A data-driven manufacturing system sustainability evaluation and improvement method

Abstract

The application discloses a data-driven manufacturing system sustainability evaluation and improvement method, which belongs to the technical field of manufacturing system sustainability evaluation and improvement, and specifically comprises the following steps: S1, obtaining process yield and material consumption data, and constructing an environmental load rate based on the material consumption data and a preset energy value conversion rate database; S2, obtaining photoresist sensitivity and production cycle coincidence, and establishing a nonlinear coupling relationship model for representing the coupling relationship between the process yield and the environmental load rate; and establishing a sustainability index calculated from the energy value output rate and the environmental load rate; S3, constructing a digital twin model with controllable process parameters as input and outputting predicted process yield and predicted environmental load rate; and then solving the optimal process parameters for maximizing the sustainability index through a reinforcement learning decision engine, and generating a process recipe to realize closed-loop control. The application solves the problem that the traditional method cannot comprehensively evaluate multidimensional performance.

Description

Technical Field

[0001] This invention relates to the field of manufacturing system sustainability evaluation and improvement, specifically a data-driven method for manufacturing system sustainability evaluation and improvement. Background Technology

[0002] With the rapid development of data science and intelligent manufacturing, the evaluation and improvement of sustainability in the manufacturing industry has become a crucial issue. In existing technologies, there is a difficult-to-optimize contradiction between the economic benefits of manufacturing systems, such as high yield rates, and their environmental costs, such as high material and energy consumption. Furthermore, the lack of effective closed-loop adaptive control means that manufacturing systems cannot autonomously adjust to maintain optimal performance when faced with external disturbances. Traditionally, the evaluation of manufacturing systems often focuses on a single dimension, such as only economic indicators, while ignoring the actual environmental pressures they impose, making it difficult for enterprises to balance economic benefits with environmental responsibility. Simultaneously, when systems operate at extreme manufacturing stages, pursuing even minor performance improvements often requires disproportionate resource consumption, disrupting the balance between economic benefits and environmental costs. These problems severely restrict the stable development of manufacturing systems towards intelligent and sustainable manufacturing.

[0003] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] The purpose of this invention is to provide a data-driven method for evaluating and improving the sustainability of manufacturing systems, in order to solve the problems mentioned in the background art.

[0005] The technical solution of the present invention includes the following specific steps:

[0006] S1. Obtain process yield and material consumption data, and construct environmental load rate based on the material consumption data and the preset energy conversion rate database;

[0007] S2. Obtain photoresist sensitivity and production cycle time compliance, establish a nonlinear coupling relationship model to characterize the coupling relationship between process yield and environmental load rate; and establish a sustainability index calculated from emergy output rate and environmental load rate.

[0008] S3. Construct a digital twin model that takes controllable process parameters as input and outputs predicted process yield and predicted environmental load rate; then solve for the optimal process parameters to maximize the sustainability index through a reinforcement learning decision engine, and generate process recipes to achieve closed-loop control.

[0009] Preferably, S1 includes:

[0010] S11. Obtain process yield from the manufacturing execution system;

[0011] S12. Obtain material consumption data from the enterprise resource planning system;

[0012] S13. Multiply the acquired material consumption data with the preset energy conversion rate database to calculate the total energy value of consumed local non-renewable resources, the total energy value of purchased resources, and the total energy value of utilized local renewable resources.

[0013] S14. Divide the sum of the total energy value of consumed local non-renewable resources and the total energy value of purchased resources by the total energy value of utilized local renewable resources to construct the environmental load rate.

[0014] Preferably, S2 includes:

[0015] S21. By performing multivariate nonlinear regression fitting on historical production data, the undetermined parameters of the nonlinear coupling relationship model are determined. The historical production dataset contains multiple sets of actual process yield and environmental load rate data collected under different photoresist sensitivity and production cycle compliance conditions.

[0016] S22. Determine the total energy contained in a unit of qualified chip and obtain the total chip processing throughput per unit time.

[0017] S23. Multiply the process yield by the total energy value contained in a unit of qualified chip and the total chip processing throughput per unit time to construct the energy value yield.

[0018] S24. Divide the constructed energy output rate by the environmental load rate to establish a sustainability index.

[0019] Preferably, the nonlinear coupling relationship model includes a first term and a second term; the first term characterizes the effect of fixed energy input being distributed among qualified products; the second term describes the exponential growth effect of marginal energy input after the process yield exceeds the dynamic threshold; the dynamic threshold is dynamically adjusted according to the photoresist sensitivity and production cycle matching corresponding to the production batch.

[0020] Preferably, the dynamic threshold is constructed as a linear model, which includes a threshold reference value and an adjustment term consisting of the difference between the photoresist sensitivity and the reference value and the difference between the production cycle compliance and the reference value.

[0021] Preferably, constructing a digital twin model includes:

[0022] S31. Using controllable process parameters as input and predicted process yield and predicted environmental load rate as output, a machine learning algorithm is trained on historical production data to construct the system.

[0023] S32. Periodically calibrate the constructed digital twin model using real-time data streams from the production line.

[0024] Preferably, solving for optimal process parameters using a reinforcement learning decision engine includes:

[0025] S33. Set the state space of the reinforcement learning decision engine. The state space includes the current controllable process parameters, photoresist sensitivity, and production cycle compliance.

[0026] S34. Set the action space of the reinforcement learning decision engine as the adjustment vector of controllable process parameters;

[0027] S35. Set the reward function of the reinforcement learning decision engine to be the increment of the sustainability index;

[0028] S36. Using a digital twin model as the simulation environment, perform virtual simulations to learn and determine the strategy for recommending the optimal adjustment actions, and then determine the optimal process parameters.

[0029] Preferably, the production process formulation to achieve closed-loop control includes:

[0030] S37. Generate a new process formula from the solved optimal process parameters and issue it for execution;

[0031] S38. Monitor the new state of the manufacturing system after executing the new process formula in real time, and use the new state as the input of the reinforcement learning decision engine to solve iteratively, so that the manufacturing system converges and maintains at the operating point with the maximum sustainability index.

[0032] This invention provides an improved data-driven method for evaluating and enhancing the sustainability of manufacturing systems, which has the following improvements and advantages compared to existing technologies:

[0033] 1. This invention establishes a technical system capable of scientifically quantifying and uniformly evaluating the economic benefits and environmental costs of manufacturing systems. Existing technologies struggle to coordinate and optimize conflicting economic and environmental objectives, primarily due to the lack of a unified benchmark. This invention acquires process yield and material consumption data from Manufacturing Execution Systems (MES) and Enterprise Resource Planning (ERP) systems, and creatively introduces emergy theory to uniformly convert diverse resource consumption from various sources into standardized emergy values. Based on this, an environmental load rate is constructed to scientifically characterize the system's dependence on scarce resources. Furthermore, an emergy output rate is constructed to quantify the rate at which the system creates effective value, and finally, the two are combined to establish a single, clear sustainability index. This index provides a clear objective function for system optimization, solving the problem that traditional methods cannot comprehensively evaluate multi-dimensional performance.

[0034] 2. This invention profoundly reveals and accurately models the inherent nonlinear relationship between economic benefits and environmental costs in high-end manufacturing processes. Traditional models often fail to accurately reflect the true cost of yield improvement under extreme manufacturing conditions. The nonlinear coupling model constructed in this invention accurately depicts the scale effect of fixed energy input being distributed among qualified products. More importantly, it describes the exponential growth effect of marginal energy input after the process yield exceeds a dynamic threshold. This design profoundly reflects the physical reality of paying a disproportionate environmental cost for pursuing minute yield improvements. Furthermore, the dynamic threshold of this model can be adaptively adjusted based on external disturbances such as photoresist sensitivity and production cycle compliance, enabling the model to closely fit the dynamic changes in the actual production line, and its descriptive accuracy far exceeds that of existing technologies.

[0035] 3. This invention constructs an active optimization and continuous improvement scheme from intelligent prediction to closed-loop control. Existing technologies mostly rely on expert experience or high-cost offline experiments, lacking online, adaptive optimization methods. This invention constructs a digital twin model with controllable process parameters as input, transforming the complex manufacturing process into a predictable and simulable white-box system, and uses a periodic calibration mechanism to ensure its long-term accuracy. Based on this, using the digital twin model as a safe virtual simulation environment, a reinforcement learning decision engine is used for efficient optimization. This engine, by perceiving the state space containing external disturbances and using the incremental sustainability index as a reward, can autonomously learn and determine the optimal process parameters. The optimal parameters are used to generate a new process formula for execution, and the new state is monitored and fed back to the decision engine in real time, forming a complete closed-loop control system of decision-making-execution-monitoring-re-decision. This system endows the manufacturing system with the ability of self-evolution and dynamic optimization, enabling it to autonomously converge and stably maintain the operating point with the maximum sustainability, realizing true intelligent and sustainable manufacturing. Attached Figure Description

[0036] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0037] Figure 1 This is a flowchart of the system of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0039] Example 1

[0040] Please see Figure 1 This invention provides a data-driven method for evaluating and improving the sustainability of manufacturing systems, the specific steps of which include:

[0041] S1. Obtain process yield and material consumption data, and construct environmental load rate based on the material consumption data and the preset energy conversion rate database;

[0042] Automation of data acquisition: In practice, data acquisition from manufacturing execution systems and enterprise resource planning systems is usually achieved by establishing data interfaces or APIs, thereby ensuring the real-time nature and accuracy of the data;

[0043] S2. Obtain photoresist sensitivity and production cycle time compliance, establish a nonlinear coupling relationship model to characterize the coupling relationship between process yield and environmental load rate; and establish a sustainability index calculated from emergy output rate and environmental load rate.

[0044] S3. Construct a digital twin model that takes controllable process parameters as input and outputs predicted process yield and predicted environmental load rate; then solve for the optimal process parameters to maximize the sustainability index through a reinforcement learning decision engine, and generate process recipes to achieve closed-loop control.

[0045] In this embodiment, the photoresist sensitivity S m Conformity τ with production cycle time p As an external disturbance variable, it is obtained through real-time data acquisition channels such as manufacturing execution systems and sensor networks, and is used to dynamically adjust the nonlinear coupling relationship model.

[0046] This embodiment provides a data-driven method for evaluating and improving the sustainability of manufacturing systems. Its purpose is to address the technical problems in existing technologies, such as the difficulty in synergistically optimizing economic benefits and environmental costs, and the lack of closed-loop adaptive control mechanisms, by accurately quantifying and proactively optimizing the sustainability of manufacturing systems. This embodiment constructs a complete technical framework from data acquisition and modeling evaluation to intelligent decision-making and closed-loop control, including the following steps:

[0047] Step S1: Acquire process yield and material consumption data, and construct an environmental load rate based on this material consumption data and a preset emergy conversion rate database. The purpose of this step is to provide basic economic and environmental data input for subsequent sustainability assessment. In this embodiment, by collecting these two types of core data from the manufacturing information system and introducing emergy theory, material consumption from different categories and sources is uniformly converted into standardized emergy values, thereby constructing an environmental load rate that can scientifically and cross-category measure the system's real environmental pressure. The construction of this indicator is the foundation for achieving multi-dimensional comprehensive performance evaluation.

[0048] Simultaneously, the photoresist sensitivity S corresponding to this production batch is obtained. m Conformity τ with production cycle time p External disturbance data;

[0049] Step S2: Obtain photoresist sensitivity and production cycle time compliance, and establish a nonlinear coupling model to characterize the relationship between process yield and environmental load rate; and establish a sustainability index calculated from energy output rate and environmental load rate; the underlying logic of this step is to accurately characterize economic benefits, with process yield Y representing the environmental cost, and environmental load rate L representing the environmental cost. E To represent the complex nonlinear relationships between components and to account for the impact of external disturbances, this embodiment establishes a nonlinear coupling model that reveals a key physical phenomenon: the disproportionate environmental cost incurred in pursuing minute yield improvements during the extreme manufacturing stage. Subsequently, to form a single, clear optimization objective, this embodiment further constructs an emergy output rate, which quantifies the rate at which the system creates effective value. Finally, a sustainability index I is established by comparing its ratio to the environmental load rate. S The physical meaning of this index is that a truly sustainable system should have a rate of effective value generation that is far higher than its environmental impact.

[0050] Step S3: Construct a digital twin model that takes controllable process parameters as input and outputs predicted process yield and predicted environmental load rate; then, use a reinforcement learning decision engine to solve for the optimal process parameters to maximize the sustainability index and generate process formulations to achieve closed-loop control; the purpose of this step is to connect the technical link from evaluation to improvement and achieve proactive system optimization; in this embodiment, by constructing a digital twin model, with controllable process parameters and the indirect system core state, Y and L, E A precise mapping relationship was established between them, enabling the prediction of its impact on the sustainability index I before process adjustments. S This makes the impact possible; based on this, by utilizing reinforcement learning decision engines—an advanced artificial intelligence decision-making technology—and using digital twin models as a simulation environment for efficient virtual deduction, it is possible to quickly and accurately solve for the factors that enable I... S Maximize the optimal process parameters; automatically generate process formulas from these optimal parameters and send them to the manufacturing system for execution, while monitoring new states in real time and feeding them back to the decision engine, forming a closed-loop control system that enables the system to autonomously and dynamically converge and stably operate at the sustainable optimal point.

[0051] This invention, through the organic combination of the above steps, establishes a complete and closed-loop technical system for the sustainable evaluation and proactive improvement of manufacturing systems. It not only solves the problem of traditional methods failing to uniformly measure economic and environmental benefits, but also profoundly reveals the inherent laws between the two through a nonlinear coupling model. Furthermore, by combining digital twins and reinforcement learning, it provides a closed-loop control method capable of autonomously optimizing and adapting to external disturbances in real time. Compared to existing technologies, this invention can significantly improve the comprehensive performance of manufacturing systems in dynamic environments, ensuring high yield while minimizing dependence on environmental resources, thus achieving truly intelligent and sustainable manufacturing.

[0052] This method focuses on key technological bottlenecks in the sustainability assessment and improvement of manufacturing systems. Through modeling and optimizing core indicators such as process yield and material consumption, it achieves significant technological breakthroughs. Although the model is simplified in real-world applications, it provides an efficient and feasible technical framework for addressing the core challenges currently facing manufacturing systems and lays the foundation for subsequent, more comprehensive, and complex model extensions.

[0053] Example 2

[0054] S1 includes:

[0055] S11. Obtain process yield from the manufacturing execution system;

[0056] S12. Obtain material consumption data from the enterprise resource planning system;

[0057] S13. Multiply the acquired material consumption data with the preset energy conversion rate database to calculate the total energy value of consumed local non-renewable resources, the total energy value of purchased resources, and the total energy value of utilized local renewable resources.

[0058] S14. Divide the sum of the total energy value of consumed local non-renewable resources and the total energy value of purchased resources by the total energy value of utilized local renewable resources to construct the environmental load rate.

[0059] This embodiment, based on the scheme of Embodiment 1, further defines and explains the implementation method of constructing the environmental load rate in step S1, aiming to ensure the reliability of the data source and the standardization of the calculation process, thereby ensuring L E The accuracy and reproducibility of the indicators;

[0060] The significance of environmental load factor: Environmental load factor L E The core significance lies in quantifying the material consumption from different sources and of different natures, such as chemicals, electricity, and water, into a unified measure of the real pressure on the environment, thereby achieving a comprehensive evaluation across categories and dimensions.

[0061] S11. Obtain process yield from the Manufacturing Execution System (MES); the MES is an information system used to monitor and manage real-time events on the production line; its function is to provide the most direct and reliable economic benefit data for this invention; in this embodiment, the process yield Y is precisely defined as the effective chip output quantity N. p Total number of chips N in wafer design t The ratio, which is automatically calculated and provided by the MES system after each batch of wafers is produced;

[0062] S12. Obtain material consumption data from the Enterprise Resource Planning (ERP) system. The ERP system is a management software platform that integrates enterprise logistics, capital flow, and information flow. Its function is to provide comprehensive and accurate production factor consumption data for this invention. In this embodiment, the material consumption data includes, but is not limited to, the specific quantities of various chemicals, ultrapure water, special gases, electricity, and other resources consumed in the production process. These data are obtained from the material management and cost accounting modules of the ERP system.

[0063] S13. Multiply the acquired material consumption data with the preset energy value conversion rate database to calculate the total energy value of consumed local non-renewable resources, the total energy value of purchased resources, and the total energy value of utilized local renewable resources. The energy value conversion rate database is a database that converts different material or energy units into a unified solar joule standard energy value unit. This database is established based on publicly available and industry-recognized data.

[0064] You can refer to the "China Energy Value Database" or authoritative international energy value analysis reports, and update them regularly to reflect the latest energy structure and resource utilization efficiency.

[0065] Sources of emergy conversion rate databases: emergy conversion rate databases are typically built based on industry-recognized data published by international authoritative institutions or research literature;

[0066] The purpose of this step is to standardize and quantify material consumption data from completely different sources and of different natures, making them additive. In this embodiment, the material consumption data obtained from the ERP system is classified into local non-renewable resources, such as locally mined minerals, purchased resources, such as electricity and chemicals imported from other regions, and local renewable resources, such as solar and wind energy in the factory area. The consumption of each resource is multiplied by its corresponding conversion rate in the energy conversion rate database, and the results are summed to obtain the total energy value E. n E o and E r ;

[0067] S14. Divide the sum of the total energy value of consumed local non-renewable resources and the total energy value of purchased resources by the total energy value of utilized local renewable resources to construct the environmental load factor; this step is L. EThe final calculation stage of the indicator; in this embodiment, the environmental load rate L E The calculation formula is:

[0068]

[0069] Among them, L E Environmental load factor measures the true stress on the environment; E n E: Total energy value of locally consumed non-renewable resources; o Total energy value of purchased resources; E r : Gross energy value of locally available renewable resources utilized;

[0070] The technical motivation behind this formula lies in the fact that the numerator represents the total consumption of existing, scarce environmental resources, while the denominator represents the capacity to utilize local, renewable resources; therefore, L E The higher the value, the more production activities rely on the consumption of high-quality energy and resources, and the greater the pressure on the environment.

[0071] When E r If the value is less than the preset threshold, it can be replaced with a small positive number, or other weighted averaging methods can be used to ensure that L... E The boundedness and physical rationality of the value;

[0072] This threshold can be preset to 10. -6 Solar Joules are used to ensure that the denominator is not zero, thereby guaranteeing L. E The boundedness and physical rationality of the value;

[0073] Compared to Example 1, this example clarifies the specific sources of process yield and material consumption data, the MES and ERP systems, and elaborates in detail on the construction of environmental load factor L based on emergy theory. E The complete calculation steps and formulas greatly enhance the operability and objectivity of the technical solution; this standardized data processing process ensures the accuracy and reliability of the sustainability evaluation results, provides a high-quality data foundation for subsequent optimization decisions, and avoids model inaccuracies caused by inconsistent data standards or ambiguous evaluation indicator definitions.

[0074] S2 includes:

[0075] S21. By performing multivariate nonlinear regression fitting on historical production data, the undetermined parameters of the nonlinear coupling relationship model are determined. The historical production dataset D hist It includes multiple sets of actual process yield and environmental load data collected under different photoresist sensitivity and production cycle compliance conditions;

[0076] S22. Determine the total energy contained in a unit of qualified chip and obtain the total chip processing throughput per unit time.

[0077] S23. Multiply the process yield by the total energy value contained in a unit of qualified chip and the total chip processing throughput per unit time to construct the energy value yield.

[0078] S24. Divide the constructed energy output rate by the environmental load rate to establish a sustainability index;

[0079] Based on the scheme of Example 1, this embodiment refines the implementation method of establishing the sustainability index in step S2, with the aim of constructing a single optimization objective with clear physical meaning that can uniformly evaluate the overall performance of the system.

[0080] S21. By performing multivariate nonlinear regression fitting on historical production data, the parameters to be determined for the nonlinear coupling relationship model are identified. To further clarify this calibration process, the historical production dataset is defined as follows:

[0081]

[0082] Where Y i With LE i It is in the i-th group of historical production conditions S m,i and τ p,i The actual process yield and environmental load rate collected below; the purpose of this step is to accurately determine the specific form of the nonlinear relationship between Y and LE through a data-driven approach; S m,i τ represents the sensitivity of the photoresist under the historical production conditions of the i-th group. p,i : Production cycle time compliance under the i-th historical production conditions; N: Total number of historical production data sets; By analyzing this dataset D hist By applying multivariate nonlinear regression algorithms such as Levenberg-Marquardt, the undetermined parameters, such as α, k, and various sensitivity coefficients in the nonlinear coupling relationship model detailed later are fitted and solved, thereby obtaining a customized model that can accurately reflect the characteristics of a specific manufacturing system.

[0083] Selection of nonlinear regression algorithm: Multiple nonlinear regression fitting can use a variety of algorithms, such as the Levenberg-Marquardt algorithm, which can efficiently solve for the undetermined parameters in the model;

[0084] Nonlinear coupling relationship model (LE(Y,S)) m ,τ p In the original text, Y and LE are derived from historical production data, while S... m and τ pThese are the external perturbation variables corresponding to these historical data points; the nonlinear coupling model is an empirical or semi-empirical model, which is designed to fit and predict the dimensionless L obtained by calculating the energy ratio in Example 2. E Therefore, the output of the entire model and each term should be dimensionless;

[0085] This model is constructed based on empirical observations and physical intuition, aiming to capture the non-linear relationship between yield improvement and environmental costs. In practical applications, the specific form and parameters of the model can be adjusted based on the fitting effect of historical data to ensure optimal physical fidelity.

[0086] The model is based on physical insights into the semiconductor manufacturing process, especially considering the diminishing marginal returns and exponential growth of resource consumption during yield improvement, thus constructing an empirical model that can simultaneously capture the scale effect and the extreme manufacturing effect.

[0087] S22. Determine the total energy contained in a unit of qualified chip and obtain the total chip processing throughput per unit time; the total energy contained in a unit of qualified chip E c E refers to the total energy value of all resources invested in producing a single qualified chip, and its function is to characterize the value content of a single product; in this embodiment, E c The value is determined by a one-time full life-cycle energy analysis of the product. Once determined, it can be considered a constant, with units of solar joules per chip; the total chip throughput Q per unit time. t This refers to the total number of chips processed by the production line per unit of time, such as one hour, regardless of whether they are good or bad. Its purpose is to characterize the physical processing rate of the production line. This parameter can be obtained in real time from the MES system, and the unit is the number of chips per hour.

[0088] S23. Multiply the process yield by the total energy contained in a unit of qualified chip and the total chip processing throughput per unit time to construct the energy yield; the purpose of this step is to construct an indicator that can represent the rate at which the system creates effective value; in this embodiment, the energy yield R E The calculation formula is clearly defined as:

[0089] R E =Y·E c ·Q t

[0090] In the formula, the process yield Y is provided in real time by the manufacturing execution system, and E... c Q is a pre-calibrated constant. t Provided in real time by the manufacturing execution system; the technical motivation for this formula lies in Y·Q tThe physical meaning is the production rate of qualified chips, with units of: number of qualified chips / hour, multiplied by the energy value E of a single qualified chip. c The obtained R E This represents the total value stream created by the system per unit time, measured in energy value, with the dimension of solar joules per hour;

[0091] Dimensions of the Sustainability Index: Sustainability Index I S The dimension of the energy yield is the same as that of solar joules per hour, because it is derived from R. E Divide by dimensionless L E Obtained;

[0092] The dimension of relevance productivity: relevance productivity R E The unit of measurement is solar joules per hour, which represents the total value stream created by the system per unit time.

[0093] S24. Divide the constructed energy output rate by the environmental load rate to establish a sustainability index; this step is a key step in integrating the two dimensions of economy and environment to form the final evaluation index; in this implementation, the sustainability index I... S The calculation formula is:

[0094]

[0095] In the formula, the energy output rate R E The environmental load rate L is calculated from the preceding step S23. E Calculated from the preceding step S1; the logic behind this formula is that for a sustainable system, the rate of value creation R... E It should be much higher than its environmental resource burden L. E ; due to L E It is dimensionless, I S Dimensions and R E The same, in solar joules per hour; this single, clear scalar I. S This provides a clear objective function that needs to be maximized for subsequent optimization decision-making steps; Y: process yield, representing economic benefits; E c Q: Total energy contained in a single qualified chip; t Total chip throughput per unit time;

[0096] Within the same production cycle, obtain the process yield Y and the total throughput Q for that batch from the MES system. t The product of the two, Y and Q t This represents the production rate of qualified chips within that time period;

[0097] This embodiment is illustrated by I. SThe construction process will integrate conflicting economic indicators Y and environmental indicators L. E Successfully integrated into a single optimization objective; by defining the energy output rate R with clear physical meaning. E This makes the final sustainability index I... S It is not only valid in mathematics, but also has profound implications in physics and economics, representing the rate of value output that can be leveraged by a unit of environmental load; this provides a high-quality reward function for subsequent reinforcement learning, making the optimization direction clearer and more efficient, and thus more accurately guiding the manufacturing system towards the direction of optimal overall performance.

[0098] Example 3

[0099] The nonlinear coupling model includes a first term and a second term; the first term characterizes the effect of fixed energy input being distributed among qualified products; the second term describes the exponential growth effect of marginal energy input after the process yield exceeds a dynamic threshold; the dynamic threshold is based on the photoresist sensitivity S corresponding to that production batch. m Conformity τ with production cycle time p Dynamic adjustment is performed; this model is a mathematical model for prediction, and its output L E The value is the same as the actual L calculated by the energy ratio in Example 2. E The values ​​are numerically equal;

[0100] The dynamic threshold is constructed as a linear model, which includes a threshold baseline and adjustment terms consisting of the difference between the photoresist sensitivity and the baseline and the difference between the production cycle compliance and the baseline.

[0101] Based on the scheme of Example 1, this embodiment provides a preferred and feasible limitation on the mathematical structure of the nonlinear coupling relationship model and the construction method of the dynamic threshold therein; this design aims to more accurately capture and quantify the special relationship between yield improvement and environmental cost under extreme manufacturing conditions.

[0102] The nonlinear coupling relationship model aims to mathematically express the environmental load rate L. E How does it vary with process yield Y, and how does it account for external disturbance S? m and τ p The regulating effect; in this embodiment, the model is used to predict the dimensionless environmental load rate L calculated in step S1. E The model is constructed in the following form:

[0103]

[0104] In the formula, L EEnvironmental load factor, a function of process yield, photoresist sensitivity, and production cycle time compliance, where Y is the real-time process yield and S... m With τ p External disturbance variables acquired in real time; parameters α, k, and functions β and Y in the model. th The undetermined coefficients were all determined through regression fitting in step S21 above, and all were dimensionless parameters; β(S m ,τ p Y: Dimensionless growth coefficient adjusted by external disturbances; th (S m ,τ p ): Dynamic threshold, the yield inflection point where marginal environmental load begins to rise sharply, dynamically adjusted by photoresist sensitivity and production cycle time compliance; k: the power of the exponential growth effect; max(0,YY) th (S m ,τ p )) k When Y>Y th When this occurs, the second item takes effect, describing the exponential growth of the marginal environmental load;

[0105] This model contains the first and second terms:

[0106] The first term, α / Y, characterizes the effect of the basic environmental load associated with fixed inputs being amortized by qualified products. Here, α is a dimensionless coefficient characterizing the basic contribution of fixed production costs to the environmental load. This coefficient is related to the total energy value of fixed inputs and a certain baseline output energy value, thereby making... Physically, this corresponds to a fixed environmental load ratio allocated to a unit of qualified products. As the number of qualified products increases, proportional to the increase in yield Y, this contribution decreases, reflecting the positive effect of economies of scale on reducing the environmental load ratio per unit of product.

[0107] The second term β(S) m ,τ p )·max(0,YY th (S m ,τ p )) k Its function is to describe the exponential growth effect of marginal environmental load after the process yield exceeds a dynamic threshold; this profoundly reveals a key technical issue: in high-end manufacturing, once the yield reaches a certain level, further small improvements, i.e., Y>Y, are impossible. th This often requires a disproportionate amount of additional resources, causing the marginal environmental load to increase dramatically in the form of a k-th power; among which, β(S m ,τ p () is the dimensionless growth coefficient that is adjusted by external disturbances;

[0108] Among them, the dynamic threshold Y th It is a core concept, referring to the yield inflection point where the marginal environmental load begins to rise sharply;

[0109] The physical significance of this dynamic adjustment mechanism lies in the fact that changes in the external environment, such as the photoresist sensitivity S... m Fluctuations or production cycle conformity τ p Deviations in Y-axis photoresist directly impact the complexity of the manufacturing process and resource consumption. For example, when the sensitivity of the photoresist decreases, maintaining yield may require increasing the exposure dose, leading to increased energy and material consumption and causing the yield inflection point to arrive earlier. Therefore, adjusting the Y-axis photoresist concentration is crucial. th The design uses dynamically adjustable variables, enabling the model to more realistically reflect the actual behavior of the manufacturing system under different external conditions.

[0110] In this embodiment, the reference value Y th,0 S m,0 and τ p,0 By analyzing historical production dataset D hist The value is determined through statistical analysis. For example, the median or mean of the respective data can be used as a benchmark. It is recommended to prioritize the median to reduce the impact of outliers on the benchmark value, thereby making the model more physically meaningful and interpretable. Similarly, the growth coefficient β(S) in the second term... m ,τ p It can also be constructed as a similar linear model, whose parameters are also obtained by fitting historical data;

[0111] Historical production dataset D can be retrieved. hist The median or mean of each data point is used as the baseline value Y. th,0 S m,0 and τ p,0 ;

[0112] The innovation of this invention lies in the fact that the threshold is not a fixed value, but rather depends on the photoresist sensitivity S. m Conformity τ with production cycle time p Dynamic adjustments were made; to ensure the feasibility of the solution, further optimizations were implemented; the dynamic threshold Y... th It was constructed as a linear model with the aim of quantifying S in a simple and effective way. m and τ p The regulating effect; the linear model includes a threshold benchmark value Y. th,0 And an adjustment term consisting of the difference between the photoresist sensitivity and the reference value, and the difference between the production cycle time compliance and the reference value; the formula is as follows:

[0113] Y th (S m ,τ p ) = Yth,0 +λ s (S m -S m,0 )+λ τ (τ p -τ p,0 )

[0114] Among them, Y th (S m ,τ p ): Dynamic threshold, a linear model that varies with photoresist sensitivity and production cycle time compliance;

[0115] In this model, Y th,0 S m,0 τ p,0 These are the baseline values ​​for the threshold, photoresist sensitivity, and production cycle time compliance, respectively, obtained from statistical analysis of historical production data; λ s and λ τ It is a dimensionless sensitivity coefficient, reflecting S m and τ p Fluctuation on threshold Y th The degree of influence; similarly, the growth coefficient β(S) in the second term m ,τ p It can also be constructed as a similar linear model;

[0116] By constructing a specific form of nonlinear coupling model that includes both apportionment and marginal exponential growth effects and ensures dimensional consistency, this invention can reflect the real physical processes of high-end manufacturing more profoundly and accurately than traditional linear or simple nonlinear models; in particular, by using the key threshold Y... th The model is designed as a variable that is dynamically adjusted by external disturbances, and a specific linearization implementation scheme is given. This makes the model not only highly interpretable, but also able to adapt to changes in the production line status in real time. This refined modeling greatly improves the prediction accuracy of the subsequent digital twin model and lays a solid model foundation for achieving precise closed-loop optimization control.

[0117] Building a digital twin model includes:

[0118] S31. Using controllable process parameters as input and predicted process yield and predicted environmental load rate as output, a machine learning algorithm is trained on historical production data to construct the system.

[0119] In this embodiment, the digital twin model can be implemented using a feedforward neural network with multiple hidden layers; the input layer receives controllable process parameters P. in The hidden layers in the middle use non-linear activation functions, such as ReLU, to handle complex mappings, while the output layers predict the process yield Y. predand environmental load factor LE pred The model includes historical data pairs (P). in ,Y actual ,LE actual The model is trained on a large dataset to minimize the mean square error between the predicted and actual values, thereby ensuring that the model can accurately map from process parameters to system state.

[0120] A feedforward neural network with three hidden layers, each containing 64, 32, and 16 neurons respectively, can be used. ReLU is used as the activation function, the Adam optimizer is employed, and the initial learning rate is set to 0.001 for training.

[0121] This digital twin model uses controllable process parameters as input to predict the process yield Y. pred And the predicted dimensionless environmental load factor (LE) pred For output; where LE pred It is calculated using a nonlinear coupling relationship model, aiming to predict actual production conditions based on the formula LE = (E n +E o ) / E r The resulting dimensionless environmental load factor; this method transforms complex physical processes into predictable mathematical models, thereby unifying the definition of LE across different embodiments;

[0122] S32. Periodically calibrate the constructed digital twin model using real-time data streams from the production line;

[0123] Based on the scheme of Example 1, this embodiment details the specific method of constructing a digital twin model in step S3, aiming to establish a high-precision prediction model that connects process parameters that can be directly controlled by front-line engineers with advanced system status indicators.

[0124] S31. Using controllable process parameters as input and predicted process yield and predicted environmental load rate as output, a digital twin model is constructed by training a machine learning algorithm on historical production data. The core purpose of the digital twin model is to answer the key question of how the yield and environmental load will change if the process parameters are adjusted, thereby enabling the prediction of the consequences of process adjustment.

[0125] Input: The input to the model is the controllable process parameter P. in ;P in It is a parameter vector whose elements can include the exposure dose D. e Etching time T e Gas flow rate F g Parameters that frontline engineers can set directly on the equipment control interface;

[0126] Output: The model output is the predicted process yield Y. pred and predicted environmental load rate L E,pred ;

[0127] Construction Method: This model is obtained by training a machine learning algorithm on historical production data; the historical production data contains a large amount of (P) in ,Y actual ,L E,actual Data pairs; in this embodiment, algorithms with strong nonlinear fitting capabilities, such as deep neural networks and gradient boosting decision trees, can be selected;

[0128] When building a digital twin model, machine learning algorithms such as deep neural networks or gradient boosting decision trees can be used. These algorithms are good at handling complex nonlinear mapping relationships.

[0129] The training process is to allow the machine learning algorithm to learn from P in To (Y) actual ,L E,actual The complex mapping relationship;

[0130] S32. Periodically calibrate the constructed digital twin model using real-time production line data streams; the manufacturing system is a dynamically evolving system. To address concept drift caused by factors such as equipment aging and material batch changes, and to ensure the long-term predictive accuracy of the digital twin model, continuous maintenance is necessary; the purpose of periodic calibration is to update or fine-tune the trained model using the latest production line data; in this embodiment, the system continuously collects new (P) data generated by the production line. in ,Y actual ,L E,actual Data; Based on a preset calibration strategy, such as a fixed time interval, such as weekly or a fixed amount of data, such as every 100 batches of data accumulated, periodic calibration can be triggered based on a preset time interval, such as weekly or a fixed amount of data, such as every 100 batches of data accumulated. The newly accumulated data in this period is used to incrementally train or fine-tune the weights of the neural network model, so that the model can continuously adapt to the latest state of the production line.

[0131] This embodiment details the construction and maintenance mechanism of the digital twin model. By using machine learning algorithms and historical data, it establishes a data-driven bridge between directly controllable process parameters and the final sustainability evaluation indicators, transforming the black-box manufacturing system into a predictable and simulable white-box model. In particular, the introduction of a periodic calibration mechanism ensures the dynamic adaptability of the model, enabling it to evolve synchronously with the physical entity. This provides a consistently accurate and reliable simulation environment for the reinforcement learning engine, which is a technical prerequisite for achieving efficient and robust closed-loop control.

[0132] Example 4

[0133] Solving for optimal process parameters using a reinforcement learning decision engine includes:

[0134] S33. Set the state space of the reinforcement learning decision engine. The state space includes the current controllable process parameters, photoresist sensitivity, and production cycle compliance.

[0135] S34. Set the action space of the reinforcement learning decision engine as the adjustment vector of controllable process parameters;

[0136] S35. Set the reward function of the reinforcement learning decision engine to be the increment of the sustainability index;

[0137] S36. Using a digital twin model as a simulation environment, perform virtual simulations to learn and determine strategies for recommending optimal adjustment actions, and then determine the optimal process parameters.

[0138] Based on the scheme of Example 1, this embodiment elaborates on the specific configuration of solving the optimal process parameters through a reinforcement learning decision engine; this step is the core of the present invention to realize the transition from evaluation to proactive improvement, and its purpose is to design an intelligent agent that can autonomously learn the optimal control strategy.

[0139] The reinforcement learning decision engine is a goal-oriented decision-making system that learns how to make optimal decisions through interaction with the environment, which in this invention is a digital twin model; the configuration is as follows:

[0140] S33. Define the state space of the reinforcement learning decision engine; the purpose of the state space is to define all the information that the agent needs to observe when making a decision; a well-defined state space should be Markovian, that is, the current state contains all the historical information needed to make the optimal decision; in this embodiment, the state space is set to include the current controllable process parameter P. in Photoresist sensitivity S m Conformity τ with production cycle time p External disturbance S m and τ p Incorporating the state space allows the agent's decisions to proactively adapt to these external changes, rather than simply optimizing for fixed process parameters; photoresist sensitivity S m Conformity τ with production cycle time p Acquired in real time from the manufacturing execution system or collected via sensor networks;

[0141] S34. The action space of the reinforcement learning decision engine is defined as the adjustment vector of the controllable process parameters; the action space defines the set of operations that the agent can perform; in this embodiment, the action space is set as the adjustment vector ΔP of the controllable process parameters; for example, the action could be to increase the exposure dose by 5 mJ / cm 2 The etching time is reduced by 0.5 seconds; the goal of the agent's learning is to output the optimal ΔP based on the current state, so that the adjusted new process parameter P... in′ =P in +ΔP can bring the greatest long-term returns;

[0142] S35. The reward function of the reinforcement learning decision engine is set to the increment of the sustainability index; the role of the reward function is to provide immediate feedback on the agent's actions and is a quantitative indicator guiding its learning direction; in this embodiment, the reward function is directly set to the increment of the sustainability index:

[0143] ΔI S =I S,new -I S,current

[0144] Where, ΔI S The increment of the sustainability index reinforces the reward function of the learning decision engine; I S,new The new sustainability index after adjusting process parameters; I S,current Current sustainability index;

[0145] When an action ΔP of the agent causes a decrease in the sustainability index I... S When the reward increases, it receives a positive reward; conversely, it receives a negative reward. The learning objective of the agent is to maximize the sum of future cumulative rewards.

[0146] S36. Using a digital twin model as the simulation environment, perform virtual simulations to learn and determine the strategy for recommending the optimal adjustment action; this step is the training process of reinforcement learning; since repeated trial and error on a real production line is extremely costly and risky, this invention utilizes a digital twin model as the simulation environment; in the virtual environment, the agent, based on the current state:

[0147] S t =(P in ,S m ,τ p )

[0148] Among them, S t : The state space of the reinforcement learning decision engine at time t; P in : Current controllable process parameters; S m Current photoresist sensitivity; τ p Current production cycle time compliance;

[0149] Perform an action A t =ΔP, the digital twin model will immediately predict Y in the new state. pred and L E,pred This leads to the calculation of a new sustainability index, I. S,pred and reward R t By conducting extensive virtual simulations in this zero-cost, high-efficiency virtual environment, intelligent agents, for example using algorithms such as deep Q-networks or proximal policy optimization, can gradually learn and determine an optimal policy π. * (S);

[0150] For discrete process parameter adjustments, deep Q-networks can be used first; for continuous parameter adjustments, algorithms based on policy gradients, such as proximal policy optimization, can be used, and appropriate discount factors and learning rates should be set to ensure the stability and convergence of training.

[0151] In reinforcement learning decision engines, advanced algorithms such as deep Q-networks and proximal policy optimization can be selected to learn the optimal policy.

[0152] This strategy is a function that, for any input state S, outputs the optimal action A that yields the maximum long-term reward; this learned strategy is then used to recommend the optimal adjustment action, thereby determining the optimal process parameters P. opt ;

[0153] This embodiment, through careful design of the three elements of reinforcement learning—state, action, and reward—and combined with a digital twin model as a simulation environment, constructs an efficient, safe, and intelligent process parameter optimization scheme. It transforms the complex, multivariate process optimization problem into a well-defined sequential decision problem. Compared to traditional methods that rely on expert experience or involve time-consuming and labor-intensive experimental design, the method of this invention can autonomously and quickly explore a vast parameter space, find non-intuitive optimal combinations of process parameters, and dynamically adapt to changes in the external environment, thereby achieving proactive and continuous improvement of the manufacturing system's sustainability.

[0154] Before deploying a model in real-world applications, it undergoes a robustness check. This process involves making predictions using input values ​​outside the range of historical data and verifying that its outputs conform to physical common sense. For example, when the process yield Y is close to 100%, the model predicts an environmental load rate L. E It should increase dramatically to conform to the physical phenomenon of marginal effect.

[0155] The process formulation for achieving closed-loop control includes:

[0156] S37. Generate a new process formula from the solved optimal process parameters and issue it for execution;

[0157] S38. Monitor the new state of the manufacturing system after executing the new process formula in real time, and use the new state as the input of the reinforcement learning decision engine to solve iteratively, so that the manufacturing system converges and maintains at the operating point with the maximum sustainability index.

[0158] Based on the scheme of Example 1, this embodiment describes the specific process of generating process formula to achieve closed-loop control; this is the last step of applying the optimization results of reinforcement learning engine to actual production and forming a continuous improvement cycle, which is the key to the implementation of the technical solution of this invention.

[0159] S37. Generate a new process formula from the solved optimal process parameters and issue it for execution; Optimal process parameters P opt This is the result obtained by the reinforcement learning decision engine in the aforementioned steps; the process recipe is a set of specific instructions that the manufacturing equipment can recognize and execute; the purpose of this step is to transform the abstract parameter vector P... opt This is converted into actual production instructions; in this implementation, the system will convert P... opt The values ​​of each parameter in, such as D e =120mJ / cm 2 T e =30s, formatted into a process formula file that conforms to the specific device communication protocol; through the factory's automation system, this new process formula is automatically distributed to the corresponding production equipment for execution in the next production batch;

[0160] S38. Monitor the new state of the manufacturing system after executing the new process formula in real time, and use the new state as the input of the reinforcement learning decision engine to solve iteratively, so that the manufacturing system converges and maintains at the operating point with the maximum sustainability index; the purpose of this step is to build a complete feedback loop to achieve continuous optimization and adaptive stability of the system.

[0161] Real-time monitoring: After the new process formula is implemented, the system will collect the actual new status of the batch of products in real time through information systems such as MES and ERP. This includes the actual process yield Y and the actual material consumption data, which are used to calculate L. E and S during the execution period m and τ p wait;

[0162] Iterative solution: This newly detected state (P) opt ,Y actual ,L E,actual ,S m ,τ p This will be fed back to the system; on the one hand, it can serve as new data points for the periodic calibration of the digital twin model; on the other hand, this new state, especially P... opt ,Sm ,τ p Part of it will serve as the new current state input for the reinforcement learning decision engine; based on this new state, the engine will recommend an optimal adjustment action again, thereby initiating a new round of optimization iterations;

[0163] Convergence and Maintenance: Through this continuous cycle of decision-making, execution, monitoring, and re-decision-making, the manufacturing system, guided by this closed-loop control mechanism, will gradually converge to a state where the sustainability index I... S The system achieves the highest stable operating point; furthermore, when external disturbances occur, the closed-loop system can automatically sense changes in state and make corresponding adjustments, thereby maintaining itself at the new optimal operating point, demonstrating strong adaptability and robustness.

[0164] This embodiment connects all the aforementioned data acquisition, modeling, evaluation, and decision-making stages into a truly operable closed-loop control system by clearly defining two key actions: formula issuance and status feedback. This not only enables the implementation of optimization results but also endows the manufacturing system with the ability to self-evolve and dynamically optimize through real-time feedback and iterative solution mechanisms. This closed-loop control architecture ensures that the technical effect of this invention is not a one-time improvement but a sustainable, online, and adaptive optimization process, thereby enabling the manufacturing system to operate stably in the long term, achieving an optimal balance between economic benefits and environmental friendliness.

[0165] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A data-driven method for evaluating and improving the sustainability of manufacturing systems, characterized in that, The specific steps include: S1. Obtain process yield and material consumption data, and construct environmental load rate based on the material consumption data and the preset energy conversion rate database; S2. Obtain photoresist sensitivity and production cycle time compliance, establish a nonlinear coupling relationship model to characterize the coupling relationship between process yield and environmental load rate; and establish a sustainability index calculated from emergy output rate and environmental load rate. S3. Construct a digital twin model that takes controllable process parameters as input and outputs predicted process yield and predicted environmental load rate; then solve for the optimal process parameters to maximize the sustainability index through a reinforcement learning decision engine, and generate process recipes to achieve closed-loop control. The nonlinear coupling model includes a first term and a second term. The first term characterizes the effect of fixed energy input being distributed among qualified products. The second term describes the exponential growth effect of marginal energy input after the process yield exceeds the dynamic threshold. The dynamic threshold is dynamically adjusted according to the photoresist sensitivity and production cycle matching corresponding to the production batch. The dynamic threshold is constructed as a linear model, which includes a threshold baseline and adjustment terms consisting of the difference between the photoresist sensitivity and the baseline and the difference between the production cycle compliance and the baseline.

2. The data-driven method for evaluating and improving the sustainability of a manufacturing system according to claim 1, characterized in that, S1 includes: S11. Obtain process yield from the manufacturing execution system; S12. Obtain material consumption data from the Enterprise Resource Planning system; S13. Multiply the acquired material consumption data with the preset energy conversion rate database to calculate the total energy value of consumed local non-renewable resources, the total energy value of purchased resources, and the total energy value of utilized local renewable resources. S14. Divide the sum of the total energy value of consumed local non-renewable resources and the total energy value of purchased resources by the total energy value of utilized local renewable resources to construct the environmental load rate.

3. The data-driven method for evaluating and improving the sustainability of a manufacturing system according to claim 1, characterized in that, S2 include: S21. By performing multivariate nonlinear regression fitting on historical production data, the undetermined parameters of the nonlinear coupling relationship model are determined. The historical production dataset contains multiple sets of actual process yield and environmental load rate data collected under different photoresist sensitivity and production cycle compliance conditions. S22. Determine the total energy contained in a unit of qualified chip and obtain the total chip processing throughput per unit time. S23. Multiply the process yield by the total energy value contained in a unit of qualified chip and the total chip processing throughput per unit time to construct the energy value yield. S24. Divide the constructed energy output rate by the environmental load rate to establish a sustainability index.

4. The data-driven method for evaluating and improving the sustainability of a manufacturing system according to claim 1, characterized in that, Building a digital twin model includes: S31. Using controllable process parameters as input and predicted process yield and predicted environmental load rate as output, a machine learning algorithm is trained on historical production data to construct the system. S32. Periodically calibrate the constructed digital twin model using real-time data streams from the production line.

5. The data-driven method for evaluating and improving the sustainability of a manufacturing system according to claim 1, characterized in that, Solving for optimal process parameters using a reinforcement learning decision engine includes: S33. Set the state space of the reinforcement learning decision engine. The state space includes the current controllable process parameters, photoresist sensitivity, and production cycle compliance. S34. Set the action space of the reinforcement learning decision engine as the adjustment vector of controllable process parameters; S35. Set the reward function of the reinforcement learning decision engine to be the increment of the sustainability index; S36. Using a digital twin model as the simulation environment, perform virtual simulations to learn and determine the strategy for recommending the optimal adjustment actions, and then determine the optimal process parameters.

6. The data-driven method for evaluating and improving the sustainability of a manufacturing system according to claim 1, characterized in that, The process formulation for achieving closed-loop control includes: S37. Generate a new process formula from the solved optimal process parameters and issue it for execution; S38. Monitor the new state of the manufacturing system after executing the new process formula in real time, and use the new state as the input of the reinforcement learning decision engine to solve iteratively, so that the manufacturing system converges and maintains at the operating point with the maximum sustainability index.

Images