Digital and Intelligent Optimization Mechanism of the Entire Supply Chain Link Based on Generative Artificial Intelligence and Neural Semantic Analysis

Abstract

In the process of supply chain optimization, data has characteristics such as multi-source heterogeneity and unstructuredness. Traditional supply chain optimization methods that rely on structured data and statistical analysis cannot achieve ideal goals. Therefore, this article studies the digital and intelligent optimization mechanism of the entire supply chain link based on generative artificial intelligence. The entire supply chain link is divided into product design and process links, product raw material procurement links, product production and manufacturing management links, product delivery links, and product retirement and recycling links. For each link, the ChatGPT large language model in generative artificial intelligence adopts a neural semantic analysis method based on an encoding–decoding architecture. For the multi-source heterogeneous general knowledge within the entire supply chain link, general corpus training, expert annotation, and special corpus training are carried out, and the semantic analysis of the general knowledge of the entire supply chain link is realized through in-depth mining and understanding. Based on the semantic analysis results, the generative adversarial network in generative artificial intelligence is used to predict complex patterns or solutions such as product design, transportation routes, and sales methods in each link of the entire supply chain, making the prediction results more accurate and more in line with the actual supply chain business. The experimental results show that this mechanism can accurately analyze the semantics of the general knowledge of the entire supply chain link, improve the accuracy of the prediction of each function of the entire supply chain link, and significantly improve the economic benefits of enterprises.

Keywords

adversarial network artificial intelligence digital and intelligent optimization entire supply chain link large language model semantic analysis

Introduction

As the critical link connecting production, distribution, and consumption, the operational efficiency and coordination capabilities of the supply chain directly determine a company’s competitiveness in complex and ever-changing market environments.¹ With intensifying market competition, rapidly evolving consumer demands, and frequent disruptions caused by external uncertainties (such as natural disasters, public health crises, and international trade frictions), traditional supply chain management models are facing unprecedented challenges.² Enterprises urgently need to leverage advanced technologies to achieve digital transformation and intelligent upgrading of their supply chains,³ thereby enhancing supply chain resilience, flexibility, and response speed to gain a competitive edge in the market. Against this backdrop, Roy et al. proposed a two-stage stochastic optimization method for supply chain optimization with simultaneous elasticity strategies,⁴ clearly defining elasticity objectives for retail supply chains under uncertainty, setting constraints, and aiming to minimize costs while considering elasticity indicators. The two-stage stochastic optimization approach involves decision-making in the first stage before uncertainty materializes and adjustments in the second stage afterward, with heuristic algorithms used for solution. However, this method faces challenges due to the high dimensionality of decision variables and the complexity of the solution process. Billal et al. proposed a Geographic Information System (GIS)-based Mixed-Integer Nonlinear Programming (MINLP) framework⁵ aimed at minimizing total cost. They utilized GIS technology for spatial analysis, identified candidate locations for Waste-to-Energy (WtE) facilities based on multi-criteria evaluation, constructed a MINLP model, and solved it using GIS data to obtain optimal decision variables. However, this framework may not fully consider supply chain uncertainty, potentially leading to poor practical application results. Jahin et al. proposed a supply chain big data prediction optimization method based on data preprocessing and machine learning,⁶ which collects data from all links, processes it into a unified format to extract feature vectors, constructs a prediction model using neural networks, trains the model, and applies it to make decisions. However, enterprises may find it difficult to understand the decision-making process and basis of the model, affecting trust and application enthusiasm. Volikatla proposed a supply chain optimization method based on reinforcement learning,⁷ with the objectives of minimizing costs and maximizing service levels. The problem is abstracted into a Markov decision process, and a deep Q-network is used to construct a model. The model is trained in a simulated environment, evaluated, and then deployed. However, deep reinforcement learning models have complex structures and numerous parameters, making training and debugging difficult and prone to overfitting and underfitting issues.

The rise of generative artificial intelligence (AI) has brought new ideas and methods to the digital and intelligent optimization of the entire supply chain.⁸ Compared with existing supply chain optimization mechanisms, generative AI, with its powerful data generation, pattern recognition, and decision-making reasoning capabilities, can deeply explore potential patterns and related information in supply chain data, simulate supply chain operation processes in various complex scenarios, and generate forward-looking and guiding optimization strategies. This study aims to explore in depth the digital optimization mechanism of the entire supply chain based on generative AI. By constructing a generative AI model suitable for supply chain scenarios and combining advanced algorithms and technological means, intelligent optimization and collaborative operation of each link in the supply chain can be achieved.

Digital and Intelligent Optimization Mechanism of the Entire Supply Chain Link

Construction of the entire supply chain link system

The digital and intelligent optimization management of the entire supply chain link, combined with various technologies in generative AI, realizes end-to-end optimization through intelligent technologies from multiple links such as planning, procurement, inventory, logistics, settlement and payment, risk control, and compliance.⁹ Figure 1 shows the entire supply chain link system.

FIG. 1.

Construction of the entire supply chain system.

Taking into account the above considerations, the supply chain is divided into six key stages based on the aforementioned supply chain lifecycle, including design and process, procurement, production and manufacturing management, delivery, sales and after-sales service, and retirement and recycling. (1)

Product design and process stage: This stage is the starting point of the supply chain lifecycle, with the primary task of designing products and planning processes based on market demand and technological developments.¹⁰ This stage requires innovation to optimize product performance and cost.

(2)

Product raw material procurement stage: By analyzing historical procurement data, market price trends, supplier information, and other data, the primary task is to predict future procurement prices and provide optimal procurement strategies; based on market supply and demand relationships and raw material price fluctuations, predict the price trend of a certain raw material over a period of time.

(3)

The product production and manufacturing management stage: This is the execution phase of product manufacturing, where production factors such as equipment, materials, personnel, and processes are managed and controlled¹¹ to ensure an efficient and orderly production process. It primarily faces challenges such as multiple involved links, high real-time requirements, and frequent abnormal situations, necessitating detailed management of personnel, machinery, materials, methods, and environment.

(4)

Product delivery phase: By analyzing traffic data, map information, cargo weight, and volume, the optimal transportation routes for vehicles are planned to reduce transportation time and costs; delivery times are predicted to provide customers with more accurate delivery information, thereby enhancing customer satisfaction.

(5)

Product sales and after-sales service phase: This phase realizes the value of the product after completion. Addressing issues such as personalized customer needs, product differentiation, and high requirements for timely after-sales response, this phase involves product promotion, sales, and customer service to provide customers with high-quality products and services.

(6)

Product retirement and recycling phase: After reaching a certain stage of use, products enter the retirement phase. During this phase, obsolete products are recycled, disassembled, and reused to maximize resource conservation and environmental protection. This phase primarily faces challenges such as diverse product types, varying conditions, and dispersed recycling channels, necessitating the establishment of reverse logistics systems and recycling networks.

To ensure real-time sharing and collaborative work of information among personnel, machines, materials, methods, and environments in production and manufacturing management, an integrated information management platform is built using generative AI technology. The platform can collect real-time data from each link, analyze and process it, and use intelligent algorithms to achieve rapid information transmission and accurate matching. At the same time, a collaborative work mechanism and warning system are set up to timely coordinate and solve abnormal problems, thereby improving the efficiency and orderliness of the production process.

In supply chain management, redefining process classification is to more accurately reflect each key link and better apply generative AI technology for digital optimization. The original broad categories such as “planning, procurement, inventory” have been refined into six stages: “design and process, procurement, production and manufacturing management, delivery, sales and after-sales, retirement and recycling” to enhance targeting, clarify specific needs and challenges; This helps to train and optimize generative AI technology for specific data and requirements, improving application effectiveness; Meanwhile, refining the process can more clearly define the interfaces and collaboration mechanisms between links, improve collaboration efficiency, and better cope with the complexity and dynamism of the supply chain, and it helps to achieve end-to-end optimization from the starting point to the end point of the supply chain, build a more complete optimization model, achieve global optimality, and promote the value enhancement of the entire supply chain.

In modern enterprise operations, supply chain management has crucial strategic value and plays a decisive role in shaping the core competitiveness of enterprises: Supply chain management that runs through production, circulation, and consumption links and through fine coordination and optimization obtains cost-effective raw materials through precise market analysis and supplier management in the procurement link, optimizes processes, plans in the production link, adopts Just-In-Time (JIT) mode to improve efficiency and reduce unit costs, optimizes transportation and distribution strategies in the logistics link to reduce costs and improve efficiency, and achieves cost control and efficiency improvement; strictly screening and controlling raw material procurement, coordinating and monitoring the production process, and ensuring timely, accurate, and effective delivery and after-sales service to ensure product quality and reliability; by sharing timely demand feedback with market research, sales channels, and other information, adjusting product strategies and production plans, and collaborating with all parties in the supply chain to obtain innovative resources, we aim to enhance market demand response and innovation capabilities; establish risk warning and emergency plans, collaborate with multiple suppliers, build inventory buffers, and pay attention to policy changes to diversify risks, strengthen information sharing and collaborative risk response, and enhance risk management and response capabilities; by providing high-quality product services to meet customer needs and expectations, improving satisfaction and loyalty, leveraging customer word-of-mouth communication to enhance brand value, and forming a virtuous cycle. It can be seen that supply chain management has a profound impact on the core competitiveness of enterprises through the abovementioned aspects and is an important guarantee for enterprises to achieve sustainable development.

Design of the digital intelligence optimization mechanism architecture

For each link in the whole supply chain system shown in Figure 1, design the digital intelligence optimization mechanism architecture for the whole supply chain based on generative AI, as shown in Table 1.

Table 1.

Architecture of digitalization optimization mechanism for the whole supply chain based on generative artificial intelligence

The entire supply chain link	General corpus training	Expert annotation	Specialized corpus training	Generative artificial intelligence input information	Generative artificial intelligence outputs information
Product design and manufacturing process	Market demand, consumer preferences, design trends, material properties, etc.	Evaluation, innovation, cost-effectiveness ratio, etc. of expert feasibility plans	Specific product design parameters, design cases, patent data, etc.	Design requirements, functional specifications, cost models, user preferences, etc.	Design innovation suggestions, process optimization plans, material selection suggestions, material optimization suggestions, etc.
Product raw material procurement process	Production management theory, process optimization theory, scheduling theory, etc.	Expert evaluation of production and equipment scheduling plans	Factory capabilities, real-time orders, resource data, scheduling strategies, etc.	Design documents, order requirements, equipment condition, process requirements, etc.	Suggestions for optimizing production plans, equipment configuration schemes, scheduling strategies, raw material procurement plans, remaining production cycles, etc.
Product production and manufacturing management process	Quality system, control standards, energy-saving technology, environmental regulations, etc.	Expert evaluation of production plans, safety contingency plans, etc.	Maintenance manual, control parameters, real-time machine feedback data, and emergency requirements, etc.	Production tasks, equipment status, process requirements, compliance requirements, etc.	Quality control suggestions, monitoring model construction plans, equipment maintenance and fault warning, parameter adjustment suggestions, etc.
Product delivery process	Inventory management methods, supply chain theory, logistics transportation regulations, etc.	Expert evaluation of inventory plans, logistics plans, supply chain planning plans, etc.	Supplier data, warehouse layout, supply chain strategy, cost composition cases, etc.	Production plan, inventory level, supply status, batch status, transportation status, etc.	Suggestions for supply chain optimization, inventory strategy, logistics plan, cost optimization, risk assessment, material shortage response strategy, etc.
Product sales and after-sales process	Market management theory, sales strategy, customer service cases, after-sales process, etc.	Expert evaluation of sales strategies, customer service plans, etc.	Consumer trends, sales data, user feedback, customer complaint cases, etc.	Market demand, sales channels, customer evaluations, after-sales feedback, etc.	Product sales forecast, marketing promotion strategy, customer service optimization plan, user demand mining, suggestions for optimizing after-sales costs, etc.
Product retirement and recycling process	Environmental regulations, principles of circular economy, waste disposal technologies, etc.	Expert evaluation of recycling process design, material reuse strategies, etc.	Recycling network, recycling technology, recycling cases, policy requirements, etc.	Product information, recycling plan, retirement capability, environmental standards, etc.	Suggestions for product recycling, optimization plan for dismantling process, waste recycling strategy, environmental impact analysis, etc.

The data sources in Table 1 are extensive, and market-related data (market demand, consumer preferences, consumption trends) are obtained from market research company reports, social media data analysis, and e-commerce platform sales data mining; design and manufacturing data (design parameters, cases, patents) are sourced from the internal design department and patent database of the enterprise; production management data (production theory, scheduling strategies, equipment status) come from the enterprise’s production management system and equipment sensors; quality and compliance data (quality system, control standards, environmental regulations) are obtained from industry associations and government regulatory websites; logistics and supply chain data (inventory management methods, logistics transportation regulations, supplier data) come from information sharing between logistics enterprise systems and supply chain partners; sales and after-sales data (sales strategy, user feedback, complaint cases) are collected through the enterprise sales system and customer service records; recycling-related data (environmental regulations, recycling technologies, case studies) are obtained from environmental organizations and recycling companies. In terms of difficulty, obtaining internal data within the enterprise is relatively easy, but it involves coordination among multiple departments and issues with data permissions; market research and industry report data require purchasing or collaborating with professional institutions, which incurs costs; the government, associations, and other public channels provide convenient access to data, but may not update it in a timely manner; the difficulty of obtaining shared data among partners depends on the depth of cooperation and level of trust. During preprocessing, text data are cleaned to remove noise, duplicates, and errors, followed by word segmentation and part-of-speech tagging. Numerical data are standardized and normalized to ensure dimensional consistency. Missing values are imputed using methods such as mean imputation and regression prediction. Outliers are identified and corrected to guarantee data quality. When data are incomplete or of poor quality, data augmentation techniques are first applied, such as synonym replacement in text data, sentence reorganization to generate additional samples, and adding reasonable noise to numerical data to simulate diverse scenarios. Transfer learning methods are then adopted, leveraging high-quality data from related fields or historical periods to train models and transfer knowledge to the current task. Additionally, based on expert experience, model outputs are reviewed and corrected by experts, and expert knowledge is used to compensate for data deficiencies and biases, ensuring system reliability.

The digital intelligence optimization process for the whole supply chain based on generative AI includes: (1)

General corpus training: Build a semantic foundation based on the large language model of generative AI technology. Collect general knowledge of the whole supply chain (such as market demand theory, production management basics) as the corpus¹² and input it into the large language model. Through the Transformer architecture,¹³ the model tokenizes the general knowledge text of the whole supply chain and calculates the attention mechanism, learning cross-link semantic associations such as “market demand-product design” and “production theory-scheduling logic,” and constructing a basic semantic understanding framework for supply chain operations, such as understanding the semantic mapping of how “consumer preferences” affect “product function specifications.”

By constructing a multidimensional semantic vector space, these semantic associations can be quantified and expressed in detail. For example, for “market demand product design,” market demand can be subdivided into sub-dimensions such as functional requirements, price sensitivity, and appearance preferences. Product design can be decomposed into sub-dimensions such as functional characteristics, cost structure, and appearance style. The model can be used to calculate the correlation coefficients between each sub-dimension, forming a semantic association matrix. At the same time, embedding vector technology is introduced to map each element of market demand and product design into high-dimensional vectors. The semantic correlation strength is quantified through numerical indicators such as cosine similarity and Euclidean distance between vectors. Finally, the quantified results of cross-link semantic correlation are presented intuitively in the form of heat maps or correlation network diagrams. (2)

Expert annotation: Strengthen the injection of professional semantics and rules. Experts annotate professional scenarios of the whole supply chain (such as product design feasibility assessment, production scheduling strategy), and transform industry knowledge (such as “cost performance ratio calculation rules”, “equipment maintenance priority logic”) into structured semantics. The large language model based on generative AI learns these annotations, corrects the semantic deviation of the general corpus, and strengthens the understanding of professional semantics. For example, distinguish the semantic differences of “production planning theory” in discrete manufacturing and process manufacturing, so that the model masters the semantic expression of expert-level business logic.

To ensure the accuracy and comprehensiveness of expert labeling, it is necessary to select experts with profound industry experience and professional knowledge, and develop detailed labeling specifications and standards; Establish a cross-review mechanism where multiple experts annotate the same content and discuss differences in depth to reach a consensus; Collect actual business cases and feedback to update the annotated content. At the same time, clarify the scope of expert knowledge coverage, use unified standards to ensure consistency, organize discussions to resolve conflicts based on case studies and business logic, establish a regular communication and dynamic update system, adjust annotations according to industry development and business feedback, and ensure timeliness and comprehensiveness. (3)

Special corpus training: Deepen the semantic adaptation of the whole supply chain scenario. Input special corpus for the sub-links of the whole supply chain. The large language model based on generative AI deeply learns the scenario-specific semantics. For example, in the product design link, master the associated semantics of “mobile phone camera parameters-cost model-user preferences”; in production scheduling, understand the semantic logic of “factory real-time orders—resource data—production scheduling strategy,”¹⁴ and achieve precise semantic adaptation from the general of the whole supply chain to the scenario-specific.

It can be verified by designing multidimensional and comprehensive testing scenarios, such as providing different combinations of mobile phone camera parameter settings, allowing the model to predict costs based on the learned cost model, and combining simulated user preference data to provide market positioning recommendations. The model output can be deeply compared with the analysis results of professional supply chain analysts based on comprehensive knowledge and experience. If the model can accurately and reasonably infer the cost impact and market positioning that conforms to user preference logic under complex and changing parameter combinations, rather than just based on fixed simple pattern associations, it can prove that it truly understands semantic associations. (4)

Generate a professional generative AI model: Integrate semantics and model capabilities. Based on the semantic knowledge learned by the large language model of generative AI, combined with the tasks of all links in the entire supply chain (design, procurement, production, sales, etc.), construct a professional generative AI model, a generative adversarial network model (GAN).¹⁵ For example, when designing the generative adversarial network model, integrate “market demand semantics + product design rule semantics” so that the generative adversarial network model can generate a full-link design plan for the product based on semantic understanding; through fine-tuning, make the model parameters adapt to the professional tasks of the supply chain and have the semantic-driven generation ability.

The paradigm breakthrough brought by generative AI is reflected in multiple dimensions. At the semantic level, it can deeply understand complex semantic associations across links, construct a precise semantic understanding framework, and achieve semantic adaptation from general to specific; In terms of data processing, multi-source heterogeneous data can be integrated to explore potential patterns; In terms of decision generation, intelligent decision-making and dynamic optimization are achieved through adversarial games between generators and discriminators to generate optimization solutions that are close to business reality, rather than relying on fixed rules. (5)

Input information of generative AI: Semantic parsing and feature extraction. When inputting general knowledge information of the entire supply chain (such as “design requirements, functional specifications”), the large language model first performs semantic parsing to identify key semantic elements (such as the user demand semantics corresponding to “consumer preferences” and the financial semantics of “cost model”), extracts features and transforms them into vector representations that can be calculated by the GAN model to prepare for subsequent predictions. For example, transform the semantic relationship of “user preference-material selection” into a feature vector and input it into the GAN.

(6)

GAN prediction: Semantic-driven adversarial generation. The generator of the GAN model generates candidate outputs such as “design innovation suggestions” and “production plan optimization plans” based on the semantic feature vectors of the general knowledge of the entire supply chain processed by the large language model. For example, in the product design link, the generator generates various material selection suggestions based on “market demand semantics + material property semantics” and simulates possible optimization plans. The discriminator uses expert-annotated professional knowledge and real business data (such as historical optimal design plans and actual feasible production plans) as the discrimination basis to judge whether the plans output by the generator are “reasonable and feasible.” For example, when discriminating the production plan, compare the historical scheduling data and evaluate whether “production scheduling strategy-remaining production cycle” conforms to the actual business logic. The generator and the discriminator continuously play games. The generator continuously optimizes the output to make the plan closer to the real business needs; the discriminator continuously strengthens the discrimination ability to eliminate unreasonable plans. Through adversarial training, finally output high-precision prediction results (such as the optimal process optimization plan and accurate supply chain optimization suggestions) to achieve prediction and optimization of all links in the entire supply chain.

In the process of data-driven optimization of the entire supply chain based on generative AI, there are differences in the application logic between large language models and generative adversarial networks at different stages of the supply chain, but there is no fundamental divergence. The former constructs a semantic understanding framework through general corpus training, expert annotation, and specialized corpus training, providing the basis for semantic parsing and feature extraction for the latter; The latter is based on the semantic feature vectors processed by the former, and through the adversarial game between the generator and discriminator, generates and optimizes a prediction scheme that is close to the actual business situation. The two work together to achieve intelligent optimization of the entire supply chain.

Semantic analysis of the entire supply chain link based on generative AI

In the digital and intelligent optimization mechanism of the entire supply chain based on generative AI, the large language model in generative AI technology can analyze the semantics of general knowledge of the entire supply chain within the entire supply chain link through semantic analysis and transform it into decision-making logic that can be understood by the GAN model.

ChatGPT is a large language model developed by OpenAI based on the Transformer architecture.¹⁶ Through training with massive amounts of data, it has acquired powerful language generation and comprehension capabilities. It can generate natural and fluent language responses based on user input such as text, images, and voice, providing users with a high-quality interactive experience. More importantly, ChatGPT employs neural semantic analysis methods to analyze the semantic meaning of input information related to the entire supply chain, understand users’ intentions and needs, and provide more precise services. The neural semantic analysis method adopts an encoding–decoding architecture,¹⁷ taking into account the diversity of language, and designs a dual encoding–decoding semantic analysis model.

Within the semantic expression encoding–decoding architecture, the general knowledge of the entire supply chain link is used as the input to the encoder. From this, the hidden layer state corresponding to the general knowledge of the entire supply chain link is output and transmitted to the decoder as the input to the decoder. Let $x = (x_{1}, x_{2}, \dots, x_{m})$ represent the general knowledge of the entire supply chain link. Usually, the recurrent neural network reads $x = (x_{1}, x_{2}, \dots, x_{m})$ in the order from front to back, and the encoder reads it sequentially. By running the recurrent neural network in both the forward and backward directions, any vocabulary $x$ in the general knowledge of the entire supply chain link can have hidden layer states in both the forward and backward directions. By concatenating the two states, the output result of the encoder can be obtained, which contains the information of the two vocabulary $x$ before and after in the information of the entire supply chain link.

Based on the encoding–decoding model, another encoder is introduced to ensure that each general knowledge of the entire supply chain link input at the input layer corresponds to a separate encoder, while sharing a decoder. Let $x$ and $x^{'}$ be two different general knowledge of the entire supply chain link, and let $y$ and $z_{i}$ represent the input of the general knowledge of the target-end supply chain link and the hidden layer state of the decoder at the current moment, respectively. By combining the attention scores corresponding to the two encoders, the input $c_{i}$ for calculating the hidden layer state of the decoder at the current moment can be obtained.

The initial hidden layer state of the decoding end can be described using equations (1) and (2):

ε x = [\begin{array}{l} {\vec{ε x_{1}}}_{} {\vec{ε x_{2}}}_{} \dots \vec{ε x_{m}} \\ {\vec{ε x_{1}}}_{} {\vec{ε x_{2}}}_{} \dots \vec{ε x_{m}} \end{array}]

(1)

ε x^{'} = [\begin{array}{l} {\vec{ε {x^{'}}_{1}}}_{} {\vec{ε {x^{'}}_{2}}}_{} \dots \vec{ε {x^{'}}_{n}} \\ {\vec{ε {x^{'}}_{1}}}_{} {\vec{ε {x^{'}}_{2}}}_{} \dots \vec{ε {x^{'}}_{n}} \end{array}]

(2)

Among them, $ε x$ and $ε x^{'}$ represent the hidden layer states of the encoder corresponding to the general knowledge $x$ and $x^{'}$ of the entire supply chain link, respectively.

After obtaining $b x$ and $ε x^{'}$ , connect the two according to the last dimension of the matrix, and use equation (3) to calculate the mean $ε$ of the first dimension of the connected matrix:

ε = m ([ε x; ε x^{'}])

(3)

Among them, $m ()$ represents the mean function.

Perform a non-linear transformation on $ε$ and a non-linear deformation of $\tanh$ on the result of the linear transformation to obtain the value of the initial hidden layer state of the decoding end:

z_{0} = \tanh (W_{c} \cdot ε)

(4)

In the semantic expression encoding–decoding architecture, let $z_{i - 1}$ and $y_{i - 1}$ represent the hidden layer state at the $i - 1$ -th moment of the top layer of the decoder and the embedding of the previous general knowledge vocabulary of the entire supply chain link, respectively. After combining the two with $c_{i}$ , the hidden layer state at the i-th moment can be obtained. The formula is described as follows:

z_{i} = f (z_{i - 1}, y_{i - 1}, c_{i})

(5)

Mapping the natural language sentence of the general knowledge of the entire supply chain link input by the input layer can obtain the sequence $(ε_{1}, ε_{2}, \dots, ε_{m})$ , which is the main basis for the calculation of $c_{i}$ . Different $ε_{j}$ contain the information of the entire input natural language sentence of the general knowledge of the supply chain link, mainly the information of the surrounding vocabulary of the j-th general knowledge vocabulary of the supply chain link in the sentence. On this basis, weighted summation can be used to obtain $c_{i}$ . The formula is described as follows:

c_{i} = \sum_{j = 1}^{m} w_{i j} ε_{j}

(6)

Among them, $w_{i j}$ represents the weights of different $ε_{j}$ . Its formula is described as follows:

{\begin{cases} w_{i j} = \frac{\exp (e_{i j})}{\sum_{k = 1}^{m} \exp (e_{i k})} \\ e_{i j} = β (z_{i - 1}, ε_{j}) \end{cases}

(7)

Among them, $β$ represents the alignment model, which describes the alignment level between the vocabulary around the j -th vocabulary in the natural language sentence of the general knowledge of the supply chain link in the input layer and the output of the i -th vocabulary through scores. The calculation of this score is mainly based on $z_{i - 1}$ and $y_{i - 1}$ .

From the above process, it can be seen that the hidden layer state of the previous moment and $ε_{j}$ are very important in determining the hidden-layer state of the current moment.

After introducing the attention mechanism into the dual encoding–decoding general knowledge semantic analysis model of the supply chain link based on ChatGPT,¹⁸ the decoder only needs to pay attention to part of the information of the input natural language sentence of the general knowledge of the supply chain link. In order to process the information of the two encoders simultaneously, the attention model needs to be improved. According to the calculation process of attention, in the process of improving the attention model, the corresponding relationships between the two general knowledge text vectors of the supply chain link and the two encoders are determined, respectively, that is, $c x_{i}$ and $c {x^{'}}_{i}$ , and thus $z_{i}$ in the original model is replaced:

z_{i} = f (z_{i - 1}, y_{i - 1}, c x_{i}, c {x^{'}}_{i})

(8)

The above dual encoding–decoding model can be easily extended to handle more language input cases, thereby realizing the neural semantic analysis from multiple languages to semantic expressions in the general knowledge of the entire supply chain link.

In the data-driven optimization of supply chain based on generative AI, evaluating the synergistic effect of semantic analysis of large language models and generative adversarial network adversarial generation can be approached from two aspects: the impact of semantic understanding accuracy on the rationality of generated solutions, and the promotion effect of generated solution feedback on semantic analysis optimization. By comparing the performance of the generated solutions before and after introducing collaborative mechanisms in terms of being close to business reality and responding to complex scenarios, as well as analyzing the depth of semantic analysis’s understanding of cross-link semantic associations in the supply chain after combining generated feedback, the synergistic effect of the two can be comprehensively judged.

Supply chain full-link prediction based on generative adversarial neural networks

The large language model in generative AI technology performs semantic analysis on the general knowledge of the entire supply chain link, providing an “understanding basis” for the GAN. Based on the semantic analysis results of the general knowledge of the entire supply chain link, the GAN makes the prediction schemes such as product design, transportation routes, and sales methods in each link of the entire supply chain link more accurate and more in line with the actual supply chain business through adversarial generation. The large-scale language model in generative AI technology provides the “understanding foundation” for generative adversarial networks, mainly including the semantic parsing ability of general knowledge throughout the supply chain, such as converting textual information such as market demand, production management, logistics parameters, etc. into structured semantic representations; At the same time, through cross link semantic association learning, mapping relationships between business logic such as “market demand product design” and “production theory scheduling logic” are established, forming a semantic understanding framework covering key indicators such as order fulfillment cycle, inventory turnover rate, and logistics cost, providing interpretable business knowledge priors for generative adversarial networks.

The generative adversarial network consists of two parts: the generator network $S$ and the discriminator network $P$ .¹⁹ Figure 2 shows the structure diagram of the generative adversarial network.

FIG. 2.

Generative adversarial neural network structure diagram.

The generator network $S$ generates new general knowledge of the entire supply chain link according to the semantic analysis results of the general knowledge of the entire supply chain link input. These general knowledge of the entire supply chain link have the characteristics of a fixed distribution and update parameters according to the evaluation of the output of the generator network by the discriminator network $P$ . $P$ is essentially a binary classifier. The actual supply chain full-link prediction scheme label and the general knowledge of the entire supply chain link generated by $S$ are, respectively, input into $P$ . $P$ outputs a real number in the range of $[0, 1]$ according to the input general knowledge of the entire supply chain link. This output can describe the possibility that the general knowledge of the entire supply chain link originates from the actual general knowledge of the entire supply chain link. $S$ needs to maximize the probability that the generated general knowledge of the entire supply chain link passes through $P$ , while $P$ is the opposite.

The loss functions of $S$ and $P$ are, respectively:

\max_{P} V_{1} (P) = E_{h \sim P_{g}} [\lg (P (h))]

(9)

\min_{S} V_{1} (S) = E_{a \sim P_{r}} [\lg (1 - P (S (a)))]

(10)

In the above formula, $h$ and $a$ , respectively, represent the prediction schemes of each function of the actual supply chain full link that the generative adversarial network needs to generate and the semantic analysis results of the random general knowledge of the entire supply chain link that satisfy the $P_{r}$ distribution; $E []$ represents the mathematical expectation of the distribution function; $P (h)$ and $P (S (a))$ , respectively, represent the probability that $h$ outputs through $P$ and the probability that the general knowledge of the entire supply chain link generated by using $a$ through $S$ outputs after passing through $P$ .

In the process of optimizing the parameters of the generative adversarial network according to the loss functions of $S$ and $P$ , generally $S$ and $P$ are alternately iterated. However, considering the instability of the training process, a conditional adversarial loss function²⁰ is introduced. Taking the sample $l$ of the general knowledge of the entire supply chain link generated as the constraint condition for $S$ and $P$ , $l$ satisfies the $P_{d}$ distribution, and the conditional adversarial loss function $V (P, S)$ of $S$ and $P$ is described by formula (11):

\min_{S} \max_{P} V (P, S) = E_{h, l \sim P_{g}, P_{r}} ([\lg (S (h, l))] + [\lg (1 - S (P (l), l))])

(11)

In formula (11), $P (h, l)$ and $P (S (l), l)$ , respectively, represent the probability of the output after $h$ and $l$ are jointly input into $P$ , and the probability of the output after the general knowledge common to the entire supply chain generated by $l$ through $S$ and $l$ are jointly input into $P$ .

The specific process of using generative adversarial networks for digital and intelligent optimization of each link in the entire supply chain is described as follows: (1)

Start: Start the GAN.

(2)

Build input set: Prepare the basic supply chain knowledge semantic analysis results for digital and intelligent optimization of each link in the entire supply chain, such as order quantity sequences, warehouse (analogous to base stations) information, transportation time parameters, inventory turnover parameters, and other learning parameters.

(3)

Build a generative adversarial network: Initialize the generative network and the discriminative network separately. The generative network is responsible for generating predictions of key indicators for the entire supply chain (such as order fulfillment cycle, inventory turnover rate, logistics costs, etc.), while the discriminative network achieves effective predictions of the operational status of the entire supply chain (such as order delivery predictions, inventory health predictions, logistics cost predictions, etc.).

(4)

Alternate training of the network: Fix the parameters of the generative network, input the predicted order fulfillment volume and predicted inventory levels generated under the supply chain scenario, along with the actual order fulfillment volume and actual inventory levels, into the discriminative network. Allow the discriminative network to learn to distinguish between real and fake data. After completing the discriminative training, update the training parameters of the discriminative network. Fix the discriminative network parameters, input the input set into the generative network to generate predicted data, and then adjust the generative network parameters based on the discriminative network’s results to make the generated data more aligned with reality, thereby improving prediction accuracy.

(5)

Evaluate prediction effectiveness: Check whether the supply chain end-to-end prediction results generated by the generation network are sufficiently similar to the actual end-to-end logistics timelines and costs. If similar, the process ends; if not, return to continue adjusting the parameters of the generation network and discriminator network training until the similarity requirements are met.

To ensure the timeliness and accuracy of data analysis during product delivery, a real-time data collection system can be built to integrate multiple data sources and automatically update them. At the same time, data cleaning and verification techniques can be used to eliminate erroneous information. In response to the impact of emergencies, it is necessary to establish a dynamic monitoring and warning mechanism, track changes in road conditions, weather, etc. in real time, combine AI algorithms to quickly generate backup route plans, and evaluate the optimal adjustment strategy through simulation to ensure flexible adaptation of transportation routes and delivery times.

Experimental Results and Analysis

This article studies the digital and intelligent optimization mechanism of the entire supply chain based on generative AI. To verify the effectiveness of the mechanism in this article, a certain new energy vehicle retired power battery enterprise supply chain is selected as the research object, and the mechanism in this article is used to conduct a full-chain digital and intelligent optimization test on its supply chain. Table 2 shows the description of the entire supply chain of the research object.

Table 2.

Supply chain description of research objects

Link	Specific information
Product Design and Process	Focusing on the comprehensive utilization of retired power batteries in the entire industry chain of new energy vehicles, we have the process technology to connect the entire industry chain of retired lithium battery recycling, procurement, warehousing and transportation, hierarchical utilization, recycling and dismantling, chemical smelting, etc.
Product raw material procurement	The procurement target is retired power batteries for new energy vehicles, and we have established cooperative relationships with numerous car companies, battery leasing companies, etc. Currently, our annual procurement volume reaches tens of thousands of tons.
Product production and manufacturing	There are three main industrial bases with an annual comprehensive processing capacity of tens of thousands of tons, leading the industry in processing capacity year after year.
Product Delivery	Deliver processed recycled materials, cascaded batteries, and other products to relevant customers such as battery manufacturers and energy storage companies, with a timely delivery rate of over 95%.
Product sales and after-sales service	Our sales network covers multiple regions both domestically and internationally, and we have established long-term cooperative relationships with numerous well-known enterprises. After-sales provide technical support and quality tracking services, with customer satisfaction reaching over 90%.
Product retirement and recycling	We specialize in the recycling and disposal of retired products, and through efficient recycling processes and advanced technology, we achieve efficient recovery and reuse of metal resources in batteries, with a resource recovery rate of over 90%.

Semantic analysis test

Semantic Analysis

Semantic analysis is the basis for the mechanism in this article to achieve digital and intelligent optimization of the entire supply chain of the research object. Therefore, it is necessary to test the semantic analysis performance of the mechanism in this article. The user inputs the general knowledge of the entire supply chain, and the mechanism in this article is used to conduct semantic analysis on the input general knowledge of the entire supply chain. Table 3 shows the semantic analysis results of the mechanism in this article.

Table 3.

Semantic analysis results of general knowledge throughout the entire supply chain

General knowledge of the entire supply chain input	Semantic analysis results
New product appearance design	Optimization research on the appearance design scheme of new products
500 sets, procurement of electronic components	Execute purchase order processing for 500 sets of electronic components of a certain model for production needs
Randomly inspect the key dimensions of semi-finished products on the production line every 2 hours	Conduct random inspections and monitoring of key dimensions of semi-finished products on the production line at 2-hour intervals
Logistics vehicles, 8:00 am, 200 finished products, warehouse of A city dealer	At 8 am tomorrow, logistics vehicles will be dispatched to deliver 200 finished products to the A city dealer warehouse for delivery processing
Customer feedback, quality defects, initiate after-sales service	Due to customer feedback on quality defects in a certain batch of products, we have initiated an after-sales recall for that batch of products
Last month, 100 units, model, reuse evaluation	Collect information on the 100 old model devices sold last month and conduct feasibility assessments for recycling and reuse

As can be seen from Table 3, the mechanism proposed in this article can effectively analyze the general knowledge in different links of the supply chain of the research object, and the semantic analysis results are roughly the same as the general knowledge of the entire supply chain link input. This shows that the mechanism proposed in this article can effectively analyze the semantics of the input general knowledge of the supply chain.

Quantitative Results of Semantic Analysis Accuracy

To quantify the accuracy of the semantic analysis of the general knowledge of the entire supply chain link by the mechanism proposed in this article, the semantic matching degree and the context-related interaction matching degree in the semantic analysis process of the mechanism proposed in this article are analyzed, and the results are shown in Figure 3.

FIG. 3.

Analysis results of semantic matching degree and context related interaction matching degree.

As can be seen from the analysis of Figure 3, when the mechanism proposed in this article is used for semantic analysis of the general knowledge of the entire supply chain link, with the increase in the number of large prediction model encoders, the semantic matching degree gradually decreases to about 97%, while the context coherence interaction matching degree decreases to about 92%. This is because, as the number of encoders in large prediction models increases, the complexity of the model increases, which may introduce more noise and interference information. At the same time, the difficulty of interaction and integration between different encoders increases, making it difficult to coordinate accurately to maintain the optimal matching state, resulting in a downward trend in semantic matching and context-related interaction matching. Although the values of both indicators show a gradually decreasing trend, they are always higher than 90%, which indicates that when the mechanism proposed in this article is used for the semantic analysis of the general knowledge of the supply chain link, it has a high semantic analysis accuracy.

Prediction results of generative adversarial model

In the mechanism proposed in this article, a generative adversarial network is used to predict various functions in different links of the entire supply chain link. Figure 4 shows the statistical results of the mean square error of the prediction results of the mechanism proposed in this article in different links of the entire supply chain link.

FIG. 4.

Mean square error of prediction results for different links in the supply chain.

As can be seen from the analysis of Figure 4, the overall mean square error of the mechanism proposed in this article is at a relatively low level, and the mean square error of most links is about 0.01, indicating that the generative adversarial model in the mechanism proposed in this article can relatively accurately predict various functions in different links of the entire supply chain link. This shows the effectiveness and reliability of the model in the supply chain prediction task. It can also be seen from the figure that the fluctuation of the mean square error between different links is small, indicating that the prediction ability of the model in each link is relatively balanced, with high stability, and there is no situation where the prediction effect of a certain link is extremely poor.

Benefit analysis

To further verify the impact of the mechanism proposed in this article on enterprises, taking the actual economic benefits as an indicator, and taking the two-stage stochastic optimization mechanism with simultaneous flexible strategies in,⁴ the optimization mechanism based on the GIS-based MINLP framework in,⁵ and the optimization mechanism proposed in⁶ based on data preprocessing and machine learning technology as comparison mechanisms, the Crystal Ball software is used to simulate and analyze the changes in the economic benefits of enterprises after adopting the mechanism proposed in this article and the comparison mechanisms, and the results are shown in Figure 5.

FIG. 5.

Changes in economic benefits.

Analysis of Figure 5 shows that after adopting the mechanism proposed in this article and the comparison mechanisms in the entire supply chain link, the economic benefits of enterprises have been significantly improved compared with those before optimization. This indicates that all kinds of optimization mechanisms adopted have played a positive role in the growth of enterprises’ economic benefits. From the data of 3 years, it can be seen that in the first year, although the economic benefits under the mechanism proposed in this article are better than the three comparative mechanisms, the advantages are not prominent. However, this does not necessarily mean that the mechanism proposed in this article has limitations in short-term applications. It may only be that the differentiated performance of short-term benefits has not been fully highlighted. In the second and third years, the economic benefits under the mechanism of this article are significantly higher than those of other comparison mechanisms. This shows that the mechanism proposed in this article can bring more considerable economic benefits to enterprises in the long-term application, has good sustainability and stability, and has more advantages than other optimization mechanisms.

Conclusion

This article focuses on in-depth research on the digital optimization mechanism of the entire supply chain based on generative AI. This article focuses on in-depth research into the digital optimization mechanism of the entire supply chain based on generative AI. It constructs a supply chain full-chain system covering six key links (design, process, procurement, etc.), designs an optimization mechanism architecture based on generative AI, and elaborates on the data situation and response strategies for each link. It analyzes the optimization process, constructs a semantic understanding framework for the large language model through general corpus training, integrates semantics and model capabilities to generate a GAN model, and leverages collaboration between the large language model and GAN to promote optimization at different stages. By introducing the ChatGPT dual encoding-decoding model and attention mechanism at the semantic analysis level, experiments show that the mechanism achieves high accuracy in semantic analysis. The prediction process utilizes GAN, yielding low mean square error and small fluctuations, indicating accurate and stable predictions. Using actual economic benefits as indicators for simulation analysis, this mechanism demonstrates significant advantages in the second and third years, with sustainability and stability. However, the research has limitations, such as difficulty in obtaining data, cost constraints on data comprehensiveness and timeliness, and potential challenges that models may face in complex scenarios. In the future, data sources can be expanded, and model algorithms optimized to provide more universal and robust solutions.

Data Sharing Agreement

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Authors’ Contributions

Conceptualization, M.Z.; methodology, M.Z. and C.D.; software, C.D.; validation, Q.X.; formal analysis, Q.X.; investigation, H.L.; resources, N.W.; data curation, H.L.; writing—original draft preparation, M.Z.; writing—review and editing, C.D., Q.X., H.L., and N.W.; visualization, Q.X.; supervision, N.W.; project administration, M.Z. All authors have read and agreed to the published version of the manuscript.

Footnotes

Author Disclosure Statement

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding Information

The authors received no financial support for the research, authorship, and/or publication of this article.

Abbreviations

References

Kazancoglu

, Mangla

, Berberoglu

, et al. Towards industry 5.0 challenges for the textile and apparel supply chain for the smart, sustainable, and collaborative industry in emerging economies. Inf Syst Front, 2024; 26(5):1857–1872.

Téllez-Ballesteros

, Torres-Mendoza

, Marmolejo-Saucedo

, et al. Sales predictive analysis for improving supply chain drug sample. Procedia Comput Sci, 2025; 253(000):1712–1719.

Yang

, Thoo

, Ab Talib

, et al. How reverse logistics and sustainable supply chain initiatives influence sustainability performance: The moderating role of organisational learning capability. JMTM, 2024; 35(1):141–163.

Roy

, Almehdawe

, Kabir

. A two-stage stochastic optimization framework for retail supply chain modeling with contemporaneous resilient strategies. Prod Eng Res Devel, 2024; 18(6):903–924.

Billal

, Sebastian

, Kumar

. Developing a GIS-based MINLP framework for optimizing the waste-to-energy supply chain with plant scale considerations. J Clean Prod, 2025; 490:144709.1–144709.19.

Jahin

, Shovon

MSH

, Shin

, et al. Big data—Supply chain management framework for forecasting: Data preprocessing and machine learning techniques. Arch Computat Methods Eng, 2024; 31(6):3619–3645.

Volikatla

. Reinforcement learning approaches for exploring discrepancy variations: Implications for supply chain optimization. J Electr Syst, 2024; 20(5):2861–2868.

Gelfand

, Orel

, Cui

, et al. Biochemical and computational characterization of haloalkane dehalogenase variants designed by generative AI: Accelerating the sn2 step. J Am Chem Soc, 2025; 147(3):2747–2755.

Bogenreuther

, Kastner

, Schneider

, et al. Biodiversity impact of food waste: Quantification for supply chain stages and products in Germany. J Ind Ecol, 2024; 28(2):355–367.

10.

Ekram

, Elmesmary

, Sakr

. Developing a framework to achieve resilience in the oil and gas supply chain during logistics disruptions: An empirical study. IJESM, 2024; 18(4):896–917.

11.

Rahmani

, Pashapour

. Dynamic pricing decision for new and returned products in a dual-channel supply chain based on customer segmentation. Soft Comput, 2024; 28(23–24):13205–13224.

12.

Ceptureanu

, Ferraro

, Ceptureanu

, et al. Impact of artificial intelligence and blockchain on supply chain resilience under influence of change management. Soft Comput, 2025; 29(7):3617–3625.

13.

Arooj

. Fedwindt: Federated learning assisted transformer architecture for collaborative and secure wind power forecasting in diverse conditions. Energy, 2024; 309:133072.

14.

Khan

, Kumar

, Liu

, et al. Modeling enablers of agile and sustainable sourcing networks in a supply chain: A case of the plastic industry. J Clean Prod, 2024; 435:140522.1–140522.17.

15.

Abady

, Barni

, Garzelli

, et al. Generation of synthetic generative adversarial network-based multispectral satellite images with improved sharpness. J Appl Rem Sens, 2024; 18(01):28.

16.

Suvarnakar

, Singh

, Reddy

, et al. S2205 lost in translation: Chatgpt’s digestive dialogue in English vs Spanish. Am J Gastroenterol, 2024; 119(10S):S1576–S1577.

17.

Levin

. Semantic analysis of compound terms in metrology. Part2: Risk in measurement and calculations. Meas Tech, 2024; 67(2):97–108.

18.

Arco

, Andrés

, Gallego-Molina Nicolás

, et al. Enhancing multimodal patterns in neuroimaging by Siamese neural networks with self-attention mechanism. Int J Neural Syst, 2023; 33(4):2350019–2350036.

19.

Zhang

, Li

, Xie

. Robust optimization method for adversarial training model integrating transfer learning. Computer Simulation, 2024; 41(05):383–389.

20.

Goceri

. Gan based augmentation using a hybrid loss function for dermoscopy images. Artif Intell Rev, 2024; 57(9):1–19.