Metal based additive manufacturing techniques such as laser powder bed fusion (LPBF) can produce parts with complex designs as compared to traditional manufacturing. The quality is affected by defects such as porosity or lack of fusion that can be reduced by online control of manufacturing parameters. The conventional way of testing is time consuming and does not allow the process parameters to be linked to the mechanical properties. In this paper, ultrasound data along with supervised learning is used to estimate the manufacturing parameters of 316L steel cubes. Nine cubes with varying manufacturing parameters (speed, hatch distance and power) are examined with ultrasound using focused transducers. The volumetric energy density (VED) is calculated from the process parameters for each cube. The ultrasound scans are performed in a dense grid in the built and transverse direction. The ultrasound data is used in partial least square regression algorithm by labelling the data with speed, hatch distance and power and then by labelling the same data with the VED. These models are computed for both measurement directions and as the samples are anisotropic, we see different behaviours of estimation in each direction. The model is then validated with an unknown set from the same 9 cubes. The manufacturing parameters are estimated and validated with a good accuracy making way for online process control.

Metal based additive manufacturing techniques such as laser powder bed fusion can produce parts with complex designs as compared to traditional manufacturing. The quality is affected by defects such as porosity or lack of fusion that can be reduced by online control of manufacturing parameters. The conventional way of testing is time consuming and does not allow the process parameters to be linked to the mechanical properties. In this paper, ultrasound data along with supervised learning is used to estimate the manufacturing parameters of 316L steel samples. The steel samples are manufactured with varying process parameters (speed, hatch distance and power) in two batches that are placed at different locations on the build plate. These samples are examined with ultrasound using a focused transducer. The ultrasound scans are performed in a dense grid in the build and transverse direction, respectively. Part of the ultrasound data are used to train a partial least squares regression algorithm by labelling the data with the corresponding manufacturing parameters (speed, hatch distance and power, and build plate location). The remaining data are used for testing of the resulting model. To assess the uncertainty of the method, a Monte-Carlo simulation approach is adopted, providing a confidence interval for the predicted manufacturing parameters. The analysis is performed in both the build and transverse direction. Since the material is anisotropic, results show that there are differences, but that the manufacturing parameters has an effect of the material microstructure in both directions.