Geometric Accuracy Prediction for Additive Manufacturing Through Machine Learning of Triangular Mesh Data

Abstract

While additive manufacturing has seen tremendous growth in recent years, a number of challenges remain, including the presence of substantial geometric differences between a three dimensional (3D) printed part, and the shape that was intended. There are a number of approaches for addressing this issue, including statistical models that seek to account for errors caused by the geometry of the object being printed. Currently, these models are largely unable to account for errors generated in freeform 3D shapes. This paper proposes a new approach using machine learning with a set of predictors based on the geometric properties of the triangular mesh file used for printing. A direct advantage of this method is the simplicity with which it can describe important properties of a 3D shape and allow for predictive modeling of dimensional inaccuracies for complex parts. To evaluate the efficacy of this approach, a sample dataset of 3D printed objects and their corresponding deviations was generated. This dataset was used to train a random forest machine learning model and generate predictions of deviation for a new object. These predicted deviations were found to compare favorably to the actual deviations, demonstrating the potential of this approach for applications in error prediction and compensation.

Predicting geometric errors and failures in additive manufacturing

“…The difference between the actual and predicted MSE was calculated for both the shuffled and the correct feature sequence. Finally, importance is given by normalizing the mean value of those differences by their standard deviation (STD) [the specifics of this process are provided in Section 4.2 of Decker and Huang (2019)]. Figure 19 shows the results of this measure.…”

“…From the results, it is clear all the features play a significant part in error prediction with importance values in [2.0, 20.0] except for the Gaussian curvature (probably due to the high correlation with the mean curvature). In Decker and Huang (2019), the importance of the predictors lies in [1.7, 5.7], where the coordinate z yields the maximum importance, whereas the elevation angles yield the minimum (≈1.7). Corresponding values for the importance of the features of our method and the importance of the predictors of Decker and Huang (2019) exhibit a consistent behavior.…”

“…In Decker and Huang (2019), the importance of the predictors lies in [1.7, 5.7], where the coordinate z yields the maximum importance, whereas the elevation angles yield the minimum (≈1.7). Corresponding values for the importance of the features of our method and the importance of the predictors of Decker and Huang (2019) exhibit a consistent behavior.…”

“…To compare our work with Decker and Huang (2019), we determined the importance of each feature for our prediction scheme by permuting the order of the values and observing how network loss is affected. Specifically, feature permutation was Figure 19 shows the results of this measure.…”

“…Decker and Huang (2019) introduced a shape-driven approach for predicting the geometric accuracy of an AM product based on machine learning that uses a set of eight predictor variables to express the geometric attributes of the triangular mesh model used for printing (.stl model). A novel method that trains a random forest to predict the geometric error of a vertex is presented.…”