Machine Learning and Digital Twin in Directed Energy Deposition: A Systematic Literature Review

Imagine a factory floor where massive metal parts are not carved out of a solid block, but built up, layer by layer, from a stream of molten metal. This is directed energy deposition, a high-tech manufacturing method used to create everything from jet engine components to medical implants. It is powerful and flexible, capable of repairing expensive machinery or printing complex shapes that would be impossible to make any other way. However, the process is notoriously difficult to control. The metal is melted by intense heat sources like lasers or electric arcs, creating a tiny, searing pool of liquid that cools and solidifies in a fraction of a second. Because this happens so fast and involves such extreme temperatures, the process is unstable. It can easily develop hidden cracks, holes, or weak spots that ruin the final part. Traditionally, fixing these issues has meant a slow, expensive cycle of trial and error: set the machine, build a part, cut it open to inspect it, find the flaw, change the settings, and try again.

To break this cycle, researchers are turning to two powerful tools: machine learning and digital twins. Machine learning is a type of computer intelligence that learns to recognize patterns from vast amounts of data, much like a human learns to spot a defect by looking at thousands of examples. A digital twin is a virtual copy of the physical machine and the part being built. It is not just a static 3D model; it is a living simulation that receives real-time data from sensors watching the actual machine, allowing it to mirror the physical world as it happens. When these two technologies are combined, they promise to transform manufacturing from a reactive process, where defects are found after they happen, into a proactive one, where problems are predicted and corrected before the metal even cools.

A team of researchers at The University of Texas Rio Grande Valley recently set out to map the current state of this technology. They conducted a systematic review of 130 scientific studies published between 2016 and 2026, analyzing how machine learning and digital twins are currently being used in directed energy deposition. Their work, published in a systematic literature review, reveals a field that is advancing rapidly but still sits on the edge of a major transition. The researchers found that while scientists have become very good at using cameras and sensors to watch the process and predict when something might go wrong, they have not yet fully mastered the ability to automatically fix those problems in real time.

The review highlights that the most common materials being studied are specific types of steel and nickel-based superalloys, which are essential for aerospace and energy industries. The most frequently used manufacturing methods are wire arc deposition, which uses an electric arc to melt metal wire, and laser-based deposition, which uses a high-powered laser to melt metal powder. In these studies, the primary focus has been on "seeing" the process. Researchers have equipped machines with high-speed cameras and thermal sensors to capture images of the molten pool. They then feed this visual data into machine learning models, often using complex neural networks that act like digital eyes. These models have shown remarkable success in identifying defects. Some studies reported that these systems could detect cracks or holes with accuracy rates exceeding 99 percent, or predict the shape of the metal bead with high precision.

However, the authors point out a critical gap between seeing a problem and solving it. Most of the digital twin systems reviewed so far function more like a "digital shadow." In this setup, data flows one way: from the physical machine to the virtual model. The virtual model can tell the operators, "Look, a defect is forming," or "The temperature is too high." But it rarely takes action on its own. The connection between the virtual prediction and the physical machine is often broken. Instead of the computer automatically adjusting the laser power or the speed of the robot arm to correct the mistake, the system usually stops and waits for a human operator to intervene. The researchers note that while a few advanced studies have demonstrated closed-loop control—where the system automatically adjusts the process to fix an error—these are the exception, not the rule.

The review also identifies significant hurdles that prevent these technologies from being used widely in industry. One major issue is the lack of data. Training a machine learning model requires thousands of examples of both good and bad parts. In directed energy deposition, creating these examples is expensive and time-consuming. It often requires destroying the finished part to inspect it for internal flaws, which means researchers cannot easily gather the massive datasets needed to teach the computers. To cope with this, some researchers are using techniques like transfer learning, where a model trained on one type of metal is adapted to work on another, or they are using physics-based simulations to generate virtual data. Yet, the authors caution that these methods are not perfect. A model trained on one machine or one material often struggles to work on a different setup, limiting its usefulness in a diverse factory environment.

Furthermore, the researchers found that the current digital twins are often too simple to handle the complexity of real-world manufacturing. A true digital twin should be able to communicate back and forth with the physical machine, constantly updating its own understanding based on new data and using that knowledge to guide the machine's actions. Currently, many systems are fragmented. They might be excellent at monitoring temperature or predicting geometry, but they lack the ability to integrate all this information into a single, autonomous decision-making process. The review suggests that for these technologies to reach their full potential, the field needs to move beyond isolated experiments. Future research must focus on creating standardized systems that can handle data from multiple sensors at once, operate in real-time without lag, and, most importantly, close the loop between prediction and action.

The authors conclude that while the foundation has been laid, the house is not yet built. The ability to predict defects is well-established, but the ability to autonomously correct them remains incomplete. The next step for the industry is to develop robust, closed-loop systems where the machine does not just watch itself, but actively steers its own course. This requires not only better algorithms but also a deeper integration of physical laws into the computer models and a commitment to sharing data across different machines and materials. Until then, directed energy deposition will remain a technology of great promise, held back by the need to bridge the gap between watching the process and truly controlling it.