A recent research paper highlights how machine learning can optimize the additive manufacturing of an advanced aluminum alloy heat exchanger. Authored by NSIRC and University of Birmingham PhD student Francesco Careri, the study describes a Powder Bed Fusion-Laser Beam on Metal (PBF-LB/M) process guided by machine learning. The approach tackled metallurgical defects, dimensional accuracy, and surface quality by using neural networks to identify ideal parameters.

Initial tests generated a process map for the A205 aluminum alloy, investigating how variations in laser parameters, beam compensation, and contour distance affect manufacturing outcomes. The optimization phase reduced thickness deviations in thin-walled lattice structures to under 2%, enabling successful fabrication of a prototype heat exchanger. This achievement suggests a feasible path toward producing complex heat exchangers with minimal porosity.

The study, co-authored with Dr. Leonardo Stella and Professor Moataz Attallah of the University of Birmingham, illustrates how integrating machine learning with PBF-LB/M supports the creation of intricate, defect-free parts. TWI’s mentorship and collaboration also underscore ongoing efforts to develop new solutions and train the next generation of specialists in advanced manufacturing technologies.

Application of machine learning in additive manufacturing of a novel Al alloy heat exchanger - The International Journal of Advanced Manufacturing Technology
Additive manufacturing (AM) of complex geometries faces limitations in the dimensional and geometrical accuracy, especially when the geometries are characterised by thin features designed to tailor the mechanical and thermal properties of novel heat exchangers (HXs). In this work, a novel, complex, thin hollow-walled lattice compact HX was fabricated using the Powder Bed Fusion-Laser Beam on Metal (PBF-LB/M) process. Given the intricate relationships between process parameters and complex design, machine learning (ML) methods were utilised to optimise the manufacturing workflow. Although new ML models would be required for different cases to ensure optimal performance, the flexibility of such approaches allows for recalibration and re-optimisation whenever there are changes to material properties, geometry, or manufacturing settings. A process map for the A205 Aluminium alloy was generated, investigating metallurgical defects and surface quality. Optimal process parameters for defect-free materials were estimated using a neural network (NN). Further, optimisation evaluated the influence of laser parameters, beam compensation (BC), and contour distance (CD), on the geometrical and dimensional accuracy of thin features, with a second NN predicting optimal BC and CD. Thickness deviations in hollow lattices were reduced to under 2%. A prototype of the novel HX using optimised parameters was successfully fabricated and characterised to evaluate manufacturing feasibility. The analysis of pores in thin features, potentially leading to leakage and part failure, was carried out through SEM analysis. While PBF-LB/M is well-established for HXs, this study demonstrates its capability for manufacturing highly complex, thin-walled designs when guided by ML-based optimisation. Graphical Abstract