Aluminum-Zinc-Magnesium-Copper alloys (AA7xxx series) are crucial in aerospace and transportation due to their high strength-to-weight ratio, corrosion resistance, and cost-effectiveness. However, the demand for even lighter and stronger materials in high-speed rail and aircraft necessitates further performance improvements. The properties of AA7xxx alloys are complexly influenced by composition, solution treatment, and aging treatment. Traditional optimization methods, such as first-principle calculations, phase field simulations, and experimental trial-and-error, face limitations when dealing with the increasing complexity of these alloys. The vast number of process combinations in a three-stage solution-aging process for instance, exceeds 10 million, making traditional trial-and-error methods impractical. Transfer learning, a machine learning technique, offers a solution by leveraging knowledge from related fields to overcome data scarcity issues. Three types of transfer learning exist: feature transfer, parameter transfer, and sample data transfer. This study focuses on sample data transfer, using existing data from similar alloys to design a new alloy's process parameters. The authors developed a new aluminum alloy, E2, with superior strength and toughness, and propose a strategy to optimize its solution-aging process using data from commercial AA7xxx alloys.

The literature review extensively covers existing research on AA7xxx series aluminum alloys, highlighting the influence of alloy composition (Mg, Cu, Zn, etc.), solution treatment, and aging treatment on their mechanical properties. Studies emphasizing the challenges of optimizing these alloys using conventional methods are presented. Additionally, the existing literature on transfer learning, particularly sample data transfer, is reviewed, citing relevant applications in materials science. The authors reference studies that have successfully used machine learning for materials design, highlighting the use of databases like OQMD and Material Project. Prior work on using transfer learning to predict material properties, particularly formation energies and thermal conductivity in polymers, is discussed.

The methodology involves a two-stage process: data preparation and machine learning modeling. First, a basic dataset (BDS) containing 1053 data points of composition, solution-aging parameters, and properties of commercial AA7xxx alloys was compiled from literature. A target dataset (TDS) was created from 20 pre-experimental data points for the new E2 alloy. The TDS was used to guide weight adjustment during the machine learning process. The TrAdaBoost algorithm, a transfer learning algorithm, was employed to construct a predictive model. TrAdaBoost iteratively refines weights of samples from the BDS and TDS, focusing on data that significantly improves the E2 alloy prediction performance. This iterative approach refines the model’s prediction accuracy for the E2 alloy. A multi-objective genetic algorithm (NSGA-II) was then used to generate a Pareto front illustrating the optimal combinations of ultimate tensile strength (UTS) and elongation (δ). Seven sets of solution-aging parameters (HT1-HT7) were selected from the Pareto front for experimental verification. The process also included developing a control sample using a single-stage solution-aging treatment (HTO) with the same solution parameters as HT2 to separate the effects of multi-stage solution and aging processes. Microstructural characterization, using SEM, EBSD, TEM, and APT, was carried out to understand the mechanism behind the observed improvements.

The study successfully optimized the solution-aging process of the E2 alloy using data transfer learning. The optimized T66R process (450 °C@1h + 470 °C@1h + 480 °C@0.5 h + 65 °C@42 h + 100 °C@16h + 135 °C@4 h) resulted in a significant increase in both UTS and elongation, increasing from 715 ± 6 MPa and 8.4 ± 0.4% to 767 ± 6 MPa and 13.4 ± 0.5%, respectively, compared to the traditional T6 process. Microstructural analysis revealed a reduction in micron-scale insoluble phases by an order of magnitude in the E2 alloy treated with the T66R process. Furthermore, the number density of finer, nearly spherical, and more uniformly dispersed nanoprecipitates increased by over 50%. Analysis of dislocation movement mechanisms showed that the T66R process favors dislocation shearing, unlike the bypassing mechanism observed in the HTO sample. This shift in mechanisms contributes to the simultaneous enhancement of both strength and ductility. The decrease in the width of the precipitation-free zone (PFZ) from 20 nm (HTO) to 16 nm (T66R) also played a crucial role in improving ductility by mitigating solute segregation and reducing the probability of micropore formation. The researchers also verified the effectiveness of the T66R process on other alloys such as AA7050, achieving similar improvements in strength and ductility.

The results demonstrate the effectiveness of using sample data transfer machine learning for rapid alloy process optimization. The substantial improvement in both strength and ductility of the E2 alloy, achieved with a minimal number of experiments, showcases the potential of this approach for accelerating materials development. The observed microstructural changes, particularly the refinement and increased density of precipitates, along with the shift in dislocation movement mechanism, confirm the relationship between processing, microstructure, and resulting mechanical properties. The comparison with other alloy processing techniques (conventional, spray deposition, powder metallurgy, severe plastic deformation) highlights that this data-driven approach can surpass the performance limitations of existing methods. The success of this method suggests the applicability of this approach to other alloy systems, reducing the time and cost associated with traditional materials development strategies.

This research successfully employed sample data transfer machine learning to optimize the solution-aging process of a novel aluminum alloy (E2), achieving a remarkable simultaneous improvement in strength and ductility. The optimized T66R process significantly reduced the number of experiments required compared to traditional trial-and-error methods. Microstructural analysis elucidated the underlying mechanisms leading to the enhanced properties. This work provides a valuable framework for leveraging existing data to accelerate the development of high-performance alloys. Future research could explore the application of this methodology to other alloy systems and further refine the machine learning models to improve prediction accuracy and efficiency. Exploring different transfer learning algorithms and larger datasets could also be avenues for future investigation.

The accuracy of the machine learning model is highly dependent on the quality and relevance of the data in the basic dataset. The selection of relevant data and the absence of bias in the dataset are critical. The generalizability of the findings to other alloy systems may be limited until further validation is performed. The microstructural analysis, though comprehensive, may not have explored all factors influencing the enhanced properties. Further investigations into the detailed interactions between precipitates and dislocations, and the precise role of grain boundaries, are required for a more comprehensive understanding.