Methodology for Data to Support Directed Energy Deposition Additive Manufacturing Qualification
TWI Industrial Member Report 1224-2026
By Georgia Schneider, PhD, MMath, CEng, MWeldI, Karan Derekar and Satish Bhattiprolu
Industrial Need
In industries such as aerospace and automotive, where precision, performance, and efficiency are paramount, Directed Energy Deposition (DED) technology has emerged as a transformative range of additive manufacturing (AM) processes. DED offers the potential to create intricate, custom-designed components with enhanced material properties. However, the reliance on laborious and expensive trial-and-error experimentation to determine optimal process parameters and toolpaths hinders the realisation of DED's full potential.
The current practice of conducting numerous developmental pre-trials to understand the resulting bead geometry based on selected parameters is not only resource-intensive but also leads to over-building of parts due to conservative estimates. There is an urgent need for a more efficient and data-driven approach to optimise DED processes, minimize wastage, and enhance overall productivity. This work provides that approach to predict bead geometry from manufacturing parameters using SciKit-Learn to determine an optimised machine learning algorithm.
Key Findings
- For each of the four prediction tasks (i.e. bead height and bead width for laser DED and arc DED), fourteen regression models were compared using Root Mean Square Error (RSME) regression models on 5-fold testing.
- The CatBoost regressor, one of the regression models considered in this work, achieved the lowest RMSE for both laser-wire DED bead width and bead height predictions and was therefore selected as the final model for both tasks.
-
Feature importance analysis for the laser DED data showed that:
o Transverse speed was the dominant factor controlling laser-wire DED bead width.
o Wire feedrate and transverse speed were the most influential parameters for bead height.
o Laser power played a minor but non-negligible role in both cases.
- The Extra Trees regressor, another of the regression models considered in this work, achieved the lowest RMSE for wire-arc bead width prediction.
- The K-nearest Neighbours regressor, another regression model considered, achieved the lowest RMSE for wire-arc bead height prediction.
- The CatBoost, Extra Trees and K-Nearest Neighbours are different machine learning algorithms. CatBoost trains models sequentially with each new model reducing the residual errors; Extra Trees builds multiple independent models and averages the predictions; K-Nearest Neighbours when given a new datapoint, scans the training dataset and find the most similar instances, averaging their target values
- Feature importance analysis identified current (A) as the primary driver of bead geometry, followed by travel speed, while voltage exhibited negligible influence.
Impact
A comprehensive machine learning framework was developed and evaluated to predict bead width and bead height for both laser-wire DED and wire-arc DED processes using manufacturing parameters. This provides a more efficient and data-driven approach to optimise DED processes, minimise wastage, and enhance overall productivity.