Artificial intelligence in the foundry: real opportunities, concrete limits.
The foundry is often perceived as a sector anchored in tradition, impermeable to new concepts such as artificial intelligence. A sector so consolidated and difficult to change: high temperatures, molten metal, rough surfaces, operators with little expertise or who refine their technique over years of field work. Yet it is precisely in this environment (apparently distant from any digital innovation) that artificial intelligence can generate one of the most concrete impacts in the manufacturing industry today.
In short: artificial intelligence in the foundry has remarkable prospects. This is not science fiction or technological marketing. We are talking about tools that, applied in the right way, make processes more controllable, repeatable, and efficient.
In this article, we delve into some important topics:
Artificial intelligence in the foundry: the role of experience.
TREBI has been working for over 40 years in robotic automation specifically for the foundry, and observes every day how the production needs of its customers change. This article stems from this meticulous daily observation.
Why we talk about artificial intelligence in the foundry today
As with any new technology entering the production field, it is necessary to refer to the main problems that such innovation aims to solve. From the perspective of increasing productivity and return on investment, the improvement of efficiency and effectiveness parameters guides evaluations. A useful insight can be found in this other article where we talk about artificial intelligence and industrial robotics.
It is worth identifying some themes that we can define as structural, which certainly provide concrete and easily recognizable reasons for those operating in the sector to support the introduction of artificial intelligence in the foundry.
The lack of specialized operators is a structural problem, not a cyclical one. According to Excelsior data from Unioncamere processed by CGIA, in 2024 Italian companies sought nearly 840,000 skilled workers — equal to 15% of total hires — but over 60% reported difficulties in finding them, with search times exceeding five months. In manufacturing, the situation is even more critical: the difficulty in finding technical profiles rose from 45% in 2023 to 48% in 2024, with an estimated peak of 70% for technical profiles by April 2026.
Specific data on the foundry sector confirms the picture. According to Assofond, in the first half of 2024, the majority of Italian foundries (73.3%) sought new personnel despite a stagnant market: the most sought-after profiles were skilled workers in 52.9% of cases, but only 13.6% of companies declared themselves fully satisfied with the search results.
Added to this is the pressure on energy costs: as highlighted by Assofond president Fabio Zanardi, Italian foundries find themselves in a situation of strong competitive disadvantage compared to international competitors who benefit from significantly lower energy costs, making structural interventions at the European level urgent.
To compensate for all this, foundries must look at new ways of doing business. Innovation is certainly a driver that can lead the way out of this situation. And today, innovation must be artificial intelligence, digital twins, IoT, and remote monitoring. The application of real-time data allows for the monitoring of critical machinery, such as melting or finishing lines, preventing breakdowns and continuously optimizing operations. Artificial intelligence in the foundry provides a new way of looking at and controlling production.
Main applications of artificial intelligence in the foundry
Intelligent quality control
AI-based vision systems allow for the automatic recognition of surface defects, classifying parts as compliant or non-compliant and reducing the subjectivity of manual inspection. The numbers are significant: AI computer vision systems applied to metal inspection have improved defect detection accuracy from 60% for human inspection to over 98%, outperforming traditional systems and significantly reducing time and costs.
On the research front, a review published in MDPI Applied Sciences analyzed 37 studies conducted over the last 15 years on the application of AI in the foundry — covering sand casting, die casting, continuous casting, and investment casting — confirming a broad and already documented application potential.
TREBI has developed and installed artificial intelligence vision systems integrated into real industrial processes, including the “Check by Flight” system for in-line quality control. These are not demonstration projects: they are solutions embedded in the production cycle, which must work on real parts, with real defects, in real industrial environments. The value lies in the integration between AI, robotics, tools, and process logic. An algorithm is not enough: you need to know what you are looking at and why.
There is also an even more important aspect: the best operators in the foundry carry with them a wealth of knowledge that is often not written down anywhere. They know how that alloy behaves in that condition, how to modify the process to obtain the highest quality with the shortest cycle time, and how to perform deburring on the part correctly without unnecessarily wearing out the tools. Artificial intelligence can become the tool to digitize this experience, making it available to the robotic system in a structured and replicable way.
TREBI does not start from the algorithm: it starts from the real production process. It is this application knowledge, matured over decades of direct processing, that allows artificial intelligence to be used in a useful and not just demonstrative way. We are working on a new AI-based system precisely in this direction: using previous process experience to enhance the performance of the machine and the operators.
Intelligent support for operators
Artificial intelligence in the foundry does not necessarily have to replace the operator. In many cases, its greatest value lies in supporting them. Intelligent on-board assistance systems can suggest processing parameters, guide tool selection, support real-time problem management, and allow for quick consultation of technical documentation and operating procedures.
Artificial intelligence in the foundry must take the operator’s experience and multiply it to make the work faster and more efficient. Small production batches should not be a problem, but a competitive advantage. This is one of the benefits that the integration of artificial intelligence in the foundry brings with it. One facet of this theme is represented by the increasingly important introduction of cobots.
According to the Confindustria labor survey of 2025, only 43.7% of Italian companies adopting AI have already taken concrete actions regarding human resources. The main criticalities are the lack of internal skills (36.7%), the technical complexity of integration, and still high costs. A technical knowledge base integrated into the systems can concretely address this difficulty, preserving and transferring know-how even in the absence of expert senior personnel. This is the clear and convinced vision of TREBI.
Intelligent support for mold design and casting simulation
Talking about artificial intelligence in the foundry also means referring to the equipment preparation phase. In particular, mold design and the definition of the casting system are phases where an error is costly: a poorly sized gating system, an insufficient cooling zone, or a poorly positioned gate result in serial defects, scrap, and significant rectification costs. For decades, the only tool to reduce this risk was the designer’s experience. Today, simulation is added, and artificial intelligence is built upon it.
Casting simulation — with software such as Magmasoft, ProCAST, or Flow-3D — has allowed for years the virtual prediction of mold filling, metal solidification, and areas at risk of porosity and shrinkage.
The path of introducing artificial intelligence in the foundry takes a further step. PhysicsX has developed and installed an AI system for an aerospace manufacturer that combines real-time defect prediction with casting system optimization: the models, trained on process data and mold geometries, operate over 100 times faster than traditional simulation methods, making information available that was previously out of reach during ongoing production. Within a few weeks of installation, the manufacturer achieved a reduction in scrap rate, a decrease in operating costs, and an improvement in the qualitative consistency of components. ModerncastingAkridata
The trajectory toward which the sector is moving is that of the intelligent mold: molds equipped with sensors, control systems, and integrated actuators, capable of incorporating embedded simulation, online inspection, and automatic real-time decisions. This is not science fiction: it is the direction that the main research centers studying how to introduce artificial intelligence in the foundry indicate as the next application frontier. MDPI
The prerequisite for the successful introduction of artificial intelligence in the foundry, we emphasize strongly, remains data quality. Process variability, which is linked to raw material quality, metal bath treatment, and mold conditions, must be incorporated into simulation models to produce reliable predictions. An AI system trained on generic data produces generic predictions. Specific process knowledge remains the discriminating factor.
Predictive maintenance
Continuous monitoring of vibrations, power consumption, temperatures, and process anomalies can anticipate failures and reduce unplanned downtime. That said, it is necessary to be honest: predictive maintenance works well when there is sufficient historical data, correctly installed sensors, and a system calibrated for that specific machine in that production context. It is not a plug-and-play solution.
In many cases, the best approach remains designing systems that are extremely reliable and easy to maintain. TREBI has always focused on this: machines designed for operational continuity, with uptime exceeding 98%. Predictive maintenance is a useful tool, but it does not replace the construction quality of the system.
We are also working on a structured data system designed to integrate easily into a factory AI system. Because the best predictive maintenance software is useless if the machines do not share data in an organized way. Introducing artificial intelligence in the foundry means designing an integrated system.
Further insights into TREBI’s approach regarding predictive maintenance are available in this article.
Intelligent analysis of production data
An artificial intelligence system in the foundry is trained on historical data to identify the relationship between process parameters, such as metal temperature, casting times, injection speed, and qualitative outputs. However, this only works if all production data is collected in an integrated system.
A concrete case comes from the aluminum wheel sector: the RONAL group combined data from low-pressure die casting with data from automatic X-ray inspection to train a neural network capable of predicting anomalies in castings even before scrap parts are produced, reducing the scrap rate with a direct impact on costs and sustainability.
TREBI has developed a structured data collection architecture that allows for the orderly and timely analysis of production: through the data read from the machine, it is possible to understand the variation of outputs as inputs vary.
From Business Intelligence to Process Intelligence: when data becomes decisions
Talking about artificial intelligence in the foundry without talking about Business Intelligence would be like installing a high-performance engine in a vehicle without a dashboard. The data is there. The problem, almost always, is that it is not collected, organized, and read in the right way.
Traditional Business Intelligence — monthly reports, aggregated Excel sheets, KPIs (Key Performance Indicators) read at the end of the shift — represented an important step. But classic BI, by definition, arrives late. It captures the past; it does not govern the future.
The natural evolution is AI-powered BI: the integration between Business Intelligence logic — structured data collection, measurable KPIs, readable dashboards — and the AI’s ability to analyze data in real time, identify hidden patterns, and suggest corrective actions before the problem manifests in the finished part, also integrating external data for realistic future projections. In summary: BI speaks of the past, AI-powered BI speaks of the future.
In this case, structured data collection is a central theme. If the data is not organized and timely, any analysis is false.
TREBI addressed this problem by building a structured data collection architecture directly on board the machine. Each system produces readable, organized, and contextualized data: not a raw stream of numerical values, but information linked to the part, the cycle, the tool, and the operator. This structure is the necessary condition for any advanced analysis system — whether it is called BI, AI, or AIpBI — and represents the true competitive advantage in the long term.
The development of artificial intelligence in the foundry is fully part of our technological paths. The attention we have dedicated to computer vision further testifies to our commitment to exploring and applying every aspect related to the introduction of artificial intelligence in the foundry.
Humanoid robots in the foundry: current scenario and real prospects
Recently, the conversation has not only been about how to integrate artificial intelligence in the foundry. In the last two years, the topic of humanoid robots has moved from a conference subject to daily news. Tesla Optimus, Figure 02, Boston Dynamics Atlas, Unitree G1: the names are multiplying, video demonstrations are going viral, and investments are growing exponentially. It is therefore worth asking: what do humanoids have to do with the foundry? How do humanoids integrate into the general development process of artificial intelligence in the foundry? And above all, where do we really stand with research and specifically with development?
The starting logic is simple. Unlike traditional automation solutions, which require a redesign of production spaces, a humanoid robot can be operational in minimal time and with minimal changes to the existing environment. In theory, therefore, it could be introduced into an existing plant — including a foundry — without having to rethink layouts, workbenches, or equipment.
In 2024, companies like Mercedes-Benz and BMW announced plans to integrate humanoid robots into their plants. BMW in particular has already started a concrete experience: at the beginning of 2025, a fleet of Figure 02 robots began working full-time at the Spartanburg plant, with the second generation of the robot performing industrial tasks at four times the speed and with seven times greater precision than the initial trial.
Impressive numbers. But it is also necessary to read the other side of the coin.
Current limitations are concrete. Regarding energy autonomy, real data is significant: Agility Digit operates in 30-minute intervals, Figure 02 reaches about 2-3 hours, Tesla Optimus Gen 2 estimates 4-5 hours. In a standard 8-hour shift, current figures do not stand up to comparison. On the economic front, a fully functional humanoid today costs between $100,000 and $150,000, with higher maintenance costs compared to traditional industrial robots. On the regulatory front, the framework is still under construction: the first specific international standard for humanoid safety, ISO 25785-112, was published only in May 2025.
There is also a technical aspect that specifically concerns the foundry. The combination of limited autonomy, low weight-to-payload ratios, and poor resistance in harsh environments means that current humanoid systems remain primarily research prototypes. The foundry is by definition a harsh environment: radiant heat, metal dust, vibrations, irregular surfaces. Traditional industrial robots are designed and certified for these conditions with specific coatings and protections. Humanoids, born for clean logistics and assembly environments, do not yet guarantee this reliability.
For fixed, high-volume processes — welding, molding, deburring, die casting — traditional 6-axis industrial robots remain the appropriate tool. Their repeatability, cycle times, and reliability records are not matched today by any humanoid platform. Humanoids instead find a more interesting logic in support operations: material handling between departments, machine feeding in flexible layouts, and auxiliary activities that currently require an operator but not decimal precision.
The overall picture is this: humanoids are a real technology, in rapid development, with application potential even in heavy manufacturing. But today they replace neither specialized industrial robots nor the expert operator who knows the process.
TREBI follows this evolution closely: when humanoids reach an industrial maturity suitable for the foundry environment — in terms of robustness, autonomy, and cost — we will already know where and how to integrate them.
The real limits of artificial intelligence in the foundry
This is the chapter that many discussions on the subject omit. We consider it the most important.
Artificial intelligence in the foundry is still a relatively new field, with very few operators actually applying it. An AI system can analyze data sets much more complex than any human could handle, but it only works if all production data is collected in an integrated way.
Every industrial process is different: a system that works perfectly on a deburring line for gray iron may not work the same way on a line for die-cast aluminum. Customization is not an option; it is a technical necessity. And results depend critically on the quality of integration: an excellent algorithm poorly integrated into a poorly designed system produces little value.
The market is full of companies selling “artificial intelligence in the foundry” without truly knowing the production process they are supposed to optimize. The risk for foundries is investing in spectacular demonstrations that do not stand the test of daily production.
Value does not come from the algorithm. It comes from knowledge of the production process. The algorithm is a tool. Competence is the prerequisite.
In other words, AI enhances knowledge and makes it more fluid. But this knowledge must exist within the company. AI makes it accessible to everyone, makes data more readable, and interpolates it with the rest of the world, but the company must have the data and must be able to produce it truly and continuously every day.
Industrial AI: cloud or proprietary systems?
This is another much-discussed topic.
The problem is how much do I invest in hardware and how many resources will the technology need? Is the cloud better so I can remove the problem of scalability and the potential change of necessary hardware platforms?
The other issue is privacy: if I put data in the cloud, is it really secure?
Compared to traditional methods, AI models operate over 100 times faster, enabling analyses that were previously out of reach during ongoing production. But processing speed is not the only variable to consider.
Cloud platforms offer computational power and continuous updates, but involve transmitting production data externally. In a sector where process parameters, casting geometries, and technical solutions developed over time constitute a real competitive advantage (as well as often a trade secret), this transfer of data is not a negligible detail.
The alternative is artificial intelligence integrated directly into the systems, with local data processing (edge computing) and systems controlled internally by the company. This approach reduces dependence on external platforms, cuts latency in real-time decisions, and protects production know-how.
However, edge computing has the major limitation of performance and learning speed. Therefore, it cannot be the solution for every problem. If complex problems using enormous models and neural networks need to be solved, an edge computing system cannot be used.
Edge computing systems are suitable for relatively simple functions (Machine Learning, chatbots…)
TREBI works in both directions, following the customer’s needs and finding the best solution for each specific case. According to our philosophy and experience, introducing artificial intelligence in the foundry does not mean marrying a fixed and limiting paradigm.
There is no single solution to this dilemma. It depends on expectations, the resources that can be invested, and the confidentiality we want (or must) maintain.
How to truly introduce artificial intelligence in the foundry
The most concrete recommendation: start from a real and measurable problem, not from a technology to be demonstrated. Here are other concrete steps to take if you are thinking of introducing or better developing artificial intelligence in the foundry.
Identify a critical process, an operation where variability is high or where manual checks are costly and unreliable.
Define a clear indicator of improvement: scrap reduction, productivity increase, downtime reduction. Start a limited pilot project, measure the real ROI, then scale.
Involving operators and production managers from the beginning is not a recommendation of principle: it is a technical necessity. Those who know the process know where the algorithm goes wrong and why.
Better a system that truly reduces scrap by 10% than a spectacular demo that is useless in production.
TREBI approaches every project with this method: listening to the customer, custom design, preventive simulation of benefits, real integration into production flows. We do not sell technology. We design solutions.
Our necessary conclusions
Artificial intelligence in the foundry represents a great opportunity, but only if applied with technical competence and real knowledge of the processes. Technology alone is not enough. Never.
The real challenge will not be replacing people. It will be transforming the experience of the best people into a shared, replicable, and protected asset over time.
TREBI sees artificial intelligence in the foundry not as a technological fad, but as a concrete evolution of industrial automation: intelligent tools designed to help people, processes, and companies work better. It is what we have been doing for more than forty years, with tools that change but always with the same approach. This is why we believe we are the most suitable technological partner to introduce or develop artificial intelligence in the foundry.
