Physics-Informed Surrogate Modelling of Thermal Behaviour in Directed Energy Deposition

Additive manufacturing is a class of production processes that enables the creation of complex parts directly from digital models, typically by depositing material layer by layer. Despite its potential, the widespread adoption of metal additive manufacturing remains hindered by difficulties in guaranteeing the quality and consistency of the final product, owing to the complex physical phenomena involved.
 

A key factor governing the mechanical properties of a metal part is its thermal history during production. The microstructure of the material, and therefore properties such as strength, hardness and fatigue resistance, is shaped by temperature gradients experienced throughout the process. Critically, research has shown that confining temperature monitoring and control to the melt pool alone is insufficient to ensure uniform material properties across complex geometries, as after-heating effects between adjacent passes can alter the microstructure of the surrounding material. This highlights the need for spatially-distributed thermal models that describe the temperature evolution of the workpiece as a whole.
 

The focus of this project is to investigate and evaluate grey-box modelling approaches for the thermal behavior of additively manufactured parts, combining physical knowledge of the process with experimental data collected via infrared camera during manufacturing on the LMD² machine. Such approaches are attractive due to their ability to respect known physical constraints, while leveraging data to compensate for
unknown or hard-to-model phenomena, such as temperature-dependent material properties.
 

Rather than pursuing high-fidelity simulations, the goal is to develop compact surrogate models that are real-time capable and suitable for control-oriented purposes. The approaches under investigation include discretized dynamic models (where the governing physics is approximated through the finite difference method) and Scientific Machine Learning approaches, with a particular interest in Physics-Informed Neural Networks (PINNs) and Deep Operator Networks (DeepONets).

M.Sc. Massimiliano Pandolfo

January 2026 - November 2027

University of Kassel

• Guilherme da Fonseca Pereira, Massimiliano Pandolfo, Lars Sommerlade, Andreas Kroll: Time-Frequency Analysis for Temperature Measurement and Control in DED-LB/M processes, 2026 IEEE Conference on Control Technology and Applications (CCTA), accepted

• Massimiliano Pandolfo, Andreas Kroll: System Identification of Physics-Informed Control-Oriented Spatio-Temporal Temperature Models for Laser Beam Direct Energy Deposition, 60th Control Engineering Colloquium (2026), Boppard, DE
• Massimiliano Pandolfo: Hot Metal, Cool Control: The Importance of Sensor Technology, Modeling and Control in Additive Manufacturing, Werk - Stoff - Idee | Materialforschung aus der Uni Kassel (2026), Kassel, DE