Application of Machine Learning Methods to Improve Dimensional Accuracy in Additive Manufacturing

Abstract

Adoption of additive manufacturing for producing end-use products faces a range of limitations. For instance, quality of AM-fabricated parts varies from run to run and from machine to machine. There is also a lack of standards developed for AM processes. Another limitation is inconsistent dimensional accuracy error, which is often out of the standard tolerancing range. To tackle these challenges, this work aims at predicting scaling ratio for each part separately depending on its placement, orientation and CAD characteristics. Recent attention to machine learning techniques as a tool for data analysis in additive manufacturing shows that such methods as classical artificial neural networks (ANN), such as multi-layer perceptron (MLP), and convolutional neural networks (CNN) have a great potential. For the data collected on polymer powder bed fusion system (EOS P395), MLP outperformed CNN based on accuracy of prediction and mean squared error. The predicted scaling ratio can be used to adjust size of the parts before fabrication.

Impregnation quality diagnosis in Resin Transfer Moulding by machine learning

“…An alternative solution is to generate synthetic data by numerical simulation based on the conventional Finite Element Method (FEM). Data scarcity is solved by synthetic data in different process optimization [19,20] and structural validation scenarios [21,22]. FEM and ML techniques are combine in order to validate prediction and optimization methodologies [19].…”

“…Data scarcity is solved by synthetic data in different process optimization [19,20] and structural validation scenarios [21,22]. FEM and ML techniques are combine in order to validate prediction and optimization methodologies [19]. To ensure that the ML prediction is reliable, the FEM model on which it is based must be as close to reality as possible, which requires a previous characterization of the process and materials.…”

Machine learning in additive manufacturing: State-of-the-art and perspectives

“…Geometric errors can be minimized by three methods, namely rescaling the entire part, modifying the original CAD, and implementing process control. The scaling ratio can be predicted through MLP or CNN to adjust the overall size of parts before fabrication [120]. The shape-dependent geometric deviations due to thermal stress can be modelled by ML algorithms so as to make necessary geometric modification in CAD file.…”

Algorithm-Driven Paradigms for Freeform Optical Engineering

“…In a basic example, a MLP was trained to model thermal stress within the manufactured structure, from which modifications to the original design were made to compensate for thermal-driven geometrical deformations . CNNs have also been shown to be capable of accurately predicting the shape of printed features as a function of their location, orientation, and geometric design …”