An interview with Marco Macchi, Full Professor at Politecnico di Milano, on the new alphabet of industrial production
In the debate on the evolution of manufacturing, terms like smart factory, sensitive factory, predictive maintenance, digital twin, and advanced asset management have entered the lexicon of innovation. The risk is that they are used as generic formulas, closer to tech marketing than to the real transformation of industrial processes. To understand when these concepts deliver value and when they remain premature promises, we need to bring them back into the real factory: its equipment, its data, its competencies, its day-to-day decisions.
This is precisely the ground covered by Professor Marco Macchi, Full Professor at Politecnico di Milano, where he teaches Industrial Technologies, Smart Maintenance and Industrial Asset Management, and Modelling and Data Analysis of Complex Systems. His profile combines scientific research, executive education, and direct engagement with the industrial world. In this interview, Professor Macchi helps distinguish between declared innovation and operational innovation.
Today, there is much talk of smart factory, but the concept of sensitive factory is also emerging more frequently. What is the difference between these two ideas of the factory?
The smart factory is certainly a more established concept, both in research and in industrial communication. It stems from a longer trajectory that was previously associated with the notion of the intelligent factory and now falls within the scope of smart manufacturing. The sensitive factory can be seen as a further evolution or variation of this paradigm.
When we speak of a smart factory, we refer to a factory capable of using connectivity, data, and digital technologies to make production processes more integrated. One of its key characteristics is the synchronization between what happens on the shop floor—on machines, equipment, and production lines—and what is decided at the managerial and organizational level. In this sense, the smart factory monitors its own operations in a pervasive way, collects information from the production process, and uses it to increase awareness and control over performance, anomalies, critical issues, and outcomes.
This availability of information enables more adaptive decision-making. If the system has better knowledge of actual operating conditions, it can react more quickly to changes, correct deviations, optimize resource use, and improve production continuity. In an initial phase, this adaptivity may be reactive: the factory observes what is happening and adjusts its decisions accordingly.
The next step is to make this capability proactive as well. This is where tools like artificial intelligence, predictive models, and simulation capabilities come into play, allowing the factory not just to respond to what is already happening, but to anticipate future scenarios. The factory can thus prepare for operating conditions that have not yet fully materialized, evaluate decision alternatives, and manage both its internal production environment and the broader supply chain context.
From this perspective, the sensitive factory can be interpreted as a factory that is even more attuned to its operating environment—a factory capable of perceiving weak signals, reading process conditions in depth, recognizing variations, and translating this sensitivity into more timely, contextual, and intelligent decisions. It is, in any case, a natural evolution of what I prefer to call the intelligent factory or smart factory.
In your experience, what are the most common obstacles that prevent companies from leveraging data from sensors, MES, ERP, and maintenance systems?
The obstacles are varied and can be grouped into three broad dimensions: technological, organizational, and cultural. These are distinct but interconnected levels, because the ability to leverage data also depends on how the company is structured, how it makes decisions, and how it works.
From a technological standpoint, the most frequent problem is fragmentation in data management. Many companies have sensors, MES, ERP, maintenance systems, or data collection platforms, but these elements are typically treated as separate components. The sensor captures one piece of information, the MES manages another, the ERP oversees business processes, while analytics or simulation tools, when present, operate on yet another level.
The critical point is that these systems do not always communicate effectively with one another. In other cases, some components are not yet in place or have been introduced at an insufficient level of maturity to support the objectives of the production process and system. The company therefore possesses individual pieces of the digital architecture, but not a truly integrated system capable of transforming raw data into useful information—and ultimately into decisions.
Then there is the organizational obstacle, which concerns the role of people, competencies, and responsibilities within the organization. To truly use data, the organization must be prepared to work according to a data-driven approach. This means being able to read the information coming from the field, interpret it correctly, connect it to production and maintenance objectives, and use it to modify behaviors, priorities, and decisions.
On top of this is the cultural dimension. In many companies, data is still perceived as a control tool or as a technical element confined to certain business functions, whereas it should become a shared resource. Leveraging data requires training, experience, and also a different habit of collaborative work across production, maintenance, quality, IT, and management.
In this regard, the generational factor can also play a role. Younger generations often have greater familiarity with digital technologies and with everyday data use, while those coming from more traditional organizational models may be influenced by established practices that are less oriented toward information sharing and evidence-based decision-making.
Predictive maintenance is often presented as an almost automatic promise of efficiency. Under what conditions does it deliver a measurable economic return?
Predictive maintenance is often portrayed as an automatically advantageous solution, but in reality it is not so straightforward to achieve the promising visions that are presented. Measuring its ROI is not immediate, and above all, it should not be considered a valid answer for every situation. Predictive maintenance is one of several possible maintenance strategies, alongside preventive, corrective, condition-based, and other approaches. It is not a panacea, but a tool that must be calibrated according to the production context, the type of asset, the criticality of the process, the economic value in terms of opportunity costs, and the risk to be kept under control.
The first thing to consider is that predictive maintenance is not free. If we understand it as a truly data-driven approach, it requires data collection infrastructure, sensors, monitoring systems, integration platforms, analytical models, data governance capabilities, and specialized competencies. When we talk about predictive maintenance, artificial intelligence and machine learning immediately come to mind, but in reality statistics, data analytics, physical modeling, process knowledge, and maintenance expertise all come into play. It is precisely the combination of these elements that makes the solution effective—but it is also what determines its actual cost.
For this reason, the economic return becomes measurable when the expected benefit clearly exceeds the cost of implementing and managing the system. Predictive maintenance makes more sense when it involves critical machinery or equipment, where a failure can lead to costly production stoppages, significant quality losses, safety issues, delivery delays, or substantial damage to assets as part of the plant infrastructure. In these cases, being able to anticipate an anomaly or degradation can produce concrete economic value, because it avoids consequences far more costly than the investment needed to monitor and analyze the production system.
What is the most common mistake you see in companies when they try to introduce smart maintenance or advanced asset management?
Smart maintenance can be considered a component of the smart factory. If the intelligent factory aims to make production processes more connected, adaptive, and data-driven, then smart maintenance applies the same logic to the management of equipment, machines, and industrial assets. It is not just about the ability to predict a failure, but also about the way the organization plans, executes, supports, and improves maintenance activities.
The most common mistake I see in companies is thinking that smart maintenance equates to purchasing a technology—or worse, introducing a single algorithm. To put it provocatively: you cannot just buy machine learning to do smart maintenance. Machine learning can be useful, artificial intelligence can play an important role—even a disruptive one in innovation—but they alone do not make a maintenance process intelligent.
Truly smart maintenance arises from the integration of technologies, processes, competencies, and organization. Technologies are certainly needed: sensors, monitoring systems, digital platforms, analytics tools, predictive models, operator support solutions. But the word intelligent should not be understood only in an algorithmic sense. Maintenance is also intelligent when it helps people make better decisions, perform interventions more effectively, reduce errors, coordinate more precisely, and better understand the real behavior of equipment.
There is also a decisive organizational aspect. Maintenance cannot be viewed as an isolated function. The degradation of a machine, for example, often depends on how the equipment is operated in service—on production loads, operating conditions, production choices, and sometimes even on material quality or production process planning. For this reason, effective smart maintenance requires collaboration among maintenance, production, quality, engineering, IT, and management. The term asset management is often used ambiguously, almost as a synonym for maintenance management, but in reality it means managing assets with both an operational and a strategic perspective—considering their role over time, the value they generate, the associated risks, and the decisions to be made throughout their entire life cycle. In this sense, the manufacturing world still has much to learn from other sectors, particularly from infrastructure, where asset management has long been a central discipline.
In the real industrial world, do physics-based approaches, data-driven models, or hybrid solutions work best? And how do you find a balance between accuracy and interpretability?
This is a complex question because it is not just about the distinction between physics-based models, data-driven approaches, and hybrid solutions. It is also about how a factory evolves toward a more intelligent, adaptive dimension, capable of making better decisions.
In a smart factory, we need to get used to the idea that models become central tools for supporting decision-making. But these models can have different natures and can operate at different scales. The factory is not a single, homogeneous system. It can be observed at the level of individual components, machines, production lines, departments, plants, and—in a broader perspective—the supply chain. Each level presents different problems, different data, different constraints, and different decision-making needs. For this reason, I do not believe there is a universally valid answer.
That said, if we look at the real industrial world as a whole, I tend to consider hybrid solutions the most promising precisely because they allow different strengths to be combined.
Physics-based models have very significant value, especially when we get close to the physical behavior of the machinery, component, or process performed by the assets. These models are built on knowledge of the physical principles governing a phenomenon and can offer a high level of interpretability. An expert can read the model, understand its logic, discuss its assumptions, and connect its results to the real behavior of the asset. In some cases, physics-based models can also be highly accurate, especially when the phenomenon is well understood and can be modeled.
The limitation is that these models require specialized expertise, deep domain knowledge, and often laborious construction, validation, and calibration. They are not always easily scalable, do not always succeed in representing the real complexity of the industrial environment, and can become difficult to maintain when operating conditions change.
Data-driven approaches, on the other hand, offer potentially greater flexibility. They can adapt to different contexts and capture relationships that are not always easily formalized through a physical model. From this point of view, data-driven models offer great potential, especially when the company has a broad, reliable, and well-structured data foundation. Here too, however, we should not oversimplify: building an effective data-driven model requires time, quality data, training and validation phases, careful algorithm selection, and adequate analytical skills. The risk with pure data-driven approaches is losing interpretability. A model may be accurate in producing a prediction but less clear in explaining why it is reaching that conclusion.
Today there is much talk of digital twins, but often the term is used vaguely. When can a digital twin be called operational rather than merely simulative?
This is a very important question because the term digital twin is often used generically—in some cases, it is an overused term. It is a topic I am pleased to discuss; I have been following it for about ten years, starting from experiences gained in European projects, and I believe it is useful to help clarify the meaning of digital twin, especially in the industrial context.
The digital twin can be applied to many domains. Even within manufacturing alone, it can refer to the product, the production system, or the services linked to the product itself. First, it should be said that there is not yet a fully shared definition between research and industry. This is one of the reasons why the concept can seem vague. However, there is a point I consider central: the digital twin is a combination of technologies. Within a digital twin, simulation, physics-based models, data-driven approaches, artificial intelligence, analytics, and geometric models can coexist. At the same time, the digital twin interfaces with both data collection systems and decision optimization tools.
The ultimate conceptual purpose of the digital twin is to build a digital representation of the physical system capable of reproducing its structure and behavior over time. When this model is connected to the real system, the digital twin can perform advanced functions. It can monitor what is happening in the production process, verify any deviations from expected behavior, predict future evolution within a certain time horizon, and support management and optimization decisions. These functionalities distinguish it from traditional systems present in the company, such as ERP, MES, or maintenance management systems.
In this context, the difference between an operational digital twin and a purely simulative model is precisely linked to the nature of the digital twin as a tool for supporting decisions. A simulative model works in a virtual environment, disconnected from the real system. It can be very useful in the design phase, scenario analysis, or preliminary studies, but if it does not receive field data and does not maintain a dynamic relationship with the physical system, it is more accurate to call it a digital model or simulator, not a digital twin in the full sense. A digital twin becomes operational when it is connected to systems that collect and manage data from the field: sensors, MES, ERP, maintenance management systems, control systems, monitoring platforms, and other information sources.
That said, this development is still in progress; further industrial experience is, in my view, necessary to shape the definition of digital twin. To make my contribution, in addition to the usual scientific journal publications and industrial projects, I am finalizing a book on digital twins of manufacturing systems that will be published soon.
Is there a minimum level of digital maturity below which talking about predictive maintenance or digital twins risks being premature?
Yes, there is certainly a minimum level of digital maturity below which talking about predictive maintenance or digital twins risks being premature. Not because these concepts are inherently inaccessible, but because they require certain baseline conditions without which there is a risk of building ambitious initiatives on still-fragile foundations.
In my experience, digital maturity should be assessed by looking at business processes as a whole. It is not enough to ask whether the company has introduced some digital technology; you need to understand how the various business areas function, what digital support they have, and how integrated they are with one another. This applies to engineering, to operations—including production, maintenance, and quality—and, more broadly, to all the processes that contribute to the real functioning of the company.
When conducting a maturity assessment, interesting situations emerge. In some cases, the organization is formally structured, roles are defined, and processes exist, but decisive transformative elements are still missing. One of these concerns people’s competencies. The transition to data-driven models requires that people are accustomed to using digital tools, reading data, interpreting it correctly, treating it methodically, and integrating it into decision-making processes.
This is a central point: digital maturity does not depend only on the platforms installed, but also on the organization’s ability to use them consistently. If data is collected but not governed, if it is available but not integrated, if it is produced but not turned into decisions, then the foundation for doing predictive maintenance or digital twins remains incomplete.
Specifically, a digital twin, by its nature, is more complex than a single technological innovation. It requires a robust digital infrastructure, good data availability, systems capable of communicating with one another, and processes mature enough to feed the model and use its outputs.
The same applies to predictive maintenance. You cannot expect to introduce it effectively if you lack reliable historical data, adequate sensors, traceability of interventions, information on asset operating conditions, and maintenance that is already sufficiently organized. Without these elements, the risk is talking about predictive maintenance when the company has not yet even consolidated the basics of good preventive or condition-based maintenance. For this reason, it is important to start with an assessment of maturity gaps across the different processes.
Is there difficulty in Italy in finding personnel with the qualifications needed to support the new industrial paradigms?
Yes, there is a real difficulty in finding personnel with the qualifications needed to support the new industrial paradigms. It is an issue that is particularly evident in Italy, because the digital transformation of industry—driven in recent years by policies linked to Industria 4.0—has generated a demand for competencies that is much broader and more varied than in the past.
The point is that purely technical profiles are not the only need. Companies need people who can work on automation, sensors, data, information systems, artificial intelligence, cybersecurity, advanced maintenance, and digital twins. But they also need people capable of connecting these competencies to how the factory actually operates. Knowing the technology is not enough: you have to understand the production process, maintenance logic, quality, work organization, and the economic implications of industrial decisions.
One of the main challenges concerns precisely this integration capability. The new industrial paradigms require hybrid competencies that no longer belong to a single discipline. You need people who can read data, but also those who can interpret it in the operational context. You need people who know algorithms, but also those who understand equipment behavior. You need people who can manage digital platforms, but also those who can engage with maintenance, production, IT, quality, and management.
In Italy, the problem is not only quantitative—that is, a lack of available people. It is also qualitative. Companies are looking for profiles with a certain level of applied maturity, but these competencies are built through education, hands-on experience, and cross-fertilization among universities, research centers, companies, and technology providers.
One final consideration: while the specific context may vary, this is not an experience confined to Italy alone—it applies to industry in transformation within global production networks.
