Industry 4.0 is currently revolutionizing the design, manufacturing and services of healthcare applications. It includes the automation and digitization of manufacturing techniques, the acquisition and analysis of process data and the communication between machines. Together, these innovations result in self-managed processes that turn traditional practices and manufacturing into so called smart factories.
What is a digital twin?
An important aspect of these smart factories are digital twins. A digital twin is commonly described as a virtual replica of a physical product or process. It is a real– time computer model that can be used to predict process outcomes, optimize a process design or monitor and control the production process. The digital twin receives real-time sensor data of the process and uses historical data or foundational assumptions (e.g. physical laws) to simulate the process evolution.
In this way, both the efficiency and quality of the process can be increased, which also generates an important economic cost-benefit. This is especially important for ATMPs, whose manufacturing processes are typically very complex and often suffer from large variability.
The computer models that are used in digital twins can be subdivided into so-called mechanistic and data-driven models. Both model types have their own strengths and weaknesses, and some digital twins may primarily or even exclusively use one or the other, depending on the application. However, the most successful digital twins typically use a combination of both.
But first, let’s take a closer look at the differences between these models.
Mechanistic model: process insight
Mechanistic models are directly based on the underlying scientific laws of the process. The biggest advantage of mechanistic models is that they can provide insight into the process because their parameters are based on the underlying physics. Therefore, a mechanistic model can also be easily adapted to changes. However, mechanistic models, in general, have a high development cost, their parameters can be difficult to validate, and they can require large computing resources, which might hamper real-time computations.
Data-driven model: speed and efficiency
The data-driven approach uses collected process data together with machine learning models to build (part of) a digital twin. In contrast to a mechanistic model, a data-driven model does not rely on physical and biological knowledge of the system, but uses process data to build a mathematical model that can simulate the process. They are sometimes called ‘black-box’ models, as they provide little to no new insight into the underlying dynamics of the process because the models only simulate patterns they learn from data but are not based on physical or biological principles.
However, a major advantage of these models is that their development is fast and efficient (provided that sufficient data is available. They can run in real-time with the process and their validation is relatively straightforward. The correlations they provide can for example be used to monitor the process Critical Quality Attributes (CQAs) non-destructively and in real-time, an application which is also known as soft sensing.
Hybrid approach: best of both worlds
In between these two extremes lies the hybrid approach that attempts to combine the best of both worlds. The fundamental idea behind this approach is to use the complementarity of the two opposing methodologies: a mechanistic process model can be used to model the part of the process about which established physical knowledge is available (e.g. fluid flow or nutrient distribution), while a data-based model is used to model the part of the process which cannot be modeled mechanistically because of the complexity or lack of available knowledge (e.g. cell growth kinetics).
Such models offer an appealing trade-off between development speed and process insight. Additionally, they can outperform mechanistic and data-based models with regard to predictive power and applicability.
Key role in process management
Digital twins play a central role in the digitization and automation of manufacturing strategies, particularly for the field of advanced therapy medicinal products (ATMPs), which are medicinal products based on cell, gene and tissue engineering processes. If successful, digital twins can be used to fast-track process development by allowing scientists and engineers to run simulation experiments as part of the process development. These kinds of experiments can be carried out much faster and at a much lower cost than typical ‘wet lab’ experiments, because the latter comes with long process times and high raw material costs.
Essential to personalized manufacturing
For ATMPs however, bear in mind that the process needs to be equipped with a flexible control strategy in order to ensure consistent product quality in the face of significant raw material variability, in particular for autologous processes. On top of this, the industry is currently moving towards personalized manufacturing in which the end product is tailored to individual patient demands, further stressing the need for a high level of non-invasive process monitoring and control. Without the power of a digital twin, this is impossible to do.