AI-Powered Network Model Tracks How Multiple Risks Cascade Through Manufacturing

High-end equipment manufacturers face a distinctive kind of vulnerability: their risks rarely arrive one at a time. A technology blockade can coincide with a pandemic, a geopolitical conflict can overlap with a supply shortage, and when several such shocks strike simultaneously, their combined damage is far greater than the sum of their parts. A new study published in Frontiers in Physics proposes a modeling framework designed to capture exactly this coupling effect, tracing how multiple risk sources propagate together through the complex web of business activities inside a manufacturing enterprise.

The research, conducted by Chen Wei, Ping-feng Liu, and Wan-ying Rao at Wuhan University of Technology, introduces what the authors call a GERT-Choquet Fuzzy model, abbreviated GERT-CF. It combines a classical network-based risk analysis technique known as the Graphical Evaluation and Review Technique, or GERT, with the Choquet fuzzy integral, a mathematical tool for combining multiple influences in a non-additive way. The goal is to identify, in advance, which business nodes are most likely to fail when several risk sources act on a company at once.

The motivation is drawn from real events. The authors cite the case of Semiconductor Manufacturing International Corporation, or SMIC, in 2020. A United States technology blockade restricted the company’s access to extreme ultraviolet lithography equipment, while the COVID-19 pandemic simultaneously disrupted global supply chains. The combined shocks forced SMIC to suspend research on advanced chips and shift toward mature manufacturing processes, while its market share gap with Taiwan Semiconductor Manufacturing Company continued to widen. The authors present this as a clear illustration of how multi-source risks interact destructively rather than merely adding up.

Building the model begins with data. The team collected risk texts through two channels: online web crawling of prospectuses and annual reports from 208 listed Chinese companies in the high-end equipment manufacturing sector, covering industries such as rail transit, marine engineering, aviation, satellite technology, and intelligent equipment, and offline field investigations at a company affiliated with CRRC Corporation Limited. The offline material came from an internal risk-management database and semi-structured interviews with managers, technical staff, and frontline employees in procurement, production, marketing, operations and maintenance, and risk management. In total, the researchers assembled 14,376 risk paragraphs, each treated as a unit describing a specific risk factor.

Artificial intelligence then took over much of the network construction. The researchers compared four large language models, ChatGPT5.6, Gemini3.5, Grok4.5, and DeepSeekV3, on a test set of 100 risk paragraphs annotated by five enterprise experts. A relationship was counted as correct only when both business activities and the direction of the connection matched the reference annotation. ChatGPT5.6 performed best, achieving a precision of 96.09 percent, a recall of 94.62 percent, and an F1-score of 95.35 percent, ahead of Gemini3.5 at 93.33 percent, Grok4.5 at 90.91 percent, and DeepSeekV3 at 88.35 percent. The winning model was then applied to all 14,376 paragraphs, with the research team manually checking and correcting the extraction results. This yielded 47 business activities linked by 56 relationships across four business segments: procurement, production, marketing, and operations and maintenance.

To label the risks traveling along those connections, the team used a topic-modeling method called KPE-BERTopic, which automatically determines the number of topics and clusters semantically similar paragraphs. It initially generated 422 preliminary risk factors, which the five experts then reviewed, standardized, and consolidated into 75 retained factors organized into seven categories, ranging from risks arising in unstable and uncertain environments, such as public health emergencies, natural disasters, and geopolitical conflicts, to research and development, procurement, production, marketing, operations and maintenance, and cash flow risks. The result is a GERT network in which each directed link between business activities carries one or more identified risk factors.

The mathematical core of the model addresses two questions: how multiple risks aggregate at a node, and how they propagate through different network structures. For aggregation, the Choquet fuzzy integral quantifies the combined risk level in a way that allows for interactions among risk factors. A positive interaction parameter indicates synergy, where several risks together produce more damage than expected, while a negative value indicates redundancy, where overlapping risks partially cancel each other’s distinct effects. For propagation, the authors derive equivalent transfer functions for five typical structures in the business process network: serial chains, mutually exclusive parallel paths, hybrid parallel structures in which any subset of upstream paths may execute simultaneously, converging parallel structures in which all upstream paths are mandatory prerequisites, and cyclic structures in which an activity can loop back on itself before proceeding.

Once the network is reduced, the model outputs two quantitative indicators for each business activity and business segment. The instability degree, based on risk volatility derived from the variance of operational cost, measures internal fluctuation during risk transmission. The uncertainty degree, based on risk entropy drawn from information theory, measures the randomness a node faces from external risks. Applied to an illustrative scenario in which a public health emergency and geopolitical tensions occur simultaneously, with operational-cost records from 2021 to 2023 collected during field investigations, the model produced a striking pattern. The instability degree of business activities rises steadily from the front to the tail of the network, meaning downstream activities inherit and amplify the accumulated effects of upstream disruptions. The uncertainty degree, by contrast, follows an inverted U-shaped curve within each segment: it starts low when risks are clear and controllable, peaks in the middle stages where potential risks emerge and information becomes ambiguous, and then declines as enterprises adapt and risk boundaries become identifiable.

The classification results pinpoint where trouble concentrates. The five activities with both high instability and high uncertainty, numbered 40 through 44, are spare parts transportation, on-site maintenance, equipment overhaul, software upgrade, and technical support, all located in the execution stage of the operations and maintenance segment at the downstream end of the network. Activities 45 through 47, at the very end of the process, show high instability but low uncertainty, a pattern the authors call the vulnerability paradox: even when uncertainty falls, residual instability remains high, so a risk event striking at the final stage could cause disproportionate damage. Low-instability, high-uncertainty activities cluster in the early stages of procurement, production planning, marketing decisions, and maintenance planning, where outcomes are uncertain but risks have not yet converted into cost fluctuations. The marketing segment overall showed both low instability and low uncertainty, an anomaly the authors attribute to its weaker dependence on the supply and production chains and to the short-term rigidity of market demand.

The authors also tested the model’s robustness. Perturbing the AI-derived risk transmission probabilities by 5, 10, and 15 percent changed the instability degree by 1.80, 4.49, and 4.74 percent on average, and the uncertainty degree by 4.56, 10.08, and 14.19 percent, yet the risk classifications of the 41 measured business activities remained over 93 percent consistent with the baseline even at the highest perturbation. Perturbations to the KPE-BERTopic risk-level estimates had smaller effects, leaving uncertainty unchanged and altering classifications in fewer than 1 percent of cases at the 15 percent level. The authors note limitations: due to limited enterprise data availability, operational cost serves as the primary activity parameter, and future work could incorporate additional parameters to characterize multi-source risk more comprehensively. They propose that enterprises use the model’s output to adopt layered and node-specific prevention strategies, allocating flexible resources to high-instability tail activities and strengthening environmental monitoring where uncertainty is high.

Subject of Research: A GERT-Choquet fuzzy model that quantifies how multiple coupled risks transmit through the business process networks of high-end equipment manufacturing enterprises.

Article Title: Multi-source risk coupling transmission in high-end equipment manufacturing enterprises under unstable and uncertain environments

Article References: Wei, C., Liu, P.-F., & Rao, W.-Y. (2026). Multi-source risk coupling transmission in high-end equipment manufacturing enterprises under unstable and uncertain environments. Frontiers in Physics, 14, Article 1873810. https://doi.org/10.3389/fphy.2026.1873810

Image Credits: AI Generated

DOI: 10.3389/fphy.2026.1873810

Keywords: risk transmission, high-end equipment manufacturing, GERT network, Choquet fuzzy integral, large language models, supply chain risk, risk entropy, operational instability, geopolitical conflict, public health emergency, KPE-BERTopic, network modeling