Single Weld Bead Dimension Prediction for TIG Arc Additive Manufacturing

Literature Overview

This paper by Zhao Peng, Lv Yanming, Zhou Wenjun, Pan Yu, and Bai Shaoyun, published in Materials for Mechanical Engineering (2020, Vol. 44, No. 11, pp. 78-82), addresses the critical challenge of predicting single weld bead geometry in TIG arc additive manufacturing (AM) of GH4169 nickel-based superalloy. The authors employed orthogonal experimental design to systematically investigate the effects of welding current, travel speed, and wire feed speed on bead width and reinforcement height. A back-propagation (BP) neural network model, optimized by genetic algorithm, was developed to predict bead dimensions with relative errors within 6%. A MATLAB-based graphical user interface was created to facilitate practical application of the prediction model.

Experimental Design and Methodology

Orthogonal Experimental Matrix

The orthogonal experimental design was used to efficiently explore the parameter space with a manageable number of experiments. The key process parameters and their levels were:

The orthogonal array L9(3⁴) was selected, requiring only 9 experimental runs to evaluate three factors at three levels each. This design efficiently captures the main effects and two-way interactions while minimizing experimental effort.

Specimen Preparation and Characterization

Weld beads were deposited on GH4169 superalloy substrate plates using a TIG arc additive manufacturing platform. The substrate plates were machined to a flat surface with a roughness of Ra < 1.6 μm to ensure consistent starting conditions. Each weld bead was characterized through:

• Visual inspection: Assessment of surface quality, bead uniformity, and the presence of defects such as undercut, spatter, or porosity.
• Cross-sectional examination: Metallographic preparation including sectioning, mounting, grinding, polishing, and etching to reveal the bead profile.
• Dimensional measurement: Bead width and reinforcement height were measured at multiple points across each bead using optical microscopy, with averages taken to minimize measurement error.

Results and Analysis

Weld Bead Width Behavior

The experimental results demonstrated clear trends in bead width behavior:

The non-monotonic behavior of bead width with respect to wire feed speed is particularly interesting and physically insightful. At low wire feed speeds, the increased metal deposition rate widens the bead by adding more material to the weld pool. However, at higher wire feed speeds, the increased solidification rate due to the larger volume of metal being deposited causes the weld pool to narrow, as the thermal energy available per unit volume of deposited metal decreases. This creates a characteristic maximum in bead width at an intermediate wire feed speed.

Weld Bead Reinforcement Height Behavior

The reinforcement height exhibited a different set of dependencies:

The linear positive correlation between reinforcement height and wire feed speed is intuitive: more metal deposited per unit length results in greater buildup above the substrate surface. The negative correlations with current and travel speed are also physically reasonable: higher current increases penetration depth relative to deposition volume, while higher travel speed reduces the heat input per unit length, both leading to flatter beads.

Prediction Model Performance

The BP neural network model, optimized by genetic algorithm, achieved the following performance metrics:

The 6% maximum relative error represents a significant improvement over empirical formulas or simple linear regression models, which typically exhibit errors of 10-15% for weld bead geometry prediction. The genetic algorithm optimization of the BP network was crucial in avoiding local minima during training and ensuring the model captured the nonlinear relationships between process parameters and bead geometry.

Engineering Application and Discussion

Practical Significance

In TIG arc additive manufacturing, maintaining consistent bead geometry across multiple layers is essential for achieving dimensional accuracy and structural integrity. Deviations in bead width or height can lead to cumulative errors that result in significant dimensional inaccuracies in the final part. The prediction model developed in this study provides a powerful tool for:

1. Process planning: Selecting appropriate process parameters to achieve target bead dimensions before starting production.
2. In-process monitoring: Comparing measured bead dimensions with predicted values to detect process drift or parameter deviations.
3. Quality assurance: Establishing statistical process control limits based on predicted bead geometry variability.
4. Multi-layer strategy optimization: Using single-bead predictions to plan layer-by-layer build strategies that account for thermal interaction between adjacent beads.

Limitations and Future Work

While the 6% prediction accuracy is impressive, several limitations should be acknowledged:

• The model was developed for a specific substrate material (GH4169) and a specific wire composition. Transfer to other materials requires retraining or recalibration.
• The model does not account for environmental factors such as ambient temperature, humidity, or gas purity, which can influence weld geometry.
• The single-bead model does not consider inter-bead or inter-layer thermal interaction, which becomes significant in multi-layer builds.
• The model assumes stable arc conditions and does not account for transient phenomena such as arc wander or wire feed irregularities.

Future work should focus on developing physics-based models that incorporate heat transfer, fluid dynamics, and solidification theory to provide more transferable prediction capabilities. Additionally, integrating the prediction model with real-time sensing systems (such as vision-based bead monitoring or thermal imaging) would enable closed-loop process control in additive manufacturing.

Study Insights

This paper makes a valuable contribution to the field of TIG arc additive manufacturing by demonstrating that data analysis approaches can effectively capture the complex nonlinear relationships between welding parameters and bead geometry. The use of genetic algorithm optimization to train the BP neural network is particularly noteworthy, as it addresses the well-known problem of local minima convergence in BP network training. The development of a MATLAB GUI interface for practical application represents a commendable effort to bridge the gap between research and industrial implementation.

For engineers working in additive manufacturing, this study underscores the importance of systematic experimental design in building reliable prediction models. The orthogonal experimental approach not only reduces the number of experiments required but also provides statistically rigorous estimates of parameter effects and interactions. The 6% prediction accuracy, while not perfect, is sufficient for many practical applications where bead geometry tolerances are on the order of ±0.5 mm or greater. However, for high-precision applications requiring tighter tolerances, additional process control measures and real-time monitoring systems will be necessary.