Do Buzzwords Dream of Clearer Substance? The Journey of 'Digital Twins' Toward Becoming an Operational Concept

A fancy concept but still quite complex to understand and define

Perhaps I’m not fully engaging my LinkedIn audience (maybe I need to add a few emojis for a touch of color ??), or it might be that the concept of 'digital twins'—though often seen as a buzzword—is still not widely understood in terms of practical, operational applications, or both! While it’s a trendy term, there seems to be a gap in understanding how digital twins can truly drive outcomes at scale.

Digital twin definition

Let’s unpack the semantics of 'Digital Twin'! It’s a term made up of two parts:

? What does 'Digital' mean? (And it can mean a lot of different things, IoT, wearables, 5G, Scada, knowledge Graph, AI algorithms...).

? What exactly are we 'Twinning'? (And no, not related to Twinings Tea—just a fleeting thought due to my hot mug full of tea! ).

Each word carries multiple meanings depending on the industrial context and specific use cases, which is why the concept of a 'Digital Twin' itself is so richly polysemic.

"A Digital Representation of a Real Object: For many, the concept of a 'digital twin' is strongly associated with the 3D representation of a real-world object, especially with the rise of augmented and virtual reality. This immersive visualization has become a defining feature in how people perceive digital twins."

As Sébastien Brasseur noted in my post (link), a 2017 study by Negri et al. found that the most frequently occurring words were "physical" and "product," indicating a strong emphasis on a physical product-centered perspective of Digital Twins (DT).

Nonetheless, alternative definitions advocate that the potential of DT should extend beyond tangible physical products, encompassing processes, systems, and even entire organizations.

A review of the literature shows that researchers and practitioners have proposed diverse definitions and applications for Digital Twins (DT), each seeking to clarify their interpretation of the term. Some envision DTs as complex, real-time digital models with predictive and prescriptive capabilities, while others consider them as simple digital representations.

In our discussions with customers about "Digital Twins," it has become clear that there is no universally accepted definition of what a Digital Twin truly is/should be, nor a shared understanding of how it can drive ROI-proven initiatives at scale within their value chains and existing organizational structures.

This is typical in a blue ocean strategy, where "making the market" and educating customers to raise their "Digital Twin" maturity are essential. However, it’s important to account for this factor, as it will inevitably extend the sales cycle.

This conceptual ambiguity, however, risks DTs being dismissed as mere hype due to their lack of clarity and genuine ROI-proven business-case going beyond the sole use-case.

Reflecting on Recent History: Remember Predictive Maintenance?

Predictive maintenance was once hailed as an obvious, high-ROI use case. Though not explicitly labeled as “Digital Twin”-based, it embodied the core concept: a digital representation of a complex physical object, aimed at forecasting part failures so operators could preemptively replace components and avoid costly production downtime. However, the business case ultimately faltered. Many comparisons optimistically assumed a leap from no maintenance at all to a fully predictive maintenance landscape—overlooking the fact that "scheduled maintenance" already existed as a robust practice to reduce downtime. This oversight significantly undermined the projected impact and led to inflated expectations.

Toward a standardized definition of Digital Twins

Digital Twins (DT) comprise two primary components: a physical entity and a digital (or logical) counterpart. Rather than focusing solely on the digital model, DT integrates both the real-world system and its virtual representation, along with the interaction between them.

The physical component, known as the Physical Object (PO), represents the actual item or system—whether a device, product, hardware, or even a physical process operating in the real world.

The digital component, or Logical Object (LO), is a virtual model of the physical system, generated through software that integrates data and algorithms. Often referred to as a digital clone or replica, this model reflects the characteristics and behavior of the physical object—within specified parameters and to an extent to be chosen when architecturing the Digital Twin.

In essence, DT links these two elements: the tangible system existing in the real world and its software-based counterpart, which can simulate and predict the system’s performance and behavior.

Properties of a Digital Twin

As described in the "Handbook of Digital Twins" (editor Zhyhan Lyu), Digital Twins should exhibit the following properties:

• Representativeness and Contextualization: The Logical Object (LO) should credibly mirror the Physical Object (PO), although capturing all facets can be complex and costly. Therefore, DT models are designed with specific objectives aligned to their application context, focusing on the essential properties, characteristics, and behaviors needed to qualify as an LO. This concept is often misunderstood, particularly in Western contexts influenced by analytical Cartesian thinking. The LO is not an "exact" replica of the PO but rather an abstraction. The level of abstraction must be carefully chosen, as higher complexity increases costs exponentially without a proportionate gain in quality or relevance.
• Reflection: This characteristic ensures that the PO’s status, attributes, and behaviors, which may vary over time, are accurately represented in the LO, with each relevant value uniquely reflected.
• Entanglement: Entanglement describes the communication link between the PO and LO, ensuring all necessary information flows to the LO for accurate representation of the PO. This link may involve a "real-time" aspect and can leverage IoT, IoMT, and wearable technologies.
• Replication: Replication is the ability to create multiple instances of a PO across different virtual environments, enabling scalability of DT operations. For example, the same DT model can be deployed across multiple data streams in various production lines to monitor and control production quality. This capability is essential for supporting scalable and viable business models.
• Persistence: While the PO may face physical limitations, the LO maintains the DT’s continuity by overcoming these restrictions to ensure constant availability. Once the LO has proven reliable and accurate, it can independently generate synthetic data, even in the absence of the PO. This capability enables simulation of future PO deployments, such as evaluating performance on a new production site.
• Memorization: The LO must retain and represent all relevant historical and current data of the DT, which is crucial for maintaining the DT throughout its lifecycle. When integrated with MLOps, this capability enables regular updates and facilitates model forking to create refined or specialized versions.
• Composability: Composability allows multiple objects to be combined into a single composite structure, enabling observation and control of both the entire composite and its individual components. This is essential for a multimodal DT approach, as it enables the integration of additional data dimensions to enhance the DT’s representation of the PO.
• Accountability/Management: Ensures comprehensive management of the DT, enabling interaction between different LOs and the assembly of larger aggregates for DT construction.
• Servitization: Provides users with services, functionalities, and data access related to the PO through DT’s tools, software, and interfaces.
• Predictability: Supports the simulation of the LO’s behavior and interactions over time or within specific contexts to predict the PO’s future performance.
• Programmability: Offers APIs that enable the programming of DT functions

Maturity models of Digital Twins

Based on the “entanglement” property, Digital Twins can be classified into various maturity stages:

1. Digital Model: At this stage, there is no automated data exchange between the Physical Object (PO) and the Logical Object (LO). The LO is created from data representing the PO, but in an asynchronous manner. This means changes to the PO do not affect the LO and vice versa. Essentially, the LO is a static digital model of the PO.
2. Digital Shadow: Here, there is one-way data synchronization from the PO to the LO. Any change in the PO is reflected in the LO, but changes in the LO do not impact the PO. For example, a fraud detection application monitoring financial data streams in real time is a "digital shadow," as it is informed by live data but cannot alter the source system.
3. Digital Twin: This is a fully interactive stage, with two-way data synchronization between the PO and the LO. The LO can function as a "virtual cockpit" to manage the PO, allowing not only for real-time data analysis but also for sending commands back to the physical object. In Industry 4.0, for instance, real-time data from IoT sensors on a production line can be analyzed by the digital model, which then sends control commands to actuators.

Some models introduce additional advanced stages, focusing on other aspects:

1. Cognitive Digital Twin: This advanced stage involves predictive capabilities, where the DT can anticipate the PO’s behavior based on AI algorithms. It can potentially correct deviations from normal operation without human intervention.
2. Collaborative Digital Twin: In this model, the DT acts as an assistant for human operators rather than automating responses. It provides insights and recommendations to support strategic decision-making but does not directly control or alter the PO.

While the first three stages primarily relate to entanglement and data synchronization, the Cognitive and Collaborative stages focus more on AI-driven insights and user experience, enhancing human collaboration and decision support.

Toward a standardized definition of Digital Twins

"A Digital Twin is a virtual representation of a "system"—whether it be a physical object, complex machine, organism, process, or organization—that mirrors its real-world characteristics and behaviors to a defined extent, with specific levels of reliability and complexity. By integrating real-time or asynchronous data and simulations, it allows for monitoring, analysis, predictions, and, in some cases, interactive feedback loops that enable users to influence and optimize the system’s performance. This interactive capability enhances decision-making and control across the system's lifecycle, creating a more dynamic user experience."

• Wide Applicability: The inclusion of a variety of potential "systems" (physical objects, machinery, organisms, processes, and organizations) reflects the broad applicability of DTs across multiple domains. This breadth is valuable for readers unfamiliar with DTs, as it highlights the versatility of the concept and expands the perception of DTs beyond industrial uses.
• Specificity of Representation: The phrase "mirrors its real-world characteristics and behaviors to a defined extent, with specific levels of reliability and complexity" underscores that DTs are tailored models rather than exact replicas. This is important because it clarifies that the fidelity of a DT depends on the requirements of the application, balancing model complexity with operational relevance and cost-effectiveness.
• Data Integration and Simulation: By mentioning both real-time and asynchronous data integration, the definition acknowledges the diverse ways DTs are implemented. This aspect makes it applicable to a range of use cases—from high-frequency industrial processes to scenarios that only need periodic updates, such as organizational or environmental simulations.
• Emphasis on Interactive Feedback Loops: Adding "interactive feedback loops" brings an important dimension to the definition. This phrase highlights that DTs are not just passive models but can influence their physical counterparts, depending on the maturity of the DT. This interactivity is especially relevant in advanced applications where the DT becomes an active tool for real-time adjustments, predictive maintenance, or automated optimization.
• User Experience Focus: The final sentence—"creating a more dynamic user experience"—positions DTs as user-centered tools. By emphasizing the UX impact, it connects DT functionality directly to its value in decision-making and operational control. This makes the definition relatable to a broader audience, including those concerned with user interfaces, decision support, and system control.

How do we implement this in TweenMe (the first universal digital twin generator) ?

TweenMe Positioning in the Digital Twin Maturity Model

TweenMe is positioned as a universal generator of both Digital Models and Digital Shadows:

• Asynchronous Data Integration: TweenMe integrates datasets asynchronously within an automated data pipeline, enabling the creation of flexible digital representations of physical objects or systems. Additionally, TweenMe includes advanced ETL capabilities to ingest data in real time from points of service when required.
• Digital Model Generation with MLOps: TweenMe produces Digital Models leveraging MLOps, supporting various use cases such as regression, clustering, segmentation, classification, dimensionality reduction, and reinforcement learning. In predictive use cases, TweenMe generates “Cognitive Digital Twins” capable of forecasting the behavior of the Physical Object (PO) represented by the Logical Object (LO) or Digital Model.
• User Experience and Collaborative Digital Twins: TweenMe packages the Digital Model into an application that optimizes User Experience (UX). For strategic decision-making, TweenMe can produce a "Collaborative Digital Twin" to explore "What if?" scenarios, aiding users in evaluating potential outcomes and guiding decisions.
• Future Integration with Bidirectional Synchronization: TweenMe's outputs can be integrated into broader systems that enable bidirectional data synchronization between the PO and the LO/Digital Model/Digital Shadow. However, such integration depends on specific use cases, existing infrastructure, and data architecture, and therefore cannot be fully automated.

This positioning highlights TweenMe's capabilities in creating flexible, predictive, and collaborative Digital Twins adaptable to various maturity levels in digital twin development.

What does make TweenMe a universal DT generator ?

A Digital Twin is a model or abstraction that must be thoughtfully designed to represent the Physical Object (PO) with a level of detail appropriate to the specific business problem, leveraging available production data. Defining the right level of abstraction and data model is critical to ensuring the Digital Twin’s business value. While TweenMe reduces the expertise required and significantly lowers the marginal cost of creating new Digital Twins, aligning the model with business goals remains essential for success.

Example: Modeling the Clinical Benefit of Paracetamol

From a clinical standpoint, I could develop a straightforward model to predict the effects of acetaminophen (paracetamol) as a pain reliever and fever reducer. The clinical question here would be: Given a certain dose of acetaminophen, what level of pain relief and fever reduction can I expect?

A basic dataset might include:

• Dosage: Amount of acetaminophen taken
• Pain Relief: Measured by patient-reported outcomes (e.g., questionnaires)
• Fever Reduction: Recorded by thermometer readings

This simple model could be expanded by incorporating additional variables for a more nuanced understanding of the clinical effects.

However, as an analytical thinker, I might also explore what happens once acetaminophen enters the bloodstream, examining its molecular actions and enzyme interactions, such as:

• Weak inhibition of COX enzymes, especially in the brain,
• Modulation of the endocannabinoid system via the metabolite AM404,
• Potential activation of TRPV1 receptors and influence on serotonergic pathways.

Further, I could delve into modeling the molecular binding between acetaminophen (or its metabolites) and specific receptors, possibly through quantum chemistry simulations to estimate binding forces.

While such detailed modeling—like visualizing the 3D docking of acetaminophen's metabolite AM404 onto CB1 receptors—might be exciting and yield insights at the molecular level, it would not directly address the initial clinical question.

This underscores the importance of defining the right question. Effective Digital Twin development requires balancing data availability, expected outcomes, and development costs against the relevance and reliability of the model, ensuring the solution aligns with the practical needs of the problem at hand to build the optimal data model to address the business question with accuracy and minimal budget.

Once the data model is designed, it can be "executed" in the TweenMe automated data pipeline to produce a Digital Model / Digital Shadow or a Digital Twin.

Example: Developing a Multimodal Model for Senology

We began with an open dataset containing features extracted by a computer vision algorithm from 569 mammograms, each labeled as either benign or malignant.

Our first step was to refine the dataset by identifying clusters of tumors with differing growth speeds, categorizing them as either slowly growing (benign or malignant) or rapidly growing. We then created additional clusters based on tumor surface characteristics, classifying tumors as either "smooth" or "rugged" based on image-derived features.

Next, we built covariance matrices to integrate synthetic genomic and proteomic data, focusing on tumor attributes like "growth speed," "smooth vs. rugged texture," and "size." For this, we applied a range of analytical techniques, including Covariance Matrices, Principal Component Analysis (PCA), Multivariate Linear Regression, Partial Least Squares (PLS) Regression, and Machine Learning models such as Support Vector Machines (SVM).

By examining these covariance patterns, we defined two new output variables: "tumor aggressiveness" and "treatment response", which we linked to proteomic markers such as ER, PR, and HER2. These derived variables provide insights into tumor behavior and potential therapeutic outcomes, enhancing the model’s clinical utility as the model answer was no more limited to only "benign" or "malignant" classification.

Beyond serving as a "Digital Twin" generator, TweenMe also creates/stores data models (ontologies) and insightful databases, which can ultimately be used to train cognitive digital models for integration into digital shadows or digital twins.

Take aways

1. Digital Twins as a Polysemic Concept: The term "Digital Twin" encompasses multiple meanings across industries, ranging from simple digital models to complex, real-time, and interactive representations of systems or processes. This semantic complexity often leads to varied interpretations, making it challenging to establish a universal understanding of its practical applications and ROI potential.
2. The Importance of Defining Scope and Abstraction: The level of detail in a Digital Twin must be carefully calibrated to fit specific business objectives. Overly complex models can increase costs without proportionate value, while simpler models may not provide the depth required for decision-making. This balance is essential for maximizing the effectiveness and efficiency of Digital Twin solutions.
3. Properties and Maturity Levels of Digital Twins: Digital Twins can vary significantly in maturity, from basic "Digital Models" with no real-time data exchange, to "Digital Shadows" with one-way updates, to full-fledged interactive Digital Twins. Advanced stages, like "Cognitive" and "Collaborative" Digital Twins, add predictive capabilities and facilitate strategic decision-making, underscoring the evolving nature of Digital Twin applications.
4. Practical Use Cases and Proven Benefits: Despite initial enthusiasm for use cases like predictive maintenance, some Digital Twin implementations have failed to deliver anticipated ROI due to overly optimistic comparisons and a lack of clear business alignment. This highlights the need to set realistic expectations and align Digital Twin capabilities with concrete business outcomes.
5. TweenMe’s Approach to Digital Twin Generation: TweenMe positions itself as a universal generator of Digital Models and Digital Shadows, reducing expertise barriers and marginal costs in Digital Twin development. However, effective implementation still requires aligning the model with specific business goals, defining the appropriate abstraction level, and understanding the limits of automated integration to maximize operational impact.