Physics-Guided Neural Network for Regularization and Learning Unbalanced Data Sets

Directed energy deposition is of interest to the aerospace and defense industries for the production of novel and complex geometries, as well as repair applications. However, variability during the build process can result in deviations in final component geometry, structure, and mechanical properties, which adds to the complexity of process planning and slows down adoption of this technology.

A representative image of the melt pool during the DED process. (Image: Army Research Laboratory)

Directed energy deposition is of interest to the aerospace and defense industries for the production of novel and complex geometries, as well as repair applications. However, variability during the build process can result in deviations in final component geometry, structure, and mechanical properties, which adds to the complexity of process planning and slows down adoption of this technology.

Directed energy deposition (DED) is a method of metal additive manufacturing (AM) by which parts are built layer by layer from 3-D computer-aided design models. In laser DED, powder is blown in front of a moving laser, which melts the powder, before it solidifies into part of the component being manufactured. AM provides a more direct pathway between the digital model and the manufactured part, which allows for more rapid design iterations and testing of parts. This further enables more advanced design methods and integrated features that would be operationally difficult by way of traditional subtractive manufacturing. In addition to being used to produce entire components, DED offers advantages through the ability to deposit additional material to existing components for the use of repair or cladding. Due to these advantages, DED is being adopted for specialized applications in defense and aerospace.

While there is great enthusiasm to use AM, process deviations are common, which can result in undesired final geometry, poor surface finish, or insufficient mechanical performance (Colosimo et al. 2018). Some of these drifts in the structure and property of metals fabricated by DED are not caused by stochastic process deviations, but by predictable trends. For example, Feenstra et al. (2020) observed changes in grain structure of 316L stainless steel with build height. Fine grains were observed in the AM structure near the build plate due to high cooling rates during building on a cold substrate. As the structure continued to build, heat buildup in the part created larger, columnar grains. At some point the grain structure was very predictable as the part reached a thermally, steady-state condition.

Thus, the deviation in microstructure and mechanical properties can be attributed to the local thermal history of the part. Observation of the melt pool during the DED process provides a means to measure an instantaneous point in the thermal history of a part. These observations can then be mapped in 3-D to the built geometry such as what was done in Kriczky et al. (2015). As a result, there is great interest in using sensors for in-situ certification of components, as well as responding to process deviations in real time, enhancing the performance of AM components, and increasing the yield of the AM process.

In this study, DED melt pool width was predicted as a function of input drivers. Physical trends in the DED melt pool width were identified in literature, predicted via modeling, and experimentally verified. These physical trends are then enforced on neural nets as constraints. Huang et al. (2019) developed models that predicted a melt pool width increase with each subsequent layer, and that as the wait time between layers increased, melt pool width decreased. Akbari and Kovacevic (2019) confirm experimentally that melt pool width only increases with increasing layer height. Liu et al. (2019) experimentally show that melt pool temperature increases with increasing laser power and—below the speed at which balling occurs— decreases with increasing laser speed. It could be argued that a hotter melt pool would be wider than a cooler one.

In this work, neural networks were trained, given process parameters and build plan information, to predict melt pool width during a DED build, given input drivers of speed, power, length, and height. Length and height are geometry dependent and may vary over the course of a build, and cause drift in melt pool width. In contrast, power and speed are process-specific and can be changed dynamically to adjust the melt pool width in response to drift caused by changes in local geometry.

This work was performed by Clara Mock, Christopher Rinderspacher, and Brandon McWilliams for the Army Research Laboratory. For more information, download the Technical Support Package (free white paper) under the Communications category. ARL-9655.



This Brief includes a Technical Support Package (TSP).
Document cover
Physics-Guided Neural Network for Regularization and Learning Unbalanced Data Sets

(reference ARL-9655) is currently available for download from the TSP library.

Don't have an account?