The purpose of this study was to develop a multiview recurrent neural network (RNN) model to identify linkage between relatively inexpensive photodiode signals and relatively expensive interlayer temperature measurements during the laser powder bed fusion additive manufacturing (PBF-LB AM) of complex geometry samples.
Two sources of in situ processing data – photodiode sensor signals synchronized to laser position and forward-looking infrared (FLIR) images – were collected. Features were extracted from photodiode heatmaps of each sample cross-section using a pre-trained deep convolutional neural network (DCNN), then used to train an RNN regression model to predict interlayer temperature taking FLIR measures as ground truth.
For the geometries investigated, the proposed model achieved high accuracy (average of 95.1%) in predicting layer-wise temperatures and success in identifying trends in temperature changes (average R2 of 0.98) along the sample build direction.
Compared to conventional image-based machine learning models (i.e. CNN), the proposed RNN-based model workflow incorporated local geometry information along with in process signals enabling the prediction of layer-wise temperature in complex geometry samples. The proposed workflow provides a promising route for predicting in process temperature from photodiode signals for real-time thermal monitoring during PBF-LB AM fabrication of complex geometry samples, having potential applications for in process feedback control as well as the development and validation of multiphysics models for AM.
