Case-Based Axiomatic Design Assistant (CADA): Combining Axiomatic Design and Case-Based Reasoning to Create a Design Knowledge Graph for Pharmaceutical Engineering

Abstract

The development of personalized drugs introduces new uncertainties and risks in production machinery design, which can be mitigated through structured workflows. As the commonly used V-Model approach has limitations in dealing with complex multi-domain problems, it is essential to address traceability and relationships between requirements and solutions in a regulated environment to ensure product quality. This study focuses on the conceptual design phase and develops a design methodology called the Case-based Axiomatic Design Assistant (CADA) to address this type of problem. It takes, as a starting point, Axiomatic Design (AD), due to its simplicity and graphical tools for quality evaluation, and Case-Based Reasoning (CBR), due to its capacity to integrate data structures and continuously improve. This combination is put into practice through a visual assistant that utilizes a knowledge graph to represent design elements comprehensively. This article describes the development, implementation, and testing process of CADA, which includes examples of the conceptual design for pharmaceutical manufacturing. The proposed CADA method facilitates systematic requirements analysis, structured reasoning, and solution evaluation, and overcomes the limitations of previous methodologies. It represents a novel approach with an intuitive workflow and advanced graphical capabilities, exemplified in the context of a conceptual design for pharmaceutical manufacturing. The inclusion of intrinsic data labeling capabilities and inference visualization enhances its relevance.

1. Introduction

The early Product Design phase is considered one of the cornerstones of successful production equipment design. Over time, various design methodologies were created due to how designers gain the specific knowledge they need to understand user requirements and make decisions for the particular use case [1,2].

From a philosophical standpoint, knowledge and design are inherently related in the form of ontology, the science of things that exist, and epistemology, which is the methodology used to achieve and correctly understand knowledge [3]. Gruber [4] found an early definition for Ontology in the context of computer science as an “explicit specification of a conceptualization”. This term is explained as “what exists is exactly that which can be represented”. Later, Studer et al. [5] expanded this definition with a domain-specific component: “an ontology is a formal, explicit specification of a shared conceptualization”. The term “conceptualization” hereby stands for “an abstract, simplified view of the world that we wish to represent for some purpose”, as a general definition for knowledge representation for a specific use case [6]. Following the ideas of these statements, a structured design process starts with something existing (Ontology). It moves into a formalized information flow of reasoning or a cognitive process (Epistemology), resulting in a new situation that becomes Ontology [3].

From a practical perspective, a design process starts with the requirements analysis and resulting requirement list, which act as design guidelines throughout the project. Designers have to distinguish between technical (functional) and non-technical (economical) attributes while working on conceptual designs [7]. Efficiently reusing knowledge to structure and organize information is a relevant topic. Designers with rich backgrounds in the specific area may handle this intrinsically, but faster-changing products and shorter innovation cycles lead to a different situation [8]. Increasingly, design methods for customized products have become necessary due to the impact of Industry 4.0 on many branches. According to Li et al. [9], these One-of-a-Kind products come with three main obstacles to product design. First, the product development timespan decreases; second, order sizes decrease; and third, product variety in production increases. To overcome these challenges, companies must find a way to reuse historical product knowledge and manage their resources very efficiently. Some researchers have focused on identifying requirements more straightforwardly and precisely through direct communication and information sharing using web technology [9]. The connection between modeling tools and design methods is seen as an enabler for high-quality products that are no longer based on the experience of designers and managers alone, and it opens possibilities like automated design [10]. Scholars have sought to structure design processes and incorporate quality assurance mechanisms for years. One way of doing this is to use Lean principles to bring product knowledge back from production to design, enhancing the manufacturability and testability and thus increasing the product quality [11].

Developing and designing new pharmaceutical devices is a challenging task requiring knowledge from various disciplines. As regulatory impact is a quality assurance method, there are helpful guidelines and tools to achieve compliance. The most recognized technique to support the high-quality development of custom applications is the V-model described in the GAMP 5 guidelines [12]. The core of the V-Model is a strictly regulated process, in which the requirements for a system must first be fully defined in the specification before work can begin in the subsequent phases [13]. According to this model, each specification must be accompanied by a test strategy to ensure that the implementation meets the requirements. The approach in GAMP 5 complements a well-known conceptual design strategy with a risk-based approach. This model has limitations because it is very general and, therefore, cannot support all applications in detail, but its applicability in software development projects is well-researched [14]. The belief that the pure V-Model accurately reflects a development process could lead to a misclassification of project complexity, so enhancing the original model may be a measure to mitigate this risk. Due to changing customer requirements, which often only arise in the implementation phase, this model had become increasingly impractical and led to debates between suppliers and representatives of the regulated industry when the latter insisted on compliance with the rules of the V-Model [13].

The primarily used V-Model approach does not provide a structure for consistent learning and information feedback. Decisions are not recorded and data are not merged into a foundation model to assist future design cases. Different tools for subparts of the design process store information in silos that are not available for joint decisions. Recent pharmaceutical regulations such as Quality by Design (QbD) or Design Space (DS) require new design methodologies with in-built data-driven design and decision-making capabilities. These risk-based approaches depend on knowledge of the product and production process to objectively assess and characterize the relevant parameters.

Special design rules and frameworks must be covered in the regulated pharmaceutical environment, which can be challenging without specific guidance. The main goal of this research is to develop a structured process incorporating existing knowledge and quality mechanisms for the early conceptual design phase. A method needs to be identified to form a valid knowledge base through information relations. Precise knowledge engineering, along with problem identification, is one of the success factors for the later use of knowledge within a model-based method [15]. Visual ways of knowledge representation shall complement the design method to keep the dataset human-readable for evaluation and validation purposes. This feedback mechanism shall also enhance the data quality for further use in training AI models. A streamlined design process from initial requirements to conceptual solutions is in the scope of this new method to meet domain-specific regulations.

This article concentrates on the theoretical development and case-based testing of a new method that can assist a designer in conceptual design tasks. The paper is organized according to the process we followed to develop a new method that has the potential to overcome the limitations of the previously mentioned V-Model approach as the mostly used one. First, the relevant literature was identified through an exploratory literature review to find out the state of the art in terms of existing methodologies. We then developed our method from these theoretical frameworks and implemented it in a structured workflow. Finally, we applied this method to a set of design problems to evaluate its applicability and usability.

2. Methods and Theories for CADA Design

The following summary of existing methods in the form of literature research is grounded in the literature published in the leading scientific databases, e.g., Web of Science (Core Collection) and Scopus.

2.1. Recognized Methods for Pharmaceutical Equipment Design

Quality by Design (QbD) is known as a systematic prospective analysis of product and process characteristics during the design phase [16]. The main goals of QbD are specifications and control parameters that enable processes to achieve predefined quality characteristics [17]. The ICH guideline Q8 (R2) on pharmaceutical development [18] instructs on using a quality-risk-based QbD approach to develop drug substances and manufacturing processes. Quality Target Product Profile (QTPP), as described in the Q8 (R2) guideline, forms the basis for design and development in the form of a set of measurable product properties most relevant to product quality. Critical Quality Attributes (CQA) can be derived based on this dataset. CQA is applied to the equipment and its internal mechanisms that are in use for the production to identify Critical Process Parameters (CPP). These physical, chemical, biological, or microbiological characteristics determine the product’s ability to meet the QTPP. They must be in a dedicated range during the production process or in the product itself.

QbD is important to remember during the method’s development as it is a fundamental description of how pharma structures data-driven development of new equipment.

Design Space (DS) is a multidimensional combination of input variables and process parameters relevant to product quality [18]. The identified attributes and their fixed ranges or a combination stated as design space must be captured during production. Process Analytical Technology (PAT) is described by the FDA [19] as a system that supports the analysis and control of a pharmaceutical manufacturing process through timely measurements. It gathers critical quality and performance information from process and material sensors. This mechanism allows better quality control and process knowledge to better coordinate individual production equipment in an integrated process [20].

DS, the principle of data-based connection between physical hardware and process parameters, is a cornerstone of the design for future-proof pharma production equipment.

2.2. Existing Methods for Structured Problem Solving in Design

The technique of order preference through similarity to an ideal solution (TOPSIS) provides an efficient evaluation process, allowing decision makers to compare and rank concepts effectively. Applying TOPSIS to support decision making along a design process can pose feasibility challenges. Subjectivity in the evaluation process may lead to biased or unreasonable outcomes [21,22].

Analytic Hierarchy Process (AHP) is an iterative decision-making process for selecting the best approach among alternatives. The methodology needs to incorporate confidence levels and uncertainties effectively in the evaluation process, potentially impacting the accuracy of the final decisions. The iterative nature of AHP can also lead to complexities, especially when dealing with many criteria or alternatives, making practical applications challenging [21,23].

Iterative solution finding as well as recognized procedure in a complex solution finding process is relevant for our design method development.

Theory of Inventive Problem Solving (TRIZ) offers a structured problem-solving approach that can enhance innovation in various industries. Its limitations due to complexity, lack of adaptability, and trend towards oversimplification need to be carefully considered when applying it [24,25,26]. TRIZ has also been applied in problem-solving activities related to Systematic Conceptual Design (SCD) and a graphical means of information representation called a Problem–Solution Network (PSN) [27].

Further development needs to take into account that graphical information mapping has a positive impact on the usability of a design method.

Axiomatic Design (AD) is well-known for its applicability and effectiveness in systematically analyzing a system and its requirements [24,28,29]. According to Suh [30], the inventor of AD, the workflow includes direct linkage of Customer Requirements (CR) to design objectives (Functional Requirements, FR) and design approaches (Design Parameters, DP) to build a problem-to-solution relationship. The fourth and last domain is Process Variables (PV). The framework has no built-in capability to keep the rationale behind the solution or other meta information that helps to understand the decision-making process inside the design matrix [31].

AD, as an easy-to-understand design workflow that is proven in a variety of domains, can form the core principle of our development.

2.3. Existing Methods for Equipment Design

Equipment design is a crucial aspect of engineering, focusing on the development of new machines or manufacturing systems. It relies on a thorough understanding of requirements and converting them into practical designs. The Manufacturing System Design Decomposition (MSDD) approach provides guidance grounded in the Manufacturing System Design (MSD) [32] methodology (based on AD theory, also described as Production System Design, PSD, in [33]) to break down requirements and designs according to specific criteria derived from LEAN principles, such as quality, output, and operational costs [34]. The aforementioned approaches are based on the AD theory outlined in Section 2.2, demonstrating that AD is a valid framework also for our method.

Integrated Product Development (IPD) (also known as concurrent engineering or simultaneous engineering) strives for parallelization of engineering tasks in multidisciplinary tasks. According to IPD, the core building blocks of every design are objects, procedure, activities, information, and methods, which interact along a specific reference model including rules on how these components interact. Alongside this model, different partial models can be created by different participants of the engineering team [35,36]. IPD clearly demonstrates how designs and their elements can be modeled and structured formally. This approach helps identify cross-relationships and dependencies, which are crucial for compliant engineering and documentation for pharmaceutical regulations.

From a product-driven perspective, the 3-Cycle Model emphasizes the importance of designing conceptual equipment during the early stages of product development. The design of the product and the production equipment can mutually influence each other, as both development cycles work together to create a virtual design that can be easily adapted [37]. This method addresses various needs, such as interdisciplinary design teams and iterative problem-solving processes. However, in pharmaceutical engineering projects, the product specifications are rigid and must not be influenced by the manufacturing equipment design. As a result, this method cannot fully utilize its potential.

The double diamond model is an example of a creative, human-centered design process. It consists of four phases: discover, define, develop, and deliver. This method encourages creativity by frequently shifting the designer’s perspective throughout the design process, alternating between divergent and convergent thinking [38,39]. We categorize this method within our research primarily as an ideation and creativity technique, it can be used alongside a modeling approach to generate ideas through an innovation process [40].

A recent approach in design science, Model-Based Systems Engineering (MBSE), has emerged in regulated industries as a holistic, data-driven model that spans from conceptual design to the end of a product’s life. MBSE promotes information exchange among disciplines and stakeholders through digital models rather than traditional documents. All system components are linked to a central model, which serves as the source for all connected services that utilize the data model. In addition to facilitating data exchange, MBSE encompasses virtual prototyping, software implementation, and simulation to validate designs [41,42]. Centralized and model-driven design addresses one of our core topics: the creation of reusable knowledge through a design workflow. Model-Based Systems Engineering (MBSE) offers a comprehensive toolkit that meets the needs of regulated industries regarding transparency and change management. However, MBSE is a framework that encompasses more principles than those necessary for our research focus. Conversely, our method should be conceptualized to fit within an MBSE environment, supporting the idea of centralized data and data-based dependencies between design elements.

2.4. Theory of Knowledge Representation

Semantic Network (SN) was first introduced in 2012 by Google. The concept of a knowledge graph improves the capability of its search engine by using the relationship between entities. From this initial event, knowledge graphs gained wide attention in research and Industry as an effective way to manage heterogeneous data [43,44]. To formulate a Knowledge Graph from heterogeneous information, data must be processed toward entities (cluster, conceptual node, or subtopic) and connections (relationship) [43]. This definition also applies to the semantic network, defined as a set of nodes representing a cluster of information connected to other objects in one graph [45].

Case-Based Reasoning (CBR) is a problem-solving method in design that reuses existing knowledge as input data [9]. Comparing previous or existing solutions to current problems is made possible when the original solution is documented in a way that keeps all relevant decisions made along the design process [9]. Continuous improvement is also a strength of CBR. By learning from the outcomes of its past recommendations (via feedback), a CBR system can continuously improve its problem-solving capabilities over time, thus enhancing its effectiveness and efficiency [46]. The number of research articles discussing CBR in AI has been steadily increasing since the 1990s. Currently, the top three applications are “planning or design”, “decision”, and “prediction” [47].

SN and CBR offer a wide range of possibilities to solve the knowledge representation within our given research task.

2.5. Existing Methods in the Use of AI for Design

Graphical representation seems to be a topic in recent design research, as there are investigations into using semantic networks to structure ideas and concepts [48].

Guebitz et al. [49] introduced a risk-based (FMEA) Ontology concept to fulfil the GAMP5 V-Model approach in a formalized way.

Some researchers investigated a combination of various methodologies in addition to the pure use of one individual methodology. Research has explored the synergies between Axiomatic Design and other tools like TRIZ in conceptual design processes, leveraging the strengths of each method to enhance problem-solving capabilities and foster innovation [50].

Modern design tools need to consider their relevance for AI-supported design. Ontology-based knowledge models to support domain-specific design extract relevant information from concepts and provide feedback through query-based interactions [51]. AI-based methods in conceptual design can facilitate design ideation or concept generation [52].

Under the term Explainable Artificial Intelligence (XAI), various research studies have identified possibilities for creating transparency during training and usage of AI models [53]. A human person is not necessarily able to understand a model’s behavior in terms of linkage between input data and output in the form of a prediction. Methods to analyze the model behavior are challenging if not intrinsically supported by the data used for training [54].

The designer plays a central role in the development of an AI-based design model. Structured and properly labeled data are crucial for training AI models that support ideation and decision-making steps. A Human-in-the-Loop approach offers a wide range of tools to generate high-quality outcomes [55].

Visualizing an AI assistant’s internal decision-making process is very important to keep humans in the loop and ensure the quality of the outcome [56]. Knowledge graph-based enhancement of AI models can help to improve models based on augmented data [44].

The structure and usability of the method need to address the exceptional environment of the regulated pharmaceutical branch. Clear and easy-to-follow decision making and content visualization are critical success factors for this tool. The visual representation of the design and its internal information structure must be capable of labeling data and showing the output of a trained model in the same way.

3. Results: Method Development, Implementation, and Testing of CADA Framework

Following the literature review, the research develops, implements, and tests a data-driven, risk-based, and structured design method that meets the needs of conceptual design for pharmaceutical equipment. For this purpose, we combined existing and function-proven design frameworks and data architecture to form a novel conceptual design method. The results of this study will then be described around the process of developing the CADA framework.

3.1. Method Development

The following Figure 1 shows the proposed method as a selection of different tools identified within the literature search. The combination and relevance of the individual building blocks are described in detail from the bottom to the top layer afterwards.

3.1.1. Semantic Network (SN)

As knowledge, or in a more general form, “data”, creates the backbone of every decision, it is important to us to have an accessible and interpretable information layer within our proposed method. Former design knowledge systems store information in fragments or tables, which may be hard for non-advanced designers to interpret without the context of the specific situation. Designers work best when they have a concrete design situation on which they can refer and base their decisions [8]. SNs implement entities and relations in an easy-to-follow way (Figure 2), which allows humans to read and follow the paths of information and connections between them within a dataset [43,44].

Knowledge stored in this way can help find similarities within design projects. If the format is clear and structured, the representation of connected information in a semantic network is easy to understand [57].

3.1.2. Case-Based Reasoning (CBR)

Reusing existing knowledge and reasoning is an efficient way to enhance the design quality. A structured and context-rich knowledge base is essential [58,59]. Our selected data layer, SN, covers this requirement. The core component of CBR is called case memory (CM), which is a database that stores all structured information and problem-solving relations. How the CM organizes its internal data can differ from domain to domain. Aamodt and Plaza [58], early adopters of CBR, describe the Generalized Episode (GE) method. It is proven to apply to generalized and case-specific knowledge as it works in a simplified way of human reasoning and learning.

Every GE has three components, namely norms, indices, and cases. Norms hold general information as the root node of an information cluster to characterize the following content. Indices build sub-nodes of a Norm in a more specific description (Figure 3). An index consists of two parts, the index name and index value, which can also be considered a search term and response (e.g., hair color = brown). These key–value pairs act as a condition to distinguish between different Cases. The Case Node keeps the information needed to form the problem solution statements. If different indices point to the same case, a new Layer is introduced as a new Norm holding its indices and cases [46,58,59].

The goal of creating a database that can support future problems is widely known in problem-solving processes in production. An operator is asked to keep cause–effect descriptions as detailed as possible, with remarks on what best describes the details of the individual combination (Figure 4). This dataset can then be fed to train a Bayesian network that can predict actions for similar problems [60]. The same method could be applied to design problems, where problem-versus-solution statements train a model.

3.1.3. Axiomatic Design (AD)

AD supports iterative thinking and separation between requirements and physical solutions. Systematic Design is made possible by enforcing the process of clearly specifying FR from the weaker formulation of pure user requirements. Engineers tend to build solutions directly from user requirements and thus miss out on the step of working on the functional requirements and their interdependencies [61].

In contrast to other methodologies, AD provides an easy-to-use, streamlined design workflow that can be used independently of the number of available concepts. It also comes with a graphical evaluation and decision-making tool. These properties led us to implement AD techniques into our proposed tool.

The theoretical basis of AD to create a good design are the following Axioms:

• Independence Axiom: maintain the independence of the functional requirements.
• Information Axiom: minimize the information content.

The design process is carried out iteratively through the “zigzagging” process between the FR and DP. This method creates a table (Design Matrix, see Figure 5) that offers the possibility of evaluating the design concept. In the best case, the Design is “uncoupled”, meaning there are only one-on-one relationships between FR and DP, which results in a diagonal matrix. Also acceptable is the “decoupled” scenario where the design matrix is triangular. Design concepts that are “coupled” indicate non-acceptable solutions. This can result from situations where not all FR are connected to at least one DP or a violation of the Independence Axiom [28,30].

If viable solutions exist, AD supports the decision-making process through the Information Axiom, which states that the best solution is the one that needs the least information to be described. Following this concept, designers strive for simplicity. This principle has the potential to be cost-effective and quality-relevant at the same time [28]. The definition of AD elements (CR, FR, DP, and PV) is very weak to keep it open for various design problems. In other words, a DP can be a system, equipment, function, or variable. This comes with the need for a clear definition available to all designers involved in the solution-finding process. A clear abstraction guideline that keeps the layer definition (what is the definition of a System, Equipment, …) is thus interchangeable information throughout different design problems and stakeholders [31]. Such a definition is crucial to support the reuse of knowledge in the context of Ontology and Epistemology. To incorporate previously designed solutions for similar problems into a new design process is an overall performance improvement of the design process. The same reasons apply to the support of data-driven design [43].

Our proposed method takes care of this, as the definition of the data objects in the underlying dataset is described in cases following the GE method of CBR.

3.2. Method Implementation

The developed method aims to streamline the design process by incorporating AD principles. The following implementation as a workflow follows a hybrid agile-waterfall approach inspired by the manifesto for agile software development. This approach allows for flexibility and structure, as it combines the iterative nature of agile methodologies with the systematic planning of the waterfall method. The workflow begins with thoroughly understanding the project requirements and objectives, using the AD elements to design the procedure. The visual representation fosters clarity and transparency, allowing for the easier identification and resolution of potential design issues.

A strong focus on existing knowledge and reusable data ensures that the design meets the functional requirements and creates a positive and intuitive user experience. Stakeholders and end-users are actively involved throughout the design process, providing valuable insights and feedback.

Our design method is mainly supported by representing knowledge (CR, FR, DP, and PV) as a node connected to other nodes in the following design stage according to AD. The nodes are connected by arrows that indicate a relationship (coupling) between them. If there are only one-on-one connections, the design is classified as uncoupled (Figure 6a), which is the best possible scenario. If arrows share the same target, the design is decoupled (Figure 6b), which is then fed back to the user by a color code. The design is coupled if connections share a source and target and thus intersect (Figure 6c). This scenario has to be avoided to satisfy the independence axiom. The design assistant can instantly give feedback on such an axiom violation by message or color code within a graphical support tool.

From there, the method emphasizes iterative design cycles, with regular feedback and iteration loops to ensure that the design meets the project’s evolving needs. The method incorporates AD principles, focusing on achieving functional requirements while minimizing system complexity.

The assistant can guide the user through an iterative process to split nodes into sub-nodes with more specific knowledge. After splitting requirements (Figure 7b), a more in-depth analysis of the possible solution can be carried out to meet the iteratively detailed requirements (Figure 7c).

This process can be carried out with the designer and client to gather knowledge in detail. This collaborative approach helps to ensure that the final design solution is user-centered and meets the needs of all stakeholders. As a result, the design framework enables the development team to adapt to changing requirements, and ultimately enhance user and regulation satisfaction.

Furthermore, using a visual framework also promotes a more iterative and adaptable design process. By clearly depicting the various design elements, it becomes easier to visualize how changes or updates may impact the overall design.

The node tree structure, which is used as a basic knowledge representation in the visual assistant, takes care of this, as it can show how a change in one design step influences all the following decisions. The nodes’ left-to-right orientation helps users understand the connections intuitively (Figure 8).

The principle of a tree-based visualization of the AD stages is also well aligned with the concept of a knowledge graph regarding the domain knowledge clustered in CR, FR, DP, or PV nodes. To fully implement a structured design process, semantic and logical reasoning between the connections of these elements needs to be added. Following the idea of CBR, a new visual design assistant component complements the AD tree structure.

The relation between the CR or FR and DP can be described through the terms used in CBR. An FR is a short paragraph containing information about selecting the DP. Our design assistant proposes a method to split the textual requirement description into a set of key–value pairs following the concept of indices from CBR. These indices form the condition or selector for the individual Case or Norm, equal to the DP in our implementation. Figure 9 shows the method with the terms used in CBR (a.) and an implementation (b.) example. The color code helps the Designer avoid becoming confused with other indicators like the arrows for the AD matrix representation.

The given example uses generic keys like Function or Object, well-known from object-oriented programming (OOP), to identify the index types. Object-centered modeling is a form of structured description in product design that copes with design evolution in an iterative process [62]. These are a good starting point for a visual design assistant that builds on a database with structured information.

Using CBR, the virtual design assistant becomes more adept at understanding and responding to design challenges, ultimately contributing to a faster and more efficient design process that learns from and builds upon previous experiences.

Method Summary and Explanation

The CADA Method aims to create a human- and machine-readable dataset in the form of SN and GEs along a design workflow for a regulated industry that depends on transparent decision making and human evaluation.

Figure 10 shows the synthesis of frameworks used to build the CADA Method. First, the design workflow with the four domains (Figure 10a) from AD assists the designer while performing a structured analysis of the design task, its decomposition into sub-tasks, and its solution. The connections between elements from various domains, illustrated as Cases (Figure 10b), reflect the crucial problem–solution relationships required for knowledge storage within Case-Based Reasoning (CBR) systems. Within a specific domain, recurring sets of information are integrated into Norms (see Figure 10c). The relationships between domain-specific datasets and Cases harbor essential information, as they constitute the AD-Matrix depicted in Figure 5. The colors assigned to these connections signify the quality of the design in relation to the independence axiom from AD. Within the framework of the CADA Method, this axiom suggests that Cases should remain independent of one another. Green connections (see Figure 10d) represent uncoupled cases, orange connections (see Figure 10e) indicate decoupled cases, and red connections (see Figure 10f) denote coupled designs.

The idea of “zig-zagging” in AD theory refers to the process of breaking down a specific task into explainable sub-elements. The CBR workflow can be used to efficiently decompose a design and its requirements along the AD domains on top of existing knowledge. In our method, key–value pairs, named Indices, are combined into a Norm (Figure 10c) to describe an information subset of a domain attribute, such as CR, FR, DP, or PV. This type of data structure enables the identification of similarities among combinations of Cases, Norms, and Domain Attributes within a CBR workflow. In the CBR cycle, the first action, called Retrieve, focuses on locating data in the case memory that corresponds to the given input. Customer requirements can result in a set of established and, therefore, reusable functional requirements, which are associated with design parameters and known process variables. In the subsequent revision phase, the designer evaluates the datasets, confirming acceptable elements, introducing new items, and adding any missing references. To complete the process, the updated dataset is saved in the case memory during the retention step.

All information and references are formatted as depicted in Figure 10. The graphical representation of data, inspired by SN, offers a clear structure and guidance for designers. Any modifications to individual datasets reflect their impact on downstream domain items, facilitating visualization even during iterative or agile solution-finding processes.

The necessary SN to create a comprehensive description from CR to PV can be quantified as a specific number of blocks (e.g., datasets) interconnected along the CADA workflow. In our design assistant, the complexity of the solutions can be articulated through the count of blocks and connections required to define the solution. When multiple solution paths are available, the designer should opt for the one that demands the least amount of description. This aligns with the second axiom, known as the Information Axiom, which asserts that the superior design is the one that can be conveyed with the least amount of information.

3.3. Method Testing

A set of pharmaceutical engineering design cases was performed to evaluate the implementation of the developed design method and implementation as workflow. The main goal of this study is to test the ease of use and quality enhancement made visible by such a visual way of structuring data and design steps. This study excluded the mapping of DP to PV as part of the general AD workflow. No other elements are in use besides those in the FR to DP mapping.

Three scenarios from different aspects of the equipment design showcase the applicability.

• Design recommendation by similar requirement statements;
• Risk-based decision-making assistance;
• Iterative design.

Design knowledge is represented as a problem-to-solution (FR to DP) connection. In the first scenario, data in the form of two FRs, including Name–Value pairs that describe requirement properties, are created. Manually identified DPs are used as a solution approach to form a valid case inside the CBR data structure.

This design problem (Figure 11) requires that linear accelerated products be counted passively (are not process-interrupting). The shape and material form the object description, and the appearance and detection mechanism detail the FR. As a solution, a capacitive sensor with a constant detection range can detect the products partly because the cylindrical shape leads to a distance deviation between the object surface and the sensor.

In a second scenario (Figure 12), a similar cylindrical product with equal FRs needs the DP to proceed in a conceptual design within the AD workflow. Through the design knowledge database within the CBR layer, the tool is able to recognize this set of FRs to recommend the DP suitable for this combination of requirements. In cases where more than one possible solution can be identified, the most used one is recommended to use the same components more often (e.g., to benefit from economies of scale).

The next test case deals with the requirement from the pharmaceutical sector where decisions are made according to the possible risk for the processed product. Our proposed workflow can help to identify risks as the combination of the FR and DP incorporates the information to evaluate possible impacts on the product quality. The scenario describes a situation that could follow the previously discussed application of counting a product. A previously detected vial made of glass is to be gripped by a robot. To keep the scene clear, only the robot tool is discussed below. The FR is described by an Object (a glass vial with a cylindrical body) and a Risk that can occur when force is applied to the surface. The effect is described as micro-cracks that can harm the product as the barrier between the atmosphere and the product may be out of specification. The first design (Figure 13) uses a commonly used parallel gripper with specially designed jaws to center and clamp the product in the gripper. The combination of glass as the product material and the working principle of a parallel gripper (being position-controlled) may lead to a force on the vial surface if the product diameter is above normal. The CBR structure can identify this risk and recommend the use of a different approach.

The graphical presentation of the combination of the FR and DP that leads to the risk helps the user create a different approach. In the consequently selected design (Figure 14), a vacuum suction cup can solve the design problem as no direct force for gripping is used. Instead of clamping, the vacuum mechanism works more like holding the product in a position. The deviation in diameter is less critical to this technology; therefore, all risk factors do not take effect.

The third scenario is an application from the robotic automation sector created for the pharmaceutical branch. As the input data (CR), the task is described in two statements. On the one hand, there is a separation problem, and the to-be-designed process should be able to take a tray filled with products as the input and hand over individual products to a downstream process. On the other hand, the downstream process requires precisely positioned products at a specific handover point.

Following the guidance of the visual assistant, the Designer takes all the core information from these statements and maps it to the predefined categories (the Object and Function for this case). By doing so, the Designer identifies two types of Objects and three Functions to be transferred to the FR Domain.

In the second step, the FR is identified along with the Object and Function definitions from the CR statements. In this case, the FR properties must match the Object level and complement the definition with a Function component. CBR offers solutions that were generated in the past. In this scenario, the separation of products was successfully completed by picking individual objects captured in the CBR definition as a norm. The assistant helps with feedback as it shows an intersection between domains, also visualized by the color code in the CBR elements.

In the third step, the Designer creates content for the DP domain. The solution-finding process is supported by the sub-elements assigned to the FRs. The visual assistant creates direct feedback on the solution quality along the three categories coming from AD. If a DP is defined and based on more than one FR and, at the same time, one FR is also connected to another DP, the design is categorized as coupled according to AD and marked with a red arrow. Figure 15 shows a coupled and decoupled DP, resulting in an unstructured mapping visualization. To gain a better design, AD enforces the splitting of the FR and DP to investigate the design problem further.

Figure 16 captures the result of the second iteration. The Designer split the FR to add more details to the intersecting requirements. This step allows the FRs to be addressed better by individual DPs and thus avoid coupled designs. The visual assistant helps identify the items where action is needed. Combined with the rationale in the CBR items, the design is structured, traceable, and extendable.

The direct visual feedback offered by this design assistant helps identify DPs related to one or more FRs. This qualitative feedback is easy to understand.

4. Discussion

This research discussed a specific design method covering the needs of a domain belonging to the pharmaceutical engineering environment. Design methods in this section need to be better represented in the design community. Our proposed method combines well-established tools and modern data architecture. The positive effect of the domains and the workflow coming with AD is also recognized in a combination of AD and TRIZ [50]. A combination of different tools like TRIZ and PSN still needs detailed instructions on how to translate the problem into a solution in a compliant way [27]. In our case, AD incorporates this method, and thus, the Designer benefits from the guidance through the domains of a structured approach within AD.

Some research design knowledge methodologies concentrate on a purely textual knowledge representation, and AD does so in its original form. Conceptual Design also builds on sketches, CAD models, and images [48]. In terms of product design knowledge on the enterprise level, a combination of TRIZ, Functional Tree Design, and relationship mapping is a core enabler for structuring design data for reuse and evaluation [63]. Our proposed method, incorporating the structure of a knowledge graph, is open to textual and graphical content. Complemented by the CBR attributes which are mandatory for our data structure, the inserted information and the basis for algorithms that produce problem–solution relations are still structured. Such a multidimensional approach is covered in the Design Knowledge Semantic Network (SKSN) by Yue et al. [64], showcasing the relevance for SN-based design knowledge representation among the technical literature. Knowledge Graphs are a proper way to structure data for visualization purposes [44]. Guidance through graphical means of data representation while creating a database for CBR is already used in similar tools which focus on different outcomes, like the iSee platform [65]. As mentioned, tools like semantic networks lack a structured process or workflow and do not support our specific domain in detail [48]. Our approach covers the need for a regulated environment, particularly transparency in decision making along the design and engineering process. Ontology-based information models for pharmaceutical engineering already exist when it comes to risk analysis [49]. Similar structures are used in the presented examples on how to use the proposed workflow.

CBR is used across several industries and research communities. A recent comparison of available frameworks shows that there is quite some interest in the development and use of CBR-based tools. According to Schultheis et al. [66], only one of the most recommended tools uses graphical means of data representation. None of the evaluated frameworks from this research intend to be embedded in a structured design workflow like AD.

The V-Model, which serves as the primary design technique in pharmaceutical equipment engineering, presents certain limitations when applied to complex, multi-domain projects. This model places a significant emphasis on detailed descriptions and specifications before implementation, a scenario that is rarely encountered in conceptual design. In reality, design usually evolves through iterative processes, where ideas are developed and concepts undergo continuous refinement. Furthermore, contemporary methods such as rapid prototyping may require expertise from various domains, which not all designers possess. Our model offers a solution for enterprise-wide design knowledge collection that can be queried to retrieve information about specific components from previous designs. The V-Model approach does not inherently support the parallel specification of modules or functions intended for integration into a single piece of equipment as it lacks modeling capabilities for interfaces between distinct development paths. In contrast, our modeling approach enables the creation of individual end-to-end concepts that can be cross-checked and interconnected within our proposed SN-inspired graphical way of modeling. Quality assurance within the V-Model framework is founded on its internal structure, where each specification must be accompanied by a corresponding test description. This concept is well-established in pharmaceutical engineering, as it guarantees high-quality outcomes with reproducible functionality through iterative testing following any changes. However, this approach may not allow for the individual solution of functional requirements (e.g., in a modular fashion), which are necessary for creating interchangeable components that can be implemented and tested independently. CADA introduces a structured workflow that enables designers to recognize coincidences or inter-relationships among building blocks, thereby promoting better concepts with a reduced likelihood of unintended effects during implementation and testing. The advancement of modern pharmaceutical equipment demands both rapid market entry and superior product quality. To comply with established regulatory standards such as Good Manufacturing Practice (GMP) and Good Engineering Practice (GEP), it is crucial to provide the comprehensive descriptions and thorough documentation of the equipment and its internal processes [12]. One effective strategy for achieving this is the construction of new solutions from well-proven and well-documented modules. This modular construction approach is gaining popularity across various engineering disciplines [67]. Our methodology supports modular engineering based on enterprise design knowledge, facilitating adherence to pharmaceutical standards from the initial stages of design. Furthermore, the integration of the CBR technique within our approach inherently fosters continuous improvement, a vital component of quality management systems. While these techniques are essential for the pharmaceutical industry, they may apply to other sectors. Our model can be used for design tasks in regulated industries beyond the pharmaceutical domain; however, for unregulated branches, it may require additional or unnecessary efforts.

CADA is designed for a continuous learning organization, ideally used by a larger number of users. As a result, it may not be as effective in start-ups or small companies. While quality feedback can still be beneficial, the recommendation feature may be limited, as it requires a substantial case memory to function effectively.

5. Final Remarks

5.1. Main Conclusions

This research addresses the individual need for a structured and visual design method to support early-phase design processes in a regulated environment. The ease of use and user guidance throughout the decision-making process are key attributes of this work.

Our method consists of three main parts built on top of each other: Semantic Networks, Case-Based Reasoning, and Axiomatic Design. It was evaluated by three cases covering pharmaceutical automation problems. More diverse studies along different customer requirements from pharmaceutical engineering or other industries should be conducted to test the applicability for various domains. Further research can identify the potential applicability of TRIZ as an information model for early-stage designs to strengthen our proposed workflow [24,25].

Its visual and, thus, qualitative feedback cannot be directly compared to quantitative evaluation and decision-making tools. The incorporated data structure following the concept of CBR is expected to be a very good dataset for AI techniques to evolve the innovation process towards an automatic design based on user requirements. Further research on the implementation of a training pipeline based on our assistant’s information structure should be carried out.

5.2. Limitations and Contributions

The present research and implementation generated a visual design assistant following the guiding structure of AD complemented by CBR and SN. Also, the literature research was exploratory and thus not proven to miss out on important information to complement the proposed theoretical method. Complex problems depend on multiple layers of problem analysis and design in the implementation process. AD does not stimulate the search for various approaches to find the best solution possible. As soon as the Design Matrix shows an uncoupled or even decoupled design, the user requirements seem to be covered and the design can go to engineering. Innovation and creativity techniques are not part of the examples provided in this article. Our proposed workflow is open to ideation techniques as its interface does not limit the Designer in using such. We recommend to incorporate further ideation techniques into the design studies to overcome this downside [24]. In our proposed method, the recommendation and evaluation are only as good as the available data, in the form of cases of already existing designs. Several design studies have to be taken into consideration until the intrinsic possibilities of CBR and the complementing methods show results in the form of recommendations or decision-making support.

This novel method is the main contribution to the theory as the result follows the principles of a knowledge graph for a natural understanding of the relation between the requirement and solution. The intrinsic workflow guides the user through an iterative solution-finding process with a visual feedback of the design quality stated in the three categories defined by AD.

In practice, this method can support decision making and tracing through its underlying graphical and easy-to-understand information model along different departments or areas of the subject. As a future work, the method needs to be implemented into user-centered software to showcase and evaluate its practical application within a structured design workflow. With such an application, empirical research with equipment designers can be carried out to test the applicability and further improve the mechanisms used in our proposed CADA method.