Laser scan strategy descriptor for defect prognosis in metal additive manufacturing using neural networks

https://doi.org/10.1016/j.jmapro.2021.05.011Get rights and content

Abstract

In situ and post-print defects, including excessive residual stresses and poor microstructural properties, are important concerns in the part design and setup stages of powder bed fusion (PBF) manufactured parts. Laser scan strategies are well known to be correlated with the development of these defects; however, due to the lack of complex scan strategy descriptors and the consequential imposition of simple scan strategies, these correlations are not well understood and are difficult to investigate. This work proposes a methodology for an intuitive quantitative descriptor of scan strategies that has the potential to provide quick prognoses to predict defects. To demonstrate this, a neural network is trained to accurately predict post-print residual stress distributions using the descriptor of a specified path. It is envisioned that this descriptor could allow for the circumvention of current costly prevention mechanisms and increase confidence and reliability in metal additive manufacturing technologies.

Introduction

Additive manufacturing has established itself as an essential set of manufacturing technologies for the making of incredibly complex parts which is critical to contemporary design goals [[1], [2], [3], [4], [5], [6], [7], [8], [9]]. In particular, metal additive manufacturing (mAM) is defined as the group of manufacturing technologies with which metal parts are manufactured in a layer by layer fashion. Among the many different mAM technologies, powder bed fusion (PBF) is the most common in which a concentrated heat source, usually a laser or electron beam, melts thin spread powder in precise locations layer by layer. This fuses the powder together at local regions creating a 3D component. Following PBF in popularity is directed energy deposition, which also involves a moving heat source where material is transported to the point where the heat source is focused on. In this instance the part is created freeform and the heating and cooling history of a component strongly depend on the geometry of the component and the heat source path.

A shared challenge among AM techniques involving moving heat sources is the selection of a scan strategy. The selection of scan strategy involves the selection of a path and speed with which a heat source moves over the powder to consolidate a cross-sectional layer [10,11]. While melting a powder layer, underlying layers also heat up and usually re-melt. This leads to multiple local heat treatments which in turn can lead to various defects. The cooling rate of the consolidated layer is also a contributor to defect development [12]. Naively planning scan strategies can lead to critical defects in printed parts including high porosity, poor microstructural properties, bad surface quality, and excessive warping. It is non-trivial to consider all possible defect mechanisms when planning scan strategies, and worse still is the lack of reliably interpretable correlations between scan strategies and defect mechanisms. A lot of scan strategies deployed today are derived from trial-and-error approaches.

The literature contains numerous experimental studies empirically demonstrating the dependency of part properties to scan strategies. Using scanning electron microscopy imagery and image processing techniques, Arısoy et al. demonstrated that PBF process parameters including laser power, scan velocity, hatch distance, and scan strategy were correlated with cooling rates, thermal gradients, and microstructural properties [13]. Through x-ray computed tomography, microscopic imaging and other techniques, Rashid et al. demonstrated that there are correlations between scan strategies and the porosity, hardness, and microstructural properties of printed parts [14]. Levkulich et al. investigated the correlations between various process parameters and the residual stresses in titanium alloy PBF parts using x-ray diffraction, hole-drilling, and contour method techniques and concluded that there are measurable correlations [15]. Lee et al. used in situ near-infrared imaging and computational modeling to reveal dependencies between part geometry, scan strategies, cracking and residual stresses. Additionally, they proposed scan strategies that could potentially mitigate cracking and warping [16].

It should be noted that challenges associated with the accuracy of experimental measurements and computational models pose difficulties in investigating correlations between process parameters and print properties. It is possible that this difficulty is exacerbated by the limited parameters imposed on scan strategies due to the a priori adoption of simple strategies such as raster or spiral scanning as shown in our schematic in Fig. 1(a). For example, in the case of the simple back and forth raster, line spacing and raster angle (the angle between the scan lines and the consolidated surface) are the only parameters needed to describe the scan strategy (excluding parameters associated with the consolidated areas geometrical constraints), whereas it is nontrivial to describe an irregular path without rigorously listing the laser’s velocity at certain times which requires a relatively significant amount of interpretive effort, even for the short and somewhat structured random path in Fig. 1(a).

In this paper, we introduce a method for quantitatively describing laser scan strategies which can be used to draw correlations between it and post-print properties. In order to demonstrate its validity and applicability, thermomechanical simulations are performed with random laser strategies and used in tandem with the descriptor to train a machine learning model to predict post-print mechanical properties using the strategy descriptor. Neural networks (NNs) are utilized to provide a measure of the detected correlations between the descriptor and post-print properties which in turn gives insight on the descriptor’s ability to communicate a scan strategy’s spatiotemporal properties.

Section snippets

The descriptor: The Relative Spacetime Proximity (RSP) map

One of the main causal links between print properties and scan strategies is the spatiotemporal temperature distribution of the consolidated surface. Simply, areas on the surface where the laser more frequently visits may be generally hotter than areas that are less frequently visited. This idea is the driving force behind the descriptor proposed in this work, named the Relative Spacetime Proximity (RSP) map. The RSP map quantifies the spatiotemporal proximity of points on the laser path with

Simulation setup

In order to investigate the correlations between the RSP map and post-print properties, 10,000 random, 8 by 8, square grid Hamiltonian paths are generated and used in weakly coupled plane-strain thermoelastoplastic finite element simulations to obtain every path’s corresponding von Mises stress distribution at the end of each print process (i.e., at room temperature). The von Mises distributions provided a good amount of distributive complexity which served the purposes of this study well.

In

Neural network setup

Considering the complexity of the physics involved in mAM, machine learning stands out as an attractive investigative tool owing to its ability to efficiently extract correlations from chaotic datasets which provides avenues for deeper understandings [[30], [31], [32], [33], [34], [35]]. Neural networks (NN), or artificial neural networks, are mathematical models that map input data to output data through the utilization of artificial neurons connected in certain ways. The Feedforward NN, used

Results and discussion

The NN has a total of 240,131 trainable parameters (total number of weights and biases) trained on a data set of 10,000 simulations. The data was split into 7500 training, 1250 validating and 1250 testing data points. Fig. 4(a) shows the training and validation mean squared error (MSE) of the neural network over the training process. In order to better investigate the predictive error, an average percentage error map was made by averaging over every test prediction’s individual error map. A

Conclusions

It has been demonstrated that there are attainable correlations between the proposed laser scan strategy descriptor, the RSP map, and the outcomes of the simulated von Mises residual stress distribution at the end of a single layer pass. A neural network is trained to obtain the correlations and perform the predictions which are inspiringly accurate, instilling confidence in the RSP map as a descriptor. This provides hope for the methodology to be extended to higher accuracy predictions at

Declaration of Competing Interest

The authors report no declarations of interest.

Acknowledgements

This work used the Extreme Science and Engineering Discovery Environment (XSEDE) Bridges system, which is supported by National Science Foundation grant number ACI-1548562. Additionally, the authors acknowledge support from 3M and the Savio computational cluster resource provided by the Berkeley Research Computing program.

References (35)

Cited by (6)

  • Research and application of machine learning for additive manufacturing

    2022, Additive Manufacturing
    Citation Excerpt :

    The proposed methodology of this study is illustrated in Fig. 11. Researchers have also made effort to investigate and study the mechanical properties of fabricated parts in AM processes using ML technologies [145,146,153–162]. For exploring substitute model for traditional numerical simulation methods, Koeppe et al. [153] introduced a strategy, combining experiments, finite element (FE) simulations, and deep learning (DL) model, to predict the maximum von Mises and equivalent principal stresses of printed lattice-cell structures in AM.

View full text