Method for ensuring product quality in production process, system for developing or improving the same, method for training data model structure, related computer program, and computer-readable storage medium

Abstract

A method for developing or improving a process for producing a product from a material includes the following steps: acquiring raw material data (21, 21a, 21b) for the production process (3) and its associated parameters (2) from at least two different sources using a data collection computer (10); performing a process mapping step using the acquired raw material data (21, 21a, 21b) related to the production process (3) using a process mapping computer (11); assigning the acquired raw material data (21, 21a, 21b) related to the associated parameters (2) of the production process to their corresponding program portions (3) by performing a data mapping step using a data mapping computer (12); analyzing the mapped process description (5) using specific software executed on an analysis computer (13), thereby identifying and verifying one or more existing characteristics (8) related to the quality or performance of the production process; and using the identified and verified characteristics (8) to develop the production process or improve its performance.

Description

[Technical Field]

[0001] This invention relates to a method and system for ensuring product quality in a process for producing a product from one or more materials.

[0002] This invention belongs to the technical field of process automation, quality control and supply chain optimization. [Previous Technology]

[0003] Modern production processes are highly complex, influenced by numerous variables related to raw material inputs, equipment and processing tools, and human interaction. Even the slightest variation can lead to significant quality problems in the finished product, resulting in deterioration or even unsaleability. This can lead to defects that are unacceptable to manufacturers and / or customers, especially in highly regulated industries such as chemicals or pharmaceuticals. Many methods are known in current state-of-the-art technologies to improve production processes and ensure the quality of manufactured products. Most of these methods are currently data-driven, relying on data collected from the production process to monitor and control relevant parameters.

[0004] One such approach is the six-standard-deviation method, which focuses in the control phase on using statistical procedure control (SPC) to identify assignable causes of variation in a procedure of interest. Assume the output of a procedure has two random variables: a population mean μ and a standard deviation σ. If we substitute samples from the procedure output and the sample size is sufficiently large, the distribution of the sample mean will be approximately normal, regardless of the distribution of the population. Furthermore, the arithmetic mean of the sample means provides a point estimate of the population mean, and the standard deviation of the sample means can be used to estimate the population standard deviation. These properties are known in statistics as the central limit theorem, and by virtue of this, we can calculate upper and lower bounds for control limits on the expected output from a given procedure. A measurement exceeding these control limits is a statistically significant event that requires further investigation to identify and address the root cause. In other words, the value deviates from the expected distribution to a degree that makes it highly unlikely to be generated by the same procedure as the previous data, and therefore there must be an assignable cause outside the procedure responsible for the deviation. The distribution of measured process variables can also be compared with an extension of known specification limits to quantify process capability. Given a specific process objective and associated error tolerances, we can directly calculate a capability index ratio that indicates the process's ability to produce acceptable quality output.

[0005] From US Patent 7,181,353 B2, a method for integrating six standard deviations into the inspection and acceptance process of one of the outsourced products is known, and the method includes the following steps: defining specification limits for product acceptance criteria; identifying primary standard products via a MES (Manufacturing Execution System) and SCADA (Supervisory Control and Data Acquisition) and reporting them to authorized personnel for disposal; preparing a report containing historical data, identifying the root cause and assigning a corrective action; separating and recording the secondary standard products in the MES; disposing of the secondary standard products; recording and documenting the corrective action in the MES; and outlining a method for recalling and eliminating a non-conforming incoming product. This method may provide an apparatus for a closed-loop corrective action (CLCA).

[0006] US Patent 6,675,135 B1 discloses a method used during a product development process, wherein the process includes a series of successive development phases, and the product includes at least two critical quality characteristics (CTQs). The method is used to generate a confidence matrix that can be used to enhance product quality through product design. A user initially provides product constraints and subsequently provides additional development information during each successive development phase. During at least two development phases and for each CTQ, the development information is used to determine a quality factor indicating the probability that the product will be within specified constraints. A confidence interval indicating the probability that the quality factor is accurate is identified. The quality factor, CTQs, and confidence factor are then configured such that the CTQs and factors are correlated.

[0007] These known methods are very closely tied to the six-standard-deviation approach, thus focusing on specific steps of the method to ensure the setting of quality standards and therefore being less suitable for use in alternative approaches that differ from this method. Furthermore, six-standard-deviation is a process improvement method that relies on the "voice of the customer" and prioritizes and filters based on expert input. It typically does not include modern data-driven approaches that utilize machine learning tools. Therefore, it has drawbacks when used with these automated methods, such as the need to repeatedly review human expert analyses. Some have also stated that the six-standard-deviation criterion is insufficient to truly guarantee defect-free products. Moreover, the proposed methods are not limited to predefined categories of items. In the definition phase of the six-standard-deviation process, in-category and out-of-category items are explicitly defined by experts involved in the project team. Here, a physical prioritization approach is used to narrow down potentially relevant areas to the voice of the customer. Out-of-category items may or may not be investigated in further projects. Therefore, even if a problem can be addressed within the project framework using the six-standard-deviation approach, there is no systematic way or method to consider out-of-category items. In the long run, this can hinder the discovery of the root cause of the problem. It is impossible to provide a sustainable solution to the impending deviation.

[0008] Another point is that in the case of different production locations with different production procedures and philosophies (e.g., such as a material provider and an actual producer), the six-standard deviation pathway leads to or preferably the need to align and / or merge those two different production procedures. This can be difficult and depends on the varying extent of the procedure resulting in many work.

[0009] Furthermore, one specific problem in this area is that undesired variations in finished product quality are driven primarily by feedstock input rather than by procedural variables only. Raw material quality is highly variable and suboptimal batch selection of product batches can result in batches produced that do not meet customer acceptance limits and require scrapping or rework.

[0010] Another specific point about this field is that the feature size of the circuit assembly has been rapidly reduced while the number of metal layers has been rapidly increased, thereby resulting in the device morphology exhibiting features that inhibit form-conserving deposition. The demand for global surface flattening of various thin film layers that make up integrated circuits (ICs) has increased substantially.

[0011] Chemical mechanical flattening is a procedure in which the surface of a substrate is smooth and flattened by the combined action of chemical and physical forces on the surface. CMP combines the advantages of both techniques while avoiding the disadvantages. However, pure grinding of the surface will result in excessive physical damage and pure chemical etching will not be able to achieve flattening, the combination of the two to produce a well flattened surface with minimal damage.

[0012] In addition to the fact that the rough morphology is reduced to a flattened state, chemical mechanical flattening (CMP) also has several advantages of semiconductor device fabrication. CMP allows device manufacturers to achieve global flattening of the entire wafer surface in a single step. The pathway can be used to flatten a wide range of materials from different metals to different oxide films. Chemical mechanical polishing is the flattening method that has been chosen by today's semiconductor industry.

[0013] A chemical mechanical flattening tool consists of a rotating table plate covered by a polishing pad. The wafer is mounted face down in a carrier pressed against the pad with a specified force. This force may be provided using either a defined and regulated gas pressure or a mechanical preamp pressure system. The wafer also rotates during the polishing procedure. The polishing pad is pumped to soak with a slurry on the pad. Polishing of the wafer surface occurs when the wafer rotates on its own axis and moves around the polishing pad while being forced against the pad. During the polishing procedure, the high points of the wafer surface naturally undergo more pressure and therefore more grinding forces. This is combined with the action of chemical etching agents to produce a one-enhanced removal rate of the material at the high point in the surface morphology with respect to the material at the low point. This produces a flattening effect in the procedure.

[0014] A chemical mechanical planarization (CMP) slurry typically includes: abrasive particles, which provide physical force on the surface; chemical additives, which promote removal rates; chemical additives, which reduce pitting; chemical additives, which reduce corrosion; oxidants, which chemically etch the material; activators or catalysts, which interact with the oxidants and promote the formation of free radicals; preservatives; stabilizers, which promote or facilitate the stability of the composition to prevent sedimentation, flocculation (including precipitation, polymerization or agglomeration of particles and the like) and decomposition; a surfactant, which helps protect the wafer surface during and after polishing to reduce defects in the wafer surface; chelating agents, which enhance the affinity of chelating ligands for metal cations and / or prevent the accumulation of metal ions on the pad (which leads to pad staining and unstable removal rates); pH adjusters; biodegradants, which control biological growth; and a solvent, such as water. CMP slurries utilize a large number of chemical materials.

[0015] In order to achieve efficient planarization at miniaturized device sizes, it is necessary to better understand and therefore better control the physical, chemical and complex interactions of the lubrication mechanical phenomena that occur at the interface between the pad and the wafer in the presence of fluid slurry.

[0016] Therefore, the production process of CMP slurry is a very complex matter, which depends on many parameters and variables regarding the materials, tools, production machines, and responsible personnel involved. Therefore, the process needs to be characterized, optimized, and modeled.

[0017] In order to improve production processes and ensure product quality, many methods are known among the most advanced technologies for monitoring the entire process. Nowadays, most of these are data-driven, in which relevant parameters and variables of the production process are periodically monitored and checked if target values ​​are not met.

[0018] Therefore, in order to comply with and enhance this state-of-the-art technology, it is desirable to find a new way to operate an automated production management system that further enhances the quality and reliability of the production process and can efficiently handle different processes.

[0019] Therefore, it is desirable to find a new way to operate an automated production management system, which further enhances the production process in terms of the quality of the resulting products that depend on the raw material input. [Summary of the Invention]

[0020] This task can be solved by a method for ensuring product quality in a process for producing a product from a material, the method comprising the following steps: acquiring raw material and finished product quality data for the production process and its related parameters from at least two different sources using a data collection computer; mapping the acquired raw material data to the production process in a process mapping step using a process mapping computer; analyzing the mapped process description using specific software executed on an analysis computer, thereby identifying and verifying one or more existing characteristics related to the quality of the produced product; and using these identified and verified characteristics to select the most suitable available raw material data to improve the quality of the resulting product. The core of this invention is to first track all potentially relevant data related to the production process to be created or improved, specifically, raw material data described by its specific parameters. This data is collected from at least two different sources, such as different production locations, internal quality control, and supplier analysis certificates (CoAs), and includes, in addition to raw material data, internal process data and customer data. One of the two different sources is usually the raw material supplier, as they possess the most knowledge about the material. Of course, different sources can also be located in the same production location and originate from, for example, two different production machines. Therefore, the data can differ in structure, format, syntax, etc. In the process mapping step, the available data related to the desired process is then used to establish and define the process structure, such as specific process steps and required process components. It also includes mapping data from the first source and other sources, thus handling two different data types. The focus here is not simply to target data from different sources to create a single process, but to keep them separate and map them together so that they can be subsequently processed and analyzed. Available raw material data may contain supplier material information, such as Certificates of Analysis (CoA, e.g., for metals), additional internal raw material (quality) measurements (e.g., trace metal impurity levels and purity levels), and batch profiles (which raw material batches and quantities) are compiled into a single data file. If process data is used, it can describe, for example, an existing production process considered to be improved by an inventive method. Alternatively, it may contain data describing a new production process to be created (e.g., available production machines and the like) to create the most efficient new production process. In the data mapping step, the acquired raw data related to its parameters is then assigned or mapped to its program counterpart from the program creation step, which is related to these parameters. The program and data mapping steps can be performed simultaneously, or the data mapping step can be performed after the program mapping step. After the program (including the assigned raw data) is successfully defined, an actual program evaluation is performed in the analysis step. During this evaluation, specialized software is fed into the mapped raw data and the final program data and analyzes their contents to search for specific patterns and dependencies that reveal the characteristics of a program, which can be used to improve the established program.This software can use different types of algorithms. It can employ, for example, artificial intelligence approaches such as supervised, unsupervised, semi-supervised, and reinforcement learning. Which is best suited depends on the available raw materials and the type of final process data. Importantly, the algorithm is trained to find patterns or identify the influencing factors of the process. This approach can be implemented using gradient-boosting decision trees, artificial neural networks (ANNs), or others. The ANN can then further improve its performance by learning from its input mapped raw materials. However, other AI software approaches are also possible. The software can also alternatively use approaches from classical statistics, provided they are suitable for determining the characteristics. Physical and mechanical models, such as mass balance, can also be used. Which approach is best suited and therefore chosen depends on the specific circumstances and the type of raw materials. After creating the mapped and analyzed process and identifying characteristics, these characteristics are applied to the production process, thus improving the production process and the quality of the resulting product. These characteristics can also be used to further improve product quality through insights gained from process experts. Specifically, based on the selection of a specific batch of raw materials with corresponding characteristics, the model will provide a prediction of the finished product quality. Different predictions of finished product quality are obtained by varying the batches of the same raw materials. By selecting the correct batch of raw materials, the most suitable final product quality for the customer can be chosen. The better the software can determine the characteristics, the less expertise from human specialists is required. All these methodological steps are executed by a computer configured to perform the methodological steps. The entire process can be automated using mathematical optimization techniques, where the user only provides finished product quality specification constraints and a weighted average of how deviations from these constraints should be penalized. The optimization algorithm then selects the most suitable batch of raw materials based on the user settings. While complete automation of the methodological steps is desirable (depending on the capabilities of the hardware and software used), human support (e.g., when evaluating acquired data or identifying characteristics) may be necessary. The minimum requirements for the computers involved include their ability to process, transmit, and display acquired raw material and final process data, as well as to perform software analysis steps. The computers themselves can be different computers in different locations, which are connected to each other via the Internet, local network, etc., or some or all of them can be the same.

[0021] Our approach is not limited to this. In the long term, learning from a use case is cascaded down to R&D for product development, and also cascaded back to procurement (suppliers). In light of this learning, experts are committed to controlling the purchase of the best materials from our suppliers, while researching the specifications of next-generation materials for new application technologies in the R&D of our customers.

[0022] The advantages of the present invention, and therefore preferred further development, arise from the associated sub-technical solutions and descriptions and associated diagrams.

[0023] One of the preferred further developments of the disclosed method includes, in order to obtain data for the production process and its related parameters, retrieving program data from a database connected to a data collection computer, creating data by observing the program using a data collection device (especially a sensor), and / or providing data by a human user. Which of these approaches, and in which combination, depends on the target production process. Generally, if at least some current data from sensors is involved, the program data has a better quality.

[0024] Another preferred further development of the disclosed method includes acquiring data by observing the production process during a previous execution of the process and / or during a current execution after the use of the identified and verified features. By doing so, it is ensured that the acquired data is always up-to-date. If AI methods (such as ANNs) use the latest information for training and application, their efficiency in using such methods is also significantly improved.

[0025] Another preferred further development of the disclosed method includes performing process mapping by describing the structure of production processes or their preceding stages from different sources (including necessary components, process sequences or process steps, ingredients (such as raw materials) and the like). By doing so, the process is defined and can subsequently be analyzed by software to create or improve its performance.

[0026] Another preferred further development of the disclosed method includes performing data mapping by assigning process parameters (such as temperature, mixing ratio of raw materials, time, and the like) obtained from all involved locations to their corresponding process components and process sequences or steps. When the process mapping defines the process structure, its necessary components, and the like, the data mapping assigns its parameters to the relevant process components defined in the process mapping. The process data thus prepared is then ready for analysis using software performance-related characteristics.

[0027] Another preferred further development of the disclosed method includes data analysis performed by software using supervised and unsupervised algorithms (including a data analysis framework with a data model that uses approaches such as multivariate analysis (e.g., PLS regression, PCA, random forest, XGBoost and artificial neural networks, PLS regression and / or random forest or similar)) or using supervised and / or unsupervised static algorithms. Both types of algorithms (supervised and AI-related or unrelated) can be used by the software. However, the more complex the procedure discussed, the more difficult it is to provide software with a non-learning approach that truly identifies all the characteristics of the desired procedure. These are more suitable for, but not limited to, less complex production procedures, or where only a specific defined part of the procedure needs to be evaluated.

[0028] Another preferred further development of the disclosed method includes analyzing data using mechanical models, physics-based models, (partial) differential equation-based models, and quantum chemical calculation-based models. The method is not limited to these model types, but rather to the most suitable model types.

[0029] Another preferred further development of the disclosed method includes the structure of the supervised algorithm being trained using the results of a program description derived from the program and data mapping, such as PLS regression, PCA, random forest, XGBoost, and artificial neural networks or similar algorithms. Artificial neural networks (ANNs) or similar algorithms are well-suited for evaluating complex production processes because they can be trained using mapped data from the program and data mapping steps and are thus adapted to production processes, no matter how complex they become. They can also be used in repeated iterations of the inventive method, and their adaptability improves as they are used more frequently to identify program characteristics.

[0030] Another preferred further development of the disclosed method includes obtaining program data manually by a user who inputs program data into the data collection computer by examining at least two different sources, or automatically by data collection software running on the data collection computer or a separate computer connected to it, wherein the data collection software transmits data to the data collection computer. Which approach is used depends on the limitations and capabilities of the hardware and software used. The more data that can be automatically obtained, the better.

[0031] Another preferred further development of the disclosed method includes using at least two different production locations as at least two different sources. These different locations may be, for example, one location for the raw materials used to produce the product (supervised by a material supplier), and the other location is the actual production plant.

[0032] Another preferred further development of the disclosed method includes characteristics that are root causes, such as maintenance problems, or previously unknown program problems related to the performance or quality of the program, such as specific setting parameters used in the production program. When the invention is used to improve an existing production program, specifically to solve one of its specific problems, the root cause is primarily relevant. However, the invention can also be used to create new production programs by identifying previously unknown connections between program parameters and / or components, thereby solving problems or unlocking previously unthinkable new potential.

[0033] Another preferred further development of the disclosed method includes production data from at least two involved production locations, including raw material data such as specific quality parameters or metal impurities and purity levels, P&ID charts, or in-process data such as sensor data including temperature, flow rate, tank level, etc. This data can typically be assigned to two different categories. One is raw material-related data, such as CoA, which is necessary for performing process mapping and thus defining the process. The other category includes process parameters, such as sensor data, which are fed into the data mapping steps.

[0034] Another preferred further development of the disclosed method includes implementing a user interface that preprocesses the acquired raw material data using a data platform on an analysis computer before applying the specific software, and writes the results to a database. A dedicated dashboard retrieves data from the database and provides it to the user for raw material viewing. The user interface is also loaded by the analysis computer. Alternatively, the dashboard is preferably organized using Tableau software; however, any other suitable software may also be used.

[0035] Another preferred further development of the disclosed method includes a user interface displaying the contribution of different raw materials to the prediction of a specific quality measure to indicate the most relevant raw material.

[0036] Another preferred further development of the disclosed method includes a semiconductor device manufacturing process that uses chemical mechanical planarization (CMP). While the inventive method can be used to improve every manufacturing process for creating a product from different raw materials, it is particularly suitable for semiconductor device manufacturing processes that preferably use CMP processes.

[0037] A further component of the present invention is a system for developing or improving a process for producing a product from a material, the system comprising: a data collection computer having a database connected to and / or at least two production sites, for acquiring program data for the production process and its related parameters from the at least two production sites; a program mapping computer for performing a program mapping step using the acquired program data related to the production process; a data mapping computer for assigning the acquired program data related to the relevant parameters of the production process to its corresponding program portion by performing a data mapping step; an analysis computer comprising specific software executed thereon using a supervised algorithm comprising a data analysis framework having a data model, wherein the software analyzes the mapped program data to identify and verify one or more existing characteristics related to the quality or performance of the production process; and a program execution computer for creating the production process and / or improving its performance at the at least two production sites by applying the identified and verified characteristics. The system performs the inventive method. As stated, the computers mentioned in the system can be built as standalone system components, or as a combination of the same computer or the like (whichever is most suitable). The analysis computer, preferably a standalone computer, should ideally have specific software. If raw material data from at least two different production locations is acquired automatically, the data collection computer needs to be connected to a computer-based automated control system at each production location. There, algorithms on the analysis computer write back to the respective system control system and the production personnel which batch IDs should be used for production. The type of computer used depends on the requirements of the method being performed. If most of the method steps are performed by a human user on a personal computer, tablet, mobile phone, or similar device with a display and some data input components or interfaces, the user can provide data to the computer and the software can be used. For more automated execution of the method, other types of computers, such as industrial PCs, microcontrollers, single-board computers, or embedded computers, can generally be used. A well-defined data interface and data transmission network (such as Ethernet, bus systems, or wireless alternatives) for automated data transmission becomes more important.

[0038] A preferred further development of the disclosed system includes at least one of at least two locations being a plant used for the production of chemicals, pharmaceuticals, or the like, and at least one of the other locations being a chemical material supplier and / or distributor. In this case, if an automated data transfer is required, both locations need to be connected to a data collection computer and / or their respective databases to provide the necessary procedural data. If secure, non-automatic data transfer is possible only due to different location owners, the connection is more indirect, for example, by transferring secure data storage and the like.

[0039] Another preferred further development of the disclosed system includes a data collection computer equipped with a computer-based digital platform for acquiring program data from at least two production locations. Another possibility for acquiring the necessary program data is to use a digital platform for data acquisition, to which all participating production locations can transmit their program-related data. The platform then manages this data and distributes it to their respective program and / or mapping computers to execute their mapping steps.

[0040] Another preferred further development of the disclosed system includes a program mapping computer and a data mapping computer supporting a human user's input terminal to perform program mapping and data mapping steps, while the analysis computer is a server loaded with software containing supervised and / or unsupervised algorithms (especially XGBoost, Random Forest, or Artificial Neural Networks), and the program execution computer is part of or the same as a computer-based control terminal at at least two production sites. As mentioned, if a human user is required to perform part of the method steps, the computer used must provide its own input and output components (such as a keyboard, mouse, screen, and the like) and its own software to handle this input. If an ANN is used by the software, suitable computer hardware is required for this ANN.

[0041] Accordingly, a further component of the present invention is an XGBoost, random forest, or artificial neural network or other AI approach, the structure of which depends on training with specific training data created by the following steps: acquiring program data of a production program and its associated parameters from at least two different sources via a data collection computer; performing a program mapping step via a program mapping computer using the acquired program data related to the production program; and assigning the acquired program data related to the associated parameters of the production program to its corresponding program parts by performing a data mapping step via a data mapping computer and creating training data from the mapped program data. The software is then trained using the training data thus created and its necessary internal structure is established, so that it can be used to analyze the mapped program data to identify desired program characteristics. By providing the software with real program data, it is further trained, and its analytical performance is improved.

[0042] Another component of the present invention is a computer program comprising instructions that cause the involved components to perform the following method steps: acquiring raw material data for the production process and its related parameters from at least two different sources using a data collection computer; performing a program mapping step using the acquired raw material data related to the production process using a program mapping computer; assigning the acquired raw material data related to the relevant parameters of the production process to its corresponding program portion by performing a data mapping step using a data mapping computer; analyzing the mapped program description program data using specific software running on an analysis computer, thereby identifying and verifying one or more existing features related to the quality or performance of the production process; and using the identified and verified features to develop the production process or improve its performance. The program portions responsible for individual method steps run on their respective computer portions. How the program itself is divided depends on the computer hardware involved. A main software may be used that runs on one of the mentioned computers or on a separate computer that controls the local client program. Other options include equivalent examples of software that communicates with each other.

[0043] The only requirement for this computer program to execute the entire described method is that the program used and its respective hardware components can execute the method completely and automatically. The program may then be stored on a computer-readable storage medium and / or data carrier signal, causing the involved components to perform the following method steps: acquiring program data for the production program and its related parameters from at least two different sources via a data collection computer; performing a program mapping step via a program mapping computer using the acquired program data related to the production program; assigning the acquired program data related to the relevant parameters of the production program to its corresponding program portion by performing a data mapping step via a data mapping computer; analyzing the mapped program data using specific software executed on an analysis computer, thereby identifying and verifying one or more existing characteristics related to the quality of the production program; and using these identified and verified characteristics via a program execution computer to create the production program and / or improve its performance. The storage medium can be stored on any suitable digital memory, such as a USB flash drive, a hard drive, a flash drive, etc. From this memory, the software can also be transmitted to its target hardware via remote communication components (such as Ethernet, wired or wireless) or any other suitable network transmission component that uses its own data carrier signal.

Implementation Method

[0050] The present invention will be described in more detail by presenting two preferred exemplary embodiments that disclose, based on data such as historical performance, customer factors, properties and material handling, raw materials and intermediate factors and properties, a proactive quality control system for materials.

[0051] Figure 1 illustrates an overview of one of the necessary method steps in two embodiments. In each exemplary embodiment, the steps themselves are performed differently depending on the conditions. Figure 2 shows a data analysis system 15 performing the method. As previously explained, its structure may also differ from embodiment to embodiment. In particular, the types of computers involved can vary greatly, depending on how many steps are performed by a human user with the assistance of a computer and application software or automatically by a specific computer using, for example, AI software.

[0052] Figure 3 illustrates a schematic overview of two example production locations and one type of process data they provide. In this case, it is a material supplier location 17 and a manufacturer location 16, wherein the material supplier location 17 provides data on raw materials 21, process-related data 1a, quality data 20a from the material supplier, other data 22a, etc. This data is collected by a data collection computer 10, which performs a data integration 19 and then provides the data to the data models 6, 7 used, such as a PLS model or an artificial neural network. The manufacturer location 16 provides data on its own production process 1b, its own quality data 20b, other data 22b, etc. This data is then transmitted via a secure data transmission connector 23. Data communication between the two locations 16, 17 is performed by website services 24a, 24b assigned to these locations. Alternatively, a third-party website service 25 may also be used. This can be centrally managed, for example, by a digital data platform.

[0053] First preferred exemplary embodiment:

[0054] This first preferred embodiment includes an active quality control system for one of the materials prior to the material being processed by a customer (in most cases, manufacturer location 16). It consists of the hardware described, as shown in Figure 3.

[0055] To ensure the quality of the resulting products, the system first needs to identify key parameters from the necessary materials. These parameters can affect customer-identifiable performance indicators, in addition to the material's Certificate of Analysis (CoA). Secondly, it needs to identify the raw material and intermediate parameters, as well as the CoA parameters, that can affect the key indicators in order to predict performance before producing a batch of materials.

[0056] To this end, the system performs the following steps.

[0057] Through a first data acquisition step, data is retrieved from multiple databases via data collection computer 10 using data extraction instruction codes. The multi-source data is cleaned, transformed, linked, coded, and normalized for customer sharing. This data includes time-series process data from the quality control system (such as sensor data, temperature, pressure, tank level, etc.), raw material data from supplier CoA (Certificate of Analysis, e.g., for metals), additional internal raw material measurements (such as trace metal impurity levels and purity levels), and batch profiles (meaning which raw material batches and quantities are compiled into a single data file).

[0058] Next, this data file is coded, preferably by assigning code names to the rows in the file and by normalizing the data between 0 and 1 for each row, except for contextual / discrete data such as date and batch number. Source data is also provided by material supplier location 17, which also contains coded and normalized contextual and time-series data. The key link between customer and material data is the material batch number.

[0059] In the next process mapping step, the manufacturing engineer and quality control personnel provide a process overview and description and input it into the process mapping computer 11.

[0060] During the subsequent data mapping step, all available data is examined by target experts, who assess the physical meaning behind the data using applied chemical / process engineering knowledge. In a workshop staffed with several process, quality, and data experts, the available data, presented as raw material parameters and / or process parameters, is mapped or linked to data from the process mapping step (such as in-line physical parameter measurements) using the data mapping computer 12. Customer data is collected from multiple wafer fabrication plants. Sometimes, using a single finished product batch in different wafer fabrication plants results in different customer indicator performance.

[0061] The most important methodological steps are related to data analysis. Here, the analysis computer 13 matches all collected and mapped data with corresponding material batch numbers based on the material batch number. Several suitable data models (such as PLS, random forest, and others) are used in different ways to predict indicators, highlight feature importance, and further develop the models used. The obtained feature importance is verified by a program expert to justify the discovered physical phenomena and their effectiveness on the data models used. After verification, the following steps are taken: 1. Create a list of feature importance for each model using SHAP or Pearson correlation values. 2. Develop a model to predict customer indicators based on effective material feature importance. 3. Develop a model to predict CoA parameters. 4. Use a randomized training and validation dataset to verify feature importance and model predictions. 5. Perform continuous bias or creep testing on the model to verify and improve model efficiency where possible.

[0062] The results can be presented to site owners and other customers. To this end, for example, suitable methods for controlling effective parameters with high feature importance are proposed. Tests have demonstrated that the proposed predictive model has proven efficiency on actual data and provides good overlap with one of the customer indicator areas of performance. Before being put into production, material batches should be internally reviewed and validated in terms of planning, manufacturing, and quality. The model should also be continuously reviewed against actual data for model-driven efficiency.

[0063] Second preferred exemplary embodiment:

[0064] This second preferred embodiment also includes an active quality control system for one of the materials. It performs the following procedural steps.

[0065] First, a batch automation is established. This means that the production of a finished product is automated, for example, by measuring the liquid level in the tank and controlling the amount of raw materials added, and by introducing a data history.

[0066] The next step is the data acquisition step. Here, data is acquired, corrected, and integrated by an internal IT and process quality expert team into an Azure Sequel Abstraction Layer (SAL) database running on the data collection computer 10. This data pertains to: • Finished product batches and batch selections from a select list SQL database, which is used to customize a web-based tool for planning batch selections for upcoming batches, thereby exchanging data with an SAP system. • Quality measurements of raw material batches and finished product batches (such as trace metals, conductivity, average particle size) and customer acceptance limits from a LIMS system Labware. • Process data from historical data, such as sensor data, such as UPW temperature, differential pressure during filtering, including insights into effects beyond raw material quality.

[0067] The process mapping step consists of providing a process overview and drafting instructions for each product under consideration. Additionally, process and quality experts use process diagrams and P&ID to identify all potential influencing factors. These quality experts are preferably supported by a process mapping computer 11.

[0068] In the data mapping step for each finished product under consideration, with the assistance of the data mapping computer 12, available data is mapped to relevant procedural steps drafted by the program / quality expert during the program mapping step. This data includes, for example, which sensors are related to which product from historical data. Moreover, a possible and better data cleaning and aggregation is also performed in this step. Data cleaning may encounter some problems, such as which batches should be excluded due to incomplete or inconsistent data, since they are still in transit. If several batches of raw materials are mixed in one batch, or if the average value of quality measurements is used in the case of several samples, then data aggregation considers a point weighting of the sum of one of the raw material quality characteristics.

[0069] Next, the core of the preferred embodiment is data analysis. Here, all data is matched to the finished product batch level by averaging values ​​over the manufacturing period of the batch, using batch numbers or, in the case of additional process data. Several data models are tested to predict various finished product quality parameters (typically about 15 parameters in total), ranging from quality balance models based on raw material batch characteristics and weights to data-based linear and nonlinear models (such as OLS, PLS, xgboost, and others). The model performance is evaluated on an invisible test set based on a time series segment of the data, attributable to the autocorrelation properties caused by the overlap in the raw material batches used. Preferably, the data models are loaded onto an analysis computer 13, which can be any suitable computer.

[0070] Then, important influencing factors were discussed with and verified by procedural and quality experts. Simplification of the final data model to use features available prior to production. The visual need to include alternative data models for potential procedural-related impacts, e.g., predicts UPW temperatures using currently unused lags and / or weather forecasting based on external temperatures.

[0071] In order to perform the inventive method, it is very preferable to use a suitable user interface. For this purpose, the final data preprocessing and application of one of the deployed predictive data models was performed using the Alteryx platform. The results were written to a SQL database on Azure, from which a dedicated Tableau dashboard extracted data for the Raw Material Review Board (RMRB). For each product and quality quality measure considered, the dashboard displays available historical quantities and forecasts, as well as customer acceptance limits for past and planned batches, and furthermore, the contribution of different feedstocks to the prediction of a given quality measure is displayed to indicate the most relevant feedstock 21c . Then, a summary is provided in the dashboard to highlight the batches that need attention due to unmet customer acceptance constraints.

[0072] Both supply chain planners and quality experts can review the dashboard to identify potential problems with planning lots. Based on quality forecasting, supply chain planners adjust batch selection as needed. Attributed to every minute of data renewal, any changes are reflected in the dashboard. During weekly meetings, supply chain planners and quality experts discuss any issues or special current circumstances (such as logistics constraints, requests from customers requesting a particular batch selection) to determine a suitable move.

[0073] With this pathway implemented as an instance in the first and second embodiments, batch automation leads to a significant reduction in finished product quality variation. This allows for transparency in the expected finished product quality based on the raw material lot selection 21c and the prediction model, thereby leading to a further refinement of one of the lot selection procedures and a reduction in the need for rework or scrapping batches.

[0074] In addition, continuous data integration allows monitoring of forecast quality and based on constantly incoming new information improves the data model used. In a further better embodiment, the system can also be extended to automate batch selection in non-edge scenarios, thus reducing the manual work of the supply chain planner. [Brief explanation of the diagram]

[0044] A better illustrative embodiment is used below to describe in more detail the system, methods and software products according to the invention and the functionally favorable developments thereof. In the schema, components corresponding to each other are provided with the same component symbol.

[0045] Graphic display:

[0046] Figure 1 is a general overview of the necessary method steps.

[0047] Figure 2 is a schematic overview of one of the system components involved.

[0048] Figure 3 is a schematic overview of the two production sites and one of the types of procedural data they provide.

[0049] Figure 4 is a flowchart of one of the procedure steps in a specific embodiment.

Claims

1. A method for ensuring product quality in a process for producing a product from a material, wherein the production process is a semiconductor device manufacturing process using chemical mechanical planarization, comprising the following steps: The process involves: acquiring raw material data (21, 21a, 21b) for the production process (3) and its associated parameters (2) from at least two different sources using a data collection computer (10); performing a program mapping step using the acquired raw material data (21, 21a, 21b) related to the production process (3) using a program mapping computer (11); assigning the acquired raw material data (21, 21a, 21b) related to its individual parameters (2) of the production process to its corresponding program parts (3) to create a mapped program description (5) by performing a data mapping step using a data mapping computer (12); and analyzing the mapped program description (5) using specific software executed on an analysis computer (13), the software including methods such as multivariate analysis, PLS regression, PCA, and random forest. The data analysis framework of one of the data models (6, 7) of Forest, XGBoost and artificial neural networks or similar approaches is used to identify and verify one or more existing characteristics (8) related to the quality of the produced product; and the identified and verified characteristics (8) are used to select the most suitable available raw material data (21c) to improve the quality of the resulting product.

2. The method of claim 1, wherein in order to obtain the data (21, 21a, 21b) for the production process (3) and its related parameters (2), the raw material data (21, 21a, 21b) is retrieved from a database (24a, 24b, 25) connected to the data collection computer (10), created by observing the process using a data collection device, especially a sensor, and / or provided by a human user.

3. The method of request 2, wherein the acquisition of the data (21, 21a, 21b) is performed during a previous execution of the program and / or during a current execution following the use of the identified and verified features (8).

4. The method of claim 1, wherein the process mapping is performed by describing the structure of the production process or its preceding stages, including necessary components, process sequences or process steps, ingredients, especially raw materials (21) and the like.

5. The method of claim 1, wherein the data mapping is performed by assigning the acquired raw material data (21, 21a, 21b) such as temperature, mixing ratio of the raw material, time and the like to their corresponding program components and program sequences or steps.

6. The method of claim 1, wherein the analysis of the data is performed using mechanical models, physics-based models, (partial) differential equation-based models, and quantum chemical calculation-based models.

7. The method of claim 6, wherein the structure of the supervised algorithm is trained using the results from the program description of the program and data mapping, such as PLS regression, PCA, random forest, XGBoost, and artificial neural network or similar.

8. The method of claim 1, wherein the data (21, 21a, 21b) is acquired manually by one of the users who input the raw data (21, 21a, 21b) into the data collection computer (10) by means of verifying the at least two different sources (16, 17), or automatically by one of the data collection software running on the data collection computer (10) or a separate computer connected to it, wherein the data collection software transmits the data to the data collection computer (10).

9. The method of claim 8, wherein at least two different production sites (16, 17) are used as at least two different sources (16, 17).

10. The method of claim 9, wherein the raw material data (21, 21a, 21b) from the at least two production sites involved (16, 17) includes specific quality parameters or metal impurities and purity levels, in-process data such as temperature, pressure, flow rate and / or P&ID chart (4).

11. The method of claim 1, wherein a user interface is implemented that preprocesses the acquired raw material data (21, 21a, 21b) using a data platform on the analysis computer (13) before the application executes the particular software, and writes the results to a database, and a dedicated dashboard retrieves the data from the database to provide it to the user for performing a raw material inspection.

12. The method of claim 11, wherein the user interface displays the contribution of different raw materials to the prediction of a specific quality measure to indicate the most relevant raw material.

13. A system (15) for developing or improving a process for producing a product from a material, comprising: A data collection computer (10) having a connected database (24a, 24b, 25) and / or at least two production locations (16, 17) for acquiring raw material data (21, 21a, 21b) for the production process (3) and its related parameters (2) from the at least two production locations (16, 17); a program mapping computer (11) for performing a program mapping step using the acquired raw material data (21, 21a, 21b) related to the production process; and a data mapping computer (12) for assigning the acquired raw material data (21, 21a, 21b) related to the relevant parameters (2) of the production process to its corresponding program part (3) by performing a data mapping step. An analysis computer (13) includes specific software executed thereon using a supervised algorithm having a data analysis framework with a data model (6, 7), wherein the software analyzes a program description (5) mapped thereon to identify and verify one or more existing characteristics (8) related to the quality of the production process; and a program execution computer (14) for creating the production process and / or improving its performance at at least two production sites (16, 17) by applying the identified and verified characteristics (8).

14. The system of claim 13, wherein at least one of the at least two locations (16, 17) is a plant (16) for the production of chemicals, pharmaceuticals or similar substances, and at least one of the other locations (17) is a chemical material supplier and / or distributor.

15. The system of request item 13, wherein the data collection computer (10) is equipped with a computer-based digital platform (19) for obtaining the raw material data (21, 21a, 21b) from the at least two production sites (16, 17).

16. The system of claim 15, wherein the program mapping computer (11) and the data mapping computer (12) support human user input terminals to perform the program mapping and data mapping steps, and the analysis computer (13) is a server that loads the software having the supervised and / or unsupervised algorithms, in particular an artificial neural network, and the program execution computer is part of or the same as the computer-based control terminals of the at least two production sites (16, 17).

17. A method for training a data model structure, the data model structure including an XGBoost, random forest, or artificial neural network (7), the structure of which depends on training using specific training data (5) created by the following steps: acquiring raw data (21, 21a, 21b) for the production process (3) and its associated parameters (2) from at least two different sources using a data collection computer (10); performing a program mapping step using the acquired raw data (21, 21a, 21b) in relation to the production process (3) using a program mapping computer (11); and assigning the acquired raw data (21, 21a, 21b) in relation to the associated parameters (2) of the production process to its corresponding program portion (3) by performing a data mapping step using a data mapping computer (12) and creating training data from the mapped program description (5).

18. A computer program comprising instructions that cause the computer concerned to perform any one of the methods described in claims 1 to 12.

19. A computer-readable storage medium having stored thereon a computer program as claimed in claim 18, the computer program causing the computer concerned to perform any one of claims 1 to 12.

Images