Research Background

Additive Friction Stir Deposition (AFSD) is an emerging and highly regarded Solid-state Additive Manufacturing (AM) technology. This method enables the fabrication of components with refined microstructures and excellent mechanical properties. However, the forming quality and mechanical performance of AFSD-produced components are directly governed by the varying thermal field distributions and material flow evolution at the contact interface, which are determined by processing parameters. These behaviors can be characterized through key parameters such as interface temperature, spindle force, and torque. Therefore, it is essential to employ various methods to obtain these parameters, thereby gaining deeper insights into thermal evolution and deformation mechanisms, and analyzing their impact on deposited component properties. Currently, machine learning (ML) has emerged as a viable approach in the AM field for effectively capturing complex correlations among diverse variables. Research indicates that ML techniques can explore the relationship between material microstructure and mechanical properties by treating process parameters and structural features as input and output variables, respectively. Furthermore, existing studies have demonstrated that artificial neural networks (ANN) and support vector machines (SVM) can accurately predict the fatigue life of 3D-printed 316 L stainless steel. However, no research has yet explored the application of machine learning techniques in the field of AFSD. Furthermore, the available data in the AFSD field currently lacks sufficient richness and stability in terms of extracted features, which significantly hinders the applicability of ML-based models. Therefore, there is an urgent need for a novel monitoring device capable of in-situ measurement of process parameters with high stability to obtain comprehensive data. This would facilitate the construction of a parameter-performance relationship network for Additive Friction Stir Deposition (AFSD) components.

Experimental Methods and Model Construction

Based on (Figure 2), the Additive Friction Stir Deposition (AFSD) equipment (a new-generation intelligent multifunctional solid phase manufacturing composite equipment, Aerospace Engineering Equipment (Suzhou) Co., Ltd.) was used to prepare AA6061 deposition parts. The fundamental process parameters were set as follows: Rotational speed (500 rpm), welding speed (140 mm/min), feed rate (11.7–35.0 mm/min), and layer thickness (0.5, 1.0, and 1.5 mm). The AFSD process yields a microstructure with uniformly arranged deposition layers and fully flowed material. During experiments, stable transmission signals (temperature, forging force, forging torque, spindle force, and spindle torque) were captured using an independently developed intelligent in-situ process monitoring toolholder (Fig. 2B, Temp-Mech-Aware toolholder, IDQ Science and Technology (Hengqin, Guangdong) Co. Ltd.). This tool holder automatically filters signals to ensure stability, with measurement ranges of 0–700 °C for temperature, 0–300 kN for force, and 0–300 Nm for torque. All embedded sensors maintain accuracy and frequency at 0.5% and 128 Hz, respectively. A series of 62 experiments yielded extensive monitoring data to validate the reliability of subsequent predictive models. Figure 3 shows real-time in-situ monitoring data corresponding to layer thicknesses of 0.5 mm (Figure 3A), 1.0 mm (Figure 3B), and 1.5 mm (Figure 3C). After AFSD testing, the microhardness and tensile strength of the deposited components were measured using a hardness tester and a tensile testing machine, respectively.

During the construction of ML models, multiple process parameters (temperature, up-thrust force, up-thrust torque, spindle force, and spindle torque) collected alongside fundamental process parameters (rotational speed, welding speed, feed rate, and layer thickness) are simultaneously incorporated as input variables. Microhardness and tensile strength are set as output variables for simulation and prediction based on ML algorithms (SVM, RF, and ANN). Furthermore, the model identifies the factors most significantly influencing the mechanical properties of the deposited components. All research findings can assist in refining machine learning models applied to the AFSD field and provide valuable technical recommendations for fabricating high-performance deposited specimens in practical engineering applications.

Figure 2-(A) Schematic diagram of the AFSD experimental process, (B) Additive Friction Stir Deposition (AFSD) equipment and physical image of the independently developed intelligent in-situ process parameter monitoring tool holder.

Figure 3—Monitoring data acquired by the intelligent in-situ process monitoring toolholder at layer thicknesses of 0.5 mm (Figure 3A), 1.0 mm (Figure 3B), and 1.5 mm (Figure 3C).

Research Findings

Figure 4 summarizes the prediction results obtained from SVM, RF, and ANN models. It indicates that, compared to SVM and RF models, the ANN model is most suitable for predicting the mechanical properties of AA6061 deposits prepared using AFSD. When predicting microhardness/tensile strength, it exhibits the highest prediction accuracy (0.9923/0.9998) alongside the lowest corresponding MAE (0.0738/0.0050) and RMSE (0.1373/0.0063). The prediction accuracy and generalization capabilities of the three ML methods ranked in the following order: ANN > RF > SVM. This is because the ANN method can recognize all interactions between input and output variables, and its training system can dynamically select the most suitable regression model for model construction, which directly improves its prediction accuracy. Furthermore, the hidden layer structure of the ANN model can effectively enhance its accuracy by fully fitting the training set.

Compared with previous research results from other groups (R² ranging from 0.21 to 0.89), this study demonstrates a significant improvement in all prediction accuracies. This advancement is largely attributed to the intelligent in-situ process parameter monitoring toolholder employed in this research, which enables stable and precise acquisition of input variables. However, previous studies primarily obtained datasets through thermocouples installed at the base of the substrate and torque sensors within external motors, while calculating spindle force signals based on the motor's current consumption. This indirect data acquisition process directly reduced data stability and accuracy, leading to poor precision in machine learning-based prediction models. The results reaffirm that the independently developed toolholder in this study provides more effective and stable data for constructing ML models, significantly enhancing corresponding prediction performance.

After verifying the optimal predictive accuracy of the ANN model, this study further evaluated the influence of input parameters on mechanical properties based on this model. As shown in Figure 5, feed rate (24.8%/24.1%) and layer thickness (25.6%/26.6%) exert dominant regulatory effects on microhardness and tensile strength, serving as key factors influencing the mechanical properties of deposited components. These two parameters directly influence the thermal evolution behavior of deposited components during the AFSD process. Increasing feed rate reduces heat accumulation effects, promotes grain refinement, and enhances the tensile properties of AA6061 deposited parts. Conversely, increasing layer thickness lowers the contact interface temperature during AFSD, leading to increased dislocation density and improved microhardness in the deposited components.

Figure 4 - Comparison of prediction results from different ML methods: (A) R², (B) MAE and RMSE, where (I) and (II) correspond to microhardness and tensile strength, respectively.

Figure 5 - Influence of Different Input Variables on (A) Microhardness and (B) Tensile Strength

Summary

This study employed various machine learning methods (SVM, RF, and ANN) to construct a network model correlating parameters with performance and predict the mechanical properties (microhardness and tensile strength) of AA6061 alloy processed via Additive Friction Stir Deposition (AFSD). The proprietary in-situ process parameter monitoring toolbar, which captures real-time process parameters, significantly enhanced the model's data richness and stability. Guided by material genomics principles, this study pioneered the application of ML model construction in AFSD through a proprietary intelligent device. This achievement provides technical guidance for high-throughput, high-precision, data-driven AFSD experiments and mass production of high-performance components, promising substantial reductions in trial-and-error costs and accelerated process development. Furthermore, this research paves the way for achieving digitalization and fully closed-loop control in additive manufacturing processes.

This work was published in the journal Materials Genome Engineering Advances. The research was jointly conducted by Professor Lap Mou Tam's team from the University of Macau and the Institute for the Development and Quality, Macau; the technical team from Aerospace Engineering Equipment (Suzhou) Co., Ltd.; and Professor Li Xiaogang's team from the University of Science and Technology Beijing.