Precision Manufacturing with Machine Learning

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Mar 29 2024

In a paper published in the journal Scientific Reports, researchers utilized machine learning (ML) to develop a model for predicting manufacturing layer height and grain size based on global energy distribution (GED) in additive manufacturing processes. The model, employing artificial neural networks (ANN), outperformed multi-linear regression in reducing error across the dataset.

Predictions for layer height achieved an impressive R2 and a root mean square error (RMSE), while grain size predictions showed an R2 and an RMSE. The study also revealed that adjustments in laser power and scanning speed led to grain refinement, a finding replicated in a different titanium alloy. These results underscore the complexity of achieving reproducibility in additive manufacturing processes due to the varied influence of constituent parameters on specific responses.

Related Work

Previous additive manufacturing (AM) research has focused on optimizing computer-aided design through layer-by-layer techniques like powder-based direct energy deposition using a laser beam (DED-LB). Variations in controlled process parameters, such as laser power and scanning speed, lead to print quality and mechanical property variability.

Energy input, often quantified as GED, influences crucial aspects like layer height and grain size. Researchers have employed ML techniques such as MLR and ANN to model these relationships. While ML has shown promise in optimizing AM processes, translating such models to bulk-build deposition remains a subject of exploration.

Powder Mixing Study

Titanium (Ti) and iron (Fe) (Ti–10Fe) alloy powders, composed of 50–100 μm diameter spherical powders, were thoroughly mixed for two hours using a Turbula Shaker to ensure homogeneity. This composition was chosen for its fully β-Ti grain microstructure, facilitating grain size measurement.

Researchers conducted DED-LB using a Trumpf TruLaser Cell 7020 with a continuous laser configuration. Process parameters were selected within the standard ranges for Ti-alloys, ensuring stable single tracks. Initial layer height approximation was determined by printing five initial layers of 10 × 10 mm at 0.2 mm layer height, followed by printing 10 × 10 × 10 mm3 cubes onto a CP-Ti substrate with CCD-dictated process parameters.

Experimentation involved 25 runs based on variable laser power, scanning speed, and spot size. The researchers measured cube heights and calculated the layer height. Grain size analysis was conducted using optical microscopy and validated by electron backscatter diffraction (EBSD).

The researchers modeled responses using MLR and ANN, with ANN employing Bayesian Regularization for robustness. They trained models on experiments 1–16 and validated them on experiments 17–25. Performance was evaluated using coefficients of determination (R2), RMSE, and mean absolute percentage error (MAPE).

Experimental Results Analysis

The results of the 25 experimental runs focused on average grain size and manufacturing layer height; across these runs, the variation in responses, such as average grain size and layer height, underscores the complexity of reproducibility solely through the GED parameter.

Layer height responses ranged from 0.32 to 0.97 mm, showing a linear trend with GED. However, there was notable variation in layer height for runs with identical GED values, highlighting the limitations of using GED alone. Researchers employed linear regression and ANN models to accurately predict layer height, with the ANN model demonstrating superior performance, particularly in capturing the nuanced influence of laser power and scanning speed.

The ANN model accurately predicted layer height responses, offering potential time savings in production. By leveraging insights from the ANN and MLR models, the study revealed that laser scanning speed significantly influenced layer height, followed by spot size and power. MLR and ANN models demonstrated high accuracy for average grain size, with the ANN slightly outperforming the MLR. Laser power emerged as the primary influencer of grain size, followed by scanning speed, with spot size playing a minor role.

The results underscore the importance of considering individual process parameters, such as laser power and scanning speed, rather than relying solely on GED. Reducing laser power and increasing scanning speed achieved grain refinement, improving mechanical properties in the tested alloy. These findings suggest broader applicability across various titanium alloys, emphasizing the significance of multi-variable optimization methodologies in achieving desired material properties, particularly in grain refinement objectives.

Conclusion

To sum up, the paper introduced a methodology for efficiently collecting and modeling build responses in additive manufacturing processes, particularly for Ti–10Fe alloy. Focusing on laser power, scanning speed, and spot size instead of the simplistic GED enabled researchers to achieve more accurate predictive models for layer height and average grain size using ANNs and MLR.

The study demonstrated that adjusting these parameters, mainly by reducing laser power and increasing scanning speed, led to significant grain refinement and production efficiency improvements. This approach offered promise for optimizing various manufacturing processes and controlling material properties beyond layer height, paving the way for advancements in additive manufacturing technology.