Digital twin entities: a matter of size and level of abstraction

While the combination of type of digital twin and its scope influence the technological stack, the scale of the physical entity being replicated directly influences the complexity, implementation timeline, and accuracy of a digital twin.

Smaller entities, like a single component or workstation, typically involve less complexity and are easier to simulate and optimize. They provide detailed insights that are critical for improving specific parts of a process or product. For example, a digital twin of a turbine blade might be used to test its durability under different operating conditions, providing valuable data that can improve design and manufacturing processes.

As the size of the entity increases, such as when dealing with a complete product system, production line, or entire supply chain, the complexity grows exponentially. This increase in complexity requires more advanced simulations, more sophisticated data integration, and more robust computational resources. For instance, a digital twin of a global supply chain must account for numerous variables, such as multiple products, different shipping types, stochastic lead times, transportation costs, and geopolitical risks, all of which interact in complex ways.

In monitoring digital twins, which track real-time data from physical entities, the complexity also translates into higher capital and operational expenditures. For example, monitoring a complex manufacturing site serving complex product mix involves tracking various machines, sensors, and processes, requiring substantial investments in IoT devices, PLCs, and data management systems.

In general, increasing the size of the digital twin – costs aside – is anyhow feasible when it comes to monitoring digital twins.

However, predictive, simulative, and prescriptive digital twins face further particular challenges as entity sizes increases.

These models rely heavily on accurate data and sophisticated algorithms to predict future outcomes or recommend optimal actions. As the entity becomes larger and more complex, the likelihood of encountering errors, double counting or misconfigurations increases.

Large-scale predictive twins can be compromised by a single misconfigured constraint or an unclear variable, leading to unreliable outputs. In facts, while very low additional value is produced by inserting more and more variables into the model, the sheer number of variables increases the risk of:

• • reaching local sub-optimal solutions 
  • to completely fail to find a solution 
  • adding white noise with redundant (i.e. two variables very similar and correlated to each other) and irrelevant variables, causing the model to fail in replicating the entity behavior
  • Overcomplicating the model into a fuzzy, unclear representation of the physical entity 

Additionally, the “black box” nature of many predictive models can make it difficult for users to trust the outputs, especially when numerous variables are involved.

On the other hand, simulative digital twins are affected by different criticalities. As simulative digital twins attempt to mimic real-world behavior by breaking down complex systems into smaller, interacting entities (for example, they can re-create a virtual replica of the global trades by the interaction over time of digital consumers and producers in the virtual environment), adding more variables does not necessarily increase the goodness of the morel results, but it might increase the discrepancy in terms of behavior and performance between the real-world entity and the virtual model. In facts, as each variable implemented into the digital twin is affected by certain error margin, adding more and more variables imply increasing the error quantity. As such errors interact with each other in the model, they lead to more volatile and randomic results, which might be therefore more positive or negative, where positive and negative is defined by the correlation of each variable error as a result of random number operations.

Moreover, as already stated for the predictive digital twin, more variables heighten the complexity of the model, the potential for mistakes and the interpretability of the outputs, leading to distrust the digital twin.

To manage this complexity, the level of abstraction in a digital twin must align with the scale of the entity being modeled. For example, in a complex end-to-end supply chain model, simulating every single machine or operator within each facility may be unnecessary and could even distort the overall results. Instead, focusing on higher-level abstractions, such as the distribution of a facility throughput or lead times, can provide more actionable insights without overcomplicating the model.

In contrast, in a narrowly focused digital twin, such as one that models a single process step, detailed simulations of individual machines or operators can add significant value. For instance, a digital twin of a specific welding process could help identify operational improvement opportunities such as discovering the optimal settings for machine parameters to achieve the highest quality welds, leading to immediate and tangible improvements in the manufacturing process.

However, there are sinergies between small entity and large entity digital twins. In facts, to exploit the best out of process digital twins, typically, outputs from smaller digital twins are used as inputs for broader ones, creating a layered and interconnected system of insights.

As an example, consider a digital twin of a single workstation in a manufacturing line, such as a CNC machine. The inputs might include tooling configuration, machine speed, and material type. The output could be an optimized cycle time or a reduction in tool wear. This twin might also predict how adjustments to tooling disposition can shave seconds off the cycle time.

A digital twin of an entire process step or manufacturing line would take the optimized cycle time from the workstation twin as an input. Other inputs could include machine availability, labor schedules, and material flow. The output might be the overall efficiency of the process step or line, identifying bottlenecks and suggesting adjustments to maximize throughput.

Scaling up further, a digital twin of an entire manufacturing site would integrate inputs from various process steps and production lines. It might also consider factors like energy consumption, logistics, and workforce distribution. The outputs would be high-level insights, such as the optimal configuration of the entire facility, strategies for reducing downtime across multiple lines, or comprehensive energy-saving recommendations.