical engineering education experts and verified using Cohen's Kappa coefficient of 0.91, indicating excellent agreement.
Duplicate entities, inconsistent relations, and incomplete records were removed through automatic validation and manual quality inspection before the data were organized into the four subgraphs using a unified ontology and entity alignment strategy. The digital twin layer is the framework's practical carrier. Internet of Things devices collect real or simulated process data, including temperature, pressure, flow rate, energy consumption, equipment status, alarm events, and operator behavior.
Virtual objects encapsulate this data into callable functional units that interact with the twin platform through APIs, while the knowledge graph supplies semantic constraints linking process knowledge, job skills, course modules, and safety regulations. The prediction model itself uses a Graph Attention Network architecture, in which neighboring nodes receive adaptive attention weights to capture spatial and semantic dependencies between production units, before two fully connected hidden layers with ReLU activation learn high-level process characteristics and the output layer predicts equipment status, process parameters, production quality, and safety risk indicators. Safety is treated with unusual rigor in the optimization objective.
Hard constraints represent mandatory operational limits such as maximum reactor temperature, pressure thresholds, and hazardous material concentration limits, and any violation incurs a large penalty that prevents infeasible operating states during optimization. Soft constraints quantify deviations from recommended process settings, including energy consumption, production efficiency, and process stability, and are incorporated as weighted penalty terms. The weighting coefficients were set through grid search at 0.50, 0.30, and 0.20, balancing process-state consistency, prediction accuracy, and safety constraint optimization.
The result is a training environment in which students can practice reactor operation, fault diagnosis, emergency response, and quality control without ever entering a high-risk production site. Governance is handled by a fuzzy collaborative network in which nodes represent government departments, industrial parks, leading enterprises, vocational colleges, digital platforms, industry associations, and employment feedback systems. Because the strength of collaboration between stakeholders is inherently uncertain, edge weights are expressed as trapezoidal fuzzy numbers estimated from a combination of domain expert evaluation and observed collaboration indicators such as enterprise participation frequency, curriculum co-development activity, and data-sharing intensity.
Three experts from nylon chemical engineering, vocational education, and industrial management independently evaluated each stakeholder pair, and a centroid method converts the fuzzy weights into clear values for computing co-centrality measures and identifying the most efficient resource transfer paths between government, enterprises, educational institutions, and platforms. On top of these layers sits a prompt-enhanced knowledge graph reasoning mechanism. For any user query, whether a company asking which courses support a polymerization position, a teacher asking which training tasks build emergency response skills, or a government agency asking which enterprises have weak collaborative ties, the system first extracts a k-hop subgraph related to the query, then encodes entities, relationships, and query semantics through a query-aware graph neural message passing mechanism with attention weights.
Domain-specific reasoning constraints covering process sequence dependencies, equipment operational limits, chemical safety regulations, and curriculum-competency mappings are built into subgraph extraction and message propagation, so candidate entities that violate process logic or safety requirements receive lower reasoning scores. The selected configuration used an embedding dimension of 128, eight attention heads, and a three-hop subgraph. The experimental evaluation compared NylonChem-IEKG against three baselines: a conventional industry-education integration management system relying on static spreadsheets and manual rules, a knowledge graph system containing only the industrial chain structure, and a standalone digital twin training platform.
As input size grew from 100 to 500 records, the proposed framework maintained the lowest response time, averaging approximately 9 seconds at 100 records and approximately 36 seconds at 500 records. A one-way analysis of variance revealed a statistically significant difference among methods (F = 29.84, p < 0.001), and post-hoc Tukey's HSD tests confirmed that NylonChem-IEKG significantly outperformed all three baselines in every pairwise comparison. An ablation study isolated the contribution of each component.
Adding the knowledge graph to a baseline graph neural network improved semantic reasoning and curriculum-job matching, incorporating the digital twin further enhanced competency-oriented training recommendations, and the fuzzy collaborative governance module strengthened multi-stakeholder collaboration through uncertainty-aware relationship modeling. The full framework with prompt-enhanced reasoning achieved the best results, raising knowledge reasoning accuracy from 93.41 to 97.82 percent, curriculum-job matching accuracy from 92.36 to 96.94 percent, and governance performance from 91.25 to 95.68 percent, while reducing average response time to 35.08 seconds. Correlation analysis across twelve key indicators also showed strong positive associations among enterprise participation, data sharing, and governance performance, with higher data-sharing levels associated with higher collaborative governance performance across polymerization, spinning, and safety management scenarios.
The authors are candid about the study's boundaries. The validation relies primarily on simulated industrial scenarios and synthesized multi-source datasets, so the results may not fully capture the dynamic characteristics, operational uncertainties, and heterogeneous data distributions of real nylon chemical production environments. The digital twin training platform and collaborative governance mechanism have not yet been verified through large-scale classroom implementation or long-term industrial deployment.
Planned future work includes integrating real industrial IoT sensor data, conducting classroom-based teaching experiments with vocational students and enterprise mentors, and performing longitudinal industrial deployment to assess long-term impacts on enterprise participation, curriculum optimization, and governance performance. Until those studies are complete, NylonChem-IEKG stands as a carefully engineered and statistically evaluated blueprint rather than a proven deployed system, but it offers process manufacturing industries an interpretable and transferable template for closing the gap between factory floors and classrooms.
Subject of Research: A knowledge graph and digital twin framework for industry-education integration in the nylon chemical industry
Article Title: An integrated knowledge graph and digital twin framework for data intelligent industry education in the nylon chemical industry
Article References: Zhang, L. (2026). An integrated knowledge graph and digital twin framework for data intelligent industry education in the nylon chemical industry. Discover Artificial Intelligence, 6(1), Article 966. https://doi.org/10.1007/s44163-026-02048-8
Image Credits: AI Generated
DOI: 10.1007/s44163-026-02048-8
Keywords: knowledge graph, digital twin, nylon chemical industry, industry-education integration, vocational education, graph attention network, collaborative governance, prompt-enhanced reasoning, chemical process training, data sharing, fuzzy network, integrated
News Source: Courtney Benton. (October 28, 2026). Research report: An integrated knowledge graph and digital twin framework for data intelligent industry education in the nylon chemical industry. Scienmag.
