August 2025 Volume 7

FORGING RESEARCH

digital twin compares the data (e.g., load-stroke from the press, 3D part geometry and 3D surface temperature) to predicted values, and, if they match within certain error, the autonomous incremental forging process continues until the part is formed to desired tolerances (Figure 14). If there is a deviation between any measured and the predicted values, a real-time adjustment is made by the ML algorithm to compensate for any difference. If the system identifies that these real-time adjustments will result in a good part, the process continues. If it does not make this judgment, then the process is halted and the forging team strives to update our models (e.g., material behavior and process conditions) so that the next forging will match prediction. Whether the forging is successful or unsuccessful, the ML algorithms are updated, and each time we make a part, the system gets smarter.

Figure 12: Microstructure variation as a function of processing across three parts formed to the same geometry via different process paths.

Figure 14: Example of digital twin (left) to provide pre-model testing, real-time monitoring, and workforce training of real Agility Forge (right). Taken together, these innovations not only enable autonomous production of forged parts that meet traditional performance standards but also give the system the capability to exceed the properties achievable through conventional forging methods. Ultimately, this technology opens the door to a new era of forging - one where high-value, low-volume components can be produced economically across a wide range of industries, from automotive and aerospace to medical and beyond. Acknowledgement The authors would like to acknowledge support from the U.S. National Science Foundation Engineering Research Center for Hybrid Autonomous Manufacturing Moving from Evolution to Revolution (ERC‐HAMMER) under Award Number EEC-2133630.

Figure 13: Example of prediction of microstructure evolution as a function of 3 different Agility Forge toolpaths. We used the DEFORM FEM code to simulate the autonomous incremental forming paths and to perform microstructure evolution modeling. DEFORM has an Application Programming Interface (API) which allows us to setup a DEFORM simulation, run it, automatically export various variables of interest (e.g., workpiece geometry and predicted microstructure) and then automatically edit simulation parameters (e.g., number of die blows, depth of die blows, speed of die blows, temperature at deformation and die geometry). We can also intelligently select these new simulation parameters via our own optimization algorithm which plots the results of different simulations and utilizes Machine Learning (ML) to choose the most efficient leaps in variable values to achieve our targets (part geometry, part performance). In this way, we setup a simulation “manually” one time and then let the ML algorithm fine tune the parameters automatically. Once the parameters for the autonomous incremental forging toolpath are selected, they are converted to G-code and the Agility Forge forges the part. It executes the heating, die selection via the automatic tool changer in the Agility Forge, forging, reheating, part translating, part rotating and more. During this process of autonomous incremental forging, the Agility Forge sensors (e.g., load-stroke data from the press, 2D thermal imaging data from cameras and 2D geometry data from a laser profilometer) reconstruct the real-time 3D geometry of the component, map its real-time 3D temperature to the surface and send all this information to a digital twin. In real time, the

FIA MAGAZINE | AUGUST 2025 73