Harnessing On-Machine Metrology Data for Prints with a Surrogate Model for Laser Powder Directed Energy Deposition

TL;DR

This work addresses the challenge of reliably predicting LP-DED outcomes under parameter uncertainty by leveraging large-scale on-machine metrology data. It develops a fast, data-driven surrogate built via Dynamic Mode Decomposition with Control (DMDc) that ingests 21 process inputs to predict melt-pool size, melt-pool temperature, and real-time working distance, with uncertainty bounds derived from Leave-3-Out CV on approximately 1.5 million data points. The results show high predictive performance for melt-pool size and temperature ($R^2 \approx 0.94$–$0.96$, RMSE on the order of $0.25$–$0.28$ mm for size and about $140$ °C for temperature), while real-time working distance is weaker ($R^2 \approx 0.64$–$0.72$) but improves substantially after imputing camera-off states. The study also identifies minimum OMM recording frequencies needed to preserve model fidelity and demonstrates ms-scale inference with real-time uncertainty bounds to enable monitoring and potential feedback control in LP-DED.

Abstract

In this study, we leverage the massive amount of multi-modal on-machine metrology data generated from Laser Powder Directed Energy Deposition (LP-DED) to construct a comprehensive surrogate model of the 3D printing process. By employing Dynamic Mode Decomposition with Control (DMDc), a data-driven technique, we capture the complex physics inherent in this extensive dataset. This physics-based surrogate model emphasizes thermodynamically significant quantities, enabling us to accurately predict key process outcomes. The model ingests 21 process parameters, including laser power, scan rate, and position, while providing outputs such as melt pool temperature, melt pool size, and other essential observables. Furthermore, it incorporates uncertainty quantification to provide bounds on these predictions, enhancing reliability and confidence in the results. We then deploy the surrogate model on a new, unseen part and monitor the printing process as validation of the method. Our experimental results demonstrate that the predictions align with actual measurements with high accuracy, confirming the effectiveness of our approach. This methodology not only facilitates real-time predictions but also operates at process-relevant speeds, establishing a basis for implementing feedback control in LP-DED.

Figures (9)

• Figure 1: Chart describing the intended prediction workflow for this study and more generally the concept of a virtual test bed. G-code or time series input signals are provided, then conditioned and input into the state-space model. The state-space iteratively predicts the next time-step for the entirety of the provided input. The predictions are collected and combined with uncertainty information. The output of the workflow is both a times series and geometric digital representation of the final object.
• Figure 2: Distributions of DED observables versus their uniform signal bounds.
• Figure 3: DMDc model error/explained variance as a response to OMM recording frequency. Melt pool temperature begins degrading below 50 Hz. Melt pool size and working distance start showing significant drop off in accuracy below 20 Hz.
• Figure 4: Experimental data spectrograms (left) and state-space model spectrograms (right) for all three observables. These spectrograms are frequency ($Hz$) as a function of laser pulse length ($s$).
• Figure 5: DMDc state-space model results for reproductive, or training, (a) and test data (b). (Left) Time series of melt pool size for both ground truth and DMDc state-space model. (Right) Histogram of residuals between the experimental data and model predictions.