To ensure thorough coverage of the literature on the application of AI, ML, and DL in Mg alloys, bibliometric data were gathered from two prominent scientific databases: Scopus (Elsevier) and Web of Science (Clarivate Analytics). The search utilized combinations of the keywords:

((TITLE-ABS-KEY (magnesium) AND ABS-KEY (alloy) AND TITLE-ABS-KEY (artificial intelligence) OR TITLE-ABS-KEY (ML) OR TITLE-ABS-KEY (DL)) AND (LIMIT-TO (LANGUAGE, “English”)) AND (LIMIT-TO (DOCTYPE, “ar”)))

In both databases, search filters were applied to include only articles in English and peer-reviewed research, excluding conference proceedings, reviews, and other non-research outputs. A total of 387 records were obtained from Scopus and 252 from Web of Science. After eliminating duplicate entries in the WoS dataset, 237 unique articles remained. The finalized datasets obtained from both databases formed the basis for the following cluster development and validation across multiple databases. Table 1 provides a comprehensive overview of the literature search findings.

To ensure the quality, consistency, and relevance of the analyzed dataset on AI, ML, and DL in Mg alloys, specific inclusion and exclusion criteria were established and applied during the literature search process. Only peer-reviewed journal articles were included, while conference proceedings, book chapters, reviews, editorials, letters, and non-research communications were excluded. The focus on journal articles ensures that the analysis relies on original research studies that have undergone thorough peer review. It leads to preserving scientific validity and methodological soundness.

Only publications in English were taken into account. This restriction was implemented to ensure consistency in language interpretation, keyword extraction, and semantic analysis during bibliometric mapping. The bibliometric tool VOSviewer heavily relies on accurate keyword co-occurrence and text mining (van Eck and Waltman, 2010; Kumar and Goel, 2021). If articles are considered in various languages, this can introduce inconsistencies due to translation differences or incomplete indexing of non-English terms. The applied filters ensure both methodological rigor and global relevance of the bibliometric findings on AI, ML, and DL in Mg alloys.

The comparative analysis between Scopus and Web of Science (WoS) datasets collected for AI, ML, and DL in Mg-alloys is depicted in Supplementary Table SM1. It demonstrates a high degree of consistency across multiple bibliometric parameters for AI, ML, and DL in Mg alloys. For author keywords, index keywords, contributing authors, journal sources, and publishing countries, the overlap ranged from 70% to 85%. This indicates a strong thematic alignment between the two databases (Scopus and WoS), despite differences in their indexing scope and coverage. This cross-database validation supports the reliability and generalizability of the cluster analysis conducted using the Scopus dataset. Figure 4 presents a side-by-side comparison of key bibliometric parameters.

The study utilized the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) framework to ensure a structured and transparent approach to conducting the systematic literature review (Page et al., 2022; Kaur S. et al., 2025). Although bibliometric data were initially retrieved from both Scopus and WoS for multi-database validation, the Scopus database was selected as the primary source for detailed cluster formation and thematic analysis due to its broader journal coverage, more extensive indexing of materials science, engineering, and computational AI-related publications, and its seamless integration with the VOSviewer bibliometric mapping tool.

The PRISMA flow diagram (Figure 5) shows 387 Scopus records initially retrieved. After screening for language, document type, and topic relevance, 254 studies remained. Of these, 162 highly cited articles were thematically clustered, and 92 relevant low-citation papers were added for comprehensive coverage, totaling 185 studies. The inclusion criteria focused on peer-reviewed, English-language journal articles applying AI/ML/DL to Mg alloys, while non-English texts, non-article types, and irrelevant studies were excluded to ensure thematic coherence and analytical quality.

VOSviewer (version 1.6) was utilized for cluster analysis, employing Scopus author keywords (N = 815, aligned with Scopus export compatibility for WoS) as input. By applying a minimum occurrence threshold of three, a total of 46 keywords were organized into nine distinct clusters. These clusters reflect thematic trends related to the structural applications of AI, ML, and DL in the context of Mg alloys. Manual labeling was subsequently conducted, wherein the presence of keywords and representative articles associated with specific topics were analyzed contextually to ensure accurate thematic representation. For example, the keywords “grain size” and “ANN” were assigned the label ‘mechanical behavior modeling using neyrak networks’. This expert-guided approach facilitates the identification of emerging research areas and existing gaps within the AI-driven literature on magnesium alloys.

• Cluster I: Mechanical behavior modeling using neural networks

Cluster I includes keywords such as aluminum alloys, ANN, constitutive model, EBSD, grain size, hardness, Mg, neural network, tensile strength, and twinning. Cluster I emphasizes predicting mechanical properties (e.g., tensile strength, hardness) and microstructural characteristics (e.g., grain size, twinning) through ANNs and constitutive modeling techniques.

• Cluster II: Alloy design and performance prediction using AI/ML

Cluster II includes keywords such as alloy design, artificial neural network, corrosion behavior, crystal structure prediction, DL, genetic algorithm, mechanical property, metals and alloys, and Mg alloy. It highlights AI/ML methods like DL, genetic algorithms, and ANNs in designing new Mg alloys and predicting properties like corrosion behavior and crystal structure.

• Cluster III: Atomic-scale simulations and ML potentials

Cluster III includes keywords such as density functional theory, dislocation, grain boundary segregation, ML potential, molecular dynamics, and tensile properties. This cluster emphasizes atomistic modeling techniques (DFT, MD) and ML-based interatomic potentials for understanding fundamental properties such as dislocations and grain boundaries.

• Cluster IV: AI in welding and advanced processing techniques

Cluster IV includes keywords such as artificial intelligence, friction stir processing, friction stir welding, magnesium alloys, and metal hydrides. Cluster IV focused on the application of AI in welding and processing technologies, such as friction stir welding, particularly for magnesium alloys.

• Cluster V: Biomaterials and microstructural optimization

Cluster V includes keywords such as biomaterials, magnesium alloy, mechanical properties, microstructure, and rare earth elements. It focuses on the biomedical applications of Mg alloys, examining how microstructure and rare-earth additions affect mechanical properties.

• Cluster VI: First-principles and active learning approaches

Cluster VI encompasses keywords such as active learning, density functional theory (DFT), first-principles calculations, and magnesium (Mg) alloys. It combines first-principles (DFT) methods with active learning, representing a data-efficient approach to modeling and material discovery in Mg alloys.

• Cluster VII: data-driven design and optimization techniques

Cluster VII encompasses feature selection, ML, and optimization keywords. It represents core ML techniques such as feature selection and optimization used in materials data analysis and model refinement.

• Cluster VIII: Corrosion modeling

Cluster VIII includes keywords such as corrosion and prediction. A small but focused cluster specifically on using AI/ML for corrosion prediction in Mg alloys.

• Cluster IX: Alloy composition and design strategies

Cluster IX includes keywords such as aluminum alloy and composition design. It focused on compositional optimization of Mg and Al alloys using computational design strategies.

This terminology will clearly define each cluster’s thematic scope while remaining focused on the overall aim of mapping advancements in AI, ML, and DL within Mg alloy research.

The initial cluster structure was refined to enhance thematic clarity and analytical rigour within this systematic review. While the original nine clusters were generated using the VOSviewer keyword co-occurrence algorithm, subsequent manual curation and semantic evaluation revealed partial overlaps and redundancies. Consequently, the clusters were reorganised to improve conceptual coherence and ensure interpretive precision.

Specifically, several clusters were identified as converging around similar research themes and methodological frameworks. For example, clusters centred on atomic-scale simulations, such as Cluster III (Atomic-Scale Simulations and ML Potentials), and first-principles learning frameworks, such as Cluster VI (First-Principles and Active Learning Approaches), both encompassed studies utilising DFT, MD, and ML potentials.

Similarly, research categorised under Cluster II (Alloy Design and Performance Prediction using AI/ML) and Cluster IX (Alloy Composition and Design Strategies) were found to share a common focus on alloy development through AI/ML models and inverse design approaches. Although initially treated as distinct due to variations in keyword density, further keyword correlation analysis and metric-based structural cohesion indicated that these clusters should be consolidated to reflect their underlying conceptual alignment.

The refinement process was conducted in three systematic stages:

1. Re-evaluation of keyword semantics and article content to identify conceptual overlaps across clusters.

This refinement fulfils two primary objectives:

• It mitigates conceptual fragmentation and redundancy within the analytic framework, thereby enhancing the structural integrity of the review.

So, the refined seven clusters for thematic classification are as follows:

• Cluster 1: Mechanical behavior modeling using neural networks

VOSviewer was chosen for its effectiveness in identifying thematic structures through bibliographic coupling and keyword co-occurrence mapping van Eck and Waltman, 2010. Its modularity optimization and local movement algorithm support precise clustering, making it particularly valuable for mapping multidisciplinary intersections like AI applications in Mg alloy research.

After generating thematic clusters, a systematic review was conducted within each group. Cluster names were assigned based on author keywords used as filters to identify relevant articles. Since some keywords overlapped, particular articles appeared in multiple clusters. To avoid redundancy, duplicates were removed, and each article was assigned to the most appropriate cluster.

Cluster 1, “mechanical behavior modeling using neural networks,” emphasizes using ML and ANNs to accurately model essential mechanical properties, such as flow stress, tensile strength, and fatigue life. These data-driven models, encompassing hybrid and optimized ANN versions, demonstrate enhanced prediction accuracy and facilitate the practical assessment and optimization of Mg alloys in various processing and loading scenarios. Mg alloys, such as AZ31 and AZ31B, are valued for their lightweight and structural potential. However, they face limitations in mechanical performance, such as low ductility and fatigue resistance. Traditional modeling approaches often fail to capture the complex, nonlinear relationships between processing, microstructure, and mechanical properties.

Recent research has shown the successful application of ML and ANN in simulating the mechanical properties of AZ31 and AZ31B Mg alloys, as shown in Supplementary Tables SM2. Yu et al. (2024) utilized a combined ML and energy-based model to forecast the low-cycle fatigue life of AZ31 alloy. This model surpassed traditional fatigue prediction techniques, exhibiting strength across various loading directions. Wang S. et al. (2024) examined the maximum thinning rate of AZ31 alloy through a hybrid BPNN-RF method, which incorporated a Genetic Algorithm (GA) and Dragonfly Optimization (DBO). The findings revealed the optimal formability conditions at 250°C and noted a prediction error as low as 0.32%, highlighting the significant impact of the resistance coefficient on formability.

Murugesan et al. (2023) focused on the flow stress behavior of AZ31B alloy using multiple ANN variants, including BP-ANN, GABP-ANN, and CFBP-ANN. The CFBP-ANN model yielded the most accurate results when trained on data from strain, temperature, and strain rate. Zhang Y. et al. (2023) utilized ANN to explore the relationship between tensile properties and crystallographic texture in AZ31 alloy. The model effectively linked texture features, such as Imax, D, PLD, and PTD, with mechanical outcomes, including yield strength and elongation, highlighting the importance of microstructural attributes. Yadav and Khurana (2022) utilized a combined ANN and GA framework to predict tensile strength, elongation, hardness, and grain size in AZ31B alloy joints. The model achieved a prediction error of less than 3% while optimizing joint performance across multiple objectives. Jaafreh et al. (2022) investigated stress concentration behavior using various ML models on pure Mg and AZ31. The ERT model demonstrated the highest accuracy, with grain size being the most significant factor; finer grains resulted in reduced stress concentration.

These studies confirm that hybrid or ensemble approaches combining ML and ANN-based models can effectively predict mechanical behaviors in Mg alloys. Models incorporating microstructural parameters, such as texture and grain size, exhibit enhanced predictive capability. The optimization techniques, such as GA and DBO, significantly improve the performance and reliability of these models, making them powerful tools for design and performance evaluation in materials engineering (Jaafreh et al., 2022; Yadav and Khurana, 2022; Murugesan et al., 2023; Zhang Y. et al., 2023; Wang S. et al., 2024; Yu et al., 2024).

Supplementary Tables SM3 depicts recent studies (2021–2025) that applied various ML and ANN techniques. It includes research aimed at predicting and enhancing the mechanical and microstructural characteristics of Al-Mg and Al-based alloys. The primary properties explored were tensile strength, flow stress, corrosion resistance, hardness, wear behavior, hydrogen storage capacity, and grain size.

Jain et al. (2024) utilized the ensemble ML model CatBoost to predict the hardness of Al-Mg-X alloys. The alloy’s composition and heat treatment data revealed exceptional performance. Kim et al. (2024) used ANN, CNN, and RF to identify new high-hardness Al-/Mg-based alloys. It also applied explainable AI to comprehend critical influencing factors. Bai et al. (2024) used backpropagation ANN and clustering to model flow stress and dislocation density in an Al-Zn-Mg-Cu-Zr alloy. The ML model achieved an R² of 98.67%.

Several studies have concentrated on Mg-based or hybrid Al-Mg alloys. Chandana and Kamesh (2024) utilized Bayesian-regularized ANN to forecast hydrogen storage and release in Mg-based alloys. It highlighted the significance of particle size and pressure. El-Garaihy et al. (2024) investigated the properties of ZX30 Mg alloy using ANN and Gaussian Process Regression, realizing that processing methods and the number of passes significantly impacted corrosion and yield strength. Zhu G. et al. (2025) confirmed high tensile strength (668 MPa) in Cu-Cr-Zr-Mg-Ti alloys by employing ANN and GA to capture micro-alloying effects. Efa (2024) applied BP-ANN with gradient descent to examine Ti-interlayer effects in AA7075 and AZ61A. It accurately predicted residual stress and peak temperature during friction stir welding (FSW).

Many studies highlighted hybrid or interpretable ML frameworks. Jiang L. et al. (2024) presented a data-driven design of advanced Al-Zn-Mg-Cu-Ti-Cr-Zr alloys. It correlated elemental additions with improved strength, toughness, and corrosion resistance using interpretable ML. Pei et al. (2024) combined ANN with internal state variable (ISV) and Arrhenius models for stress-strain prediction in Mg-Gd-Y-Zr alloys under hot forming. It showed excellent agreement with experimental data. Li G. et al. (2023) used GA-BPNN to predict the slurry temperature in ECP-processed 7075 Al and achieved R² = 0.9988. Mokhtari et al. (2024) compared DL architectures, ANN, GRU, LSTM, and RNN. The results revealed that memory-enabled models, RNN and LSTM, predicted better tensile strength in Al-Mg alloys.

Gurugubelli et al. (2023); Wang S. et al. (2024) revealed that AI integrated with simulations or hybrid optimization GA-DBO was able to predict formability and tensile strength with reduced experimental dependency. Xue et al. (2023) predicted corrosion resistance in Al-Mg alloys using minimal input data. The study highlighted Mg content and annealing parameters as key features. Mishra (2024a) applied 6 ML models to FSW Mg alloys, with XGBoost achieved the highest R² (0.81), while Troncoso et al. (2023) used ML to detect twinning interfaces in Mg-10%Al alloys via stress field analysis. Byun et al. (2024) and Medghalchi et al. (2023) used advanced DL models (GAN-GRU, YOLOv5) for anisotropic deformation and micro-damage detection in ZK60 and Ca-alloyed Mg, respectively.

Jagadeesh and Setti (2023) used ANN to predict wear and tribological behavior for Mg Ze41A alloys, while Nithin et al. (2022) used CNNs to predict grain evolution in Al-Si-Mg alloys with over 90% accuracy. Masood Chaudry et al. (2021) and Shu et al. (2022) employed Gradient-Boost Trees and graph-based neural networks (HGGAT) to model unexplored Al alloys and microstructure-property relationships, respectively.

ANN and hybrid ML approaches (e.g., GA-BPNN, PSO-BP ANN, ensemble ML) have remarkably succeeded in modeling and predicting complex mechanical behaviors in Al-Mg and Al-based systems. Hybrid or optimized ANN models integrate additional techniques like genetic algorithms, fuzzy logic, or swarm intelligence to enhance prediction accuracy and convergence speed. Compared to traditional ANNs, these models offer improved generalization and robustness in capturing the complex nonlinear behavior of Mg alloy properties. Models integrating microstructural features (e.g., grain size, DRX, texture) consistently demonstrate higher accuracy and better generalization. Optimization algorithms, such as GA, DBO, and PSO, enhance prediction performance and aid in solving multi-objective problems. DL techniques (CNNs, GRUs, GANs) are proving to be essential when handling complex microstructural data or small datasets, as demonstrated in the work by (Nithin et al., 2022; Byun et al., 2024). ML models replace traditional empirical models (Arrhenius) and accelerate materials discovery, process optimization, and property prediction. ML-driven modeling is transforming alloy design and characterization in the Al-Mg system. Integrating domain knowledge, microstructural data, and advanced ML frameworks is paving the way for high-performance alloy development, rapid prototyping, and reduced experimental burden.

Supplementary Tables SM4 illustrates the application of cutting-edge ML and DL models, CNNs, GBT, and Graph Neural Networks (GNNs) to predict and analyze Al and Mg-based alloys’ properties and microstructural characteristics. Kim et al. (2024) presented a comprehensive ensemble ML framework that integrates ANN, CNN, and RF models to forecast hardness in Al-/Mg-based alloys. The research identified new high-hardness alloy compositions and utilized explainable AI methods to clarify the influence of individual elements and processing parameters. Lopes Marinho et al. (2024) explored various CNN architectures (U-Net, HR-Net, nnU-Net) to assess biodegradation and osseointegration in Mg-based biomedical implants. Nithin et al. (2022) concentrated on grain prediction in Al-Si-Mg alloys using a conventional CNN model and achieved over 90% accuracy in identifying α-Al grains. It confirmed the CNN’s efficiency in image-based microstructural evaluation. Masood Chaudry et al. (2021) focused on structured data and employed GBT to predict hardness in Al-Cu-Mg-x alloys with less-explored compositions. The GBT model leveraged curated literature data and demonstrated exceptional predictive capability.

The studies in Supplementary Tables SM4 highlight the growing incorporation of DL and graph-based models in the design and analysis of alloys’ microstructure. CNNs (Nithin et al., 2022; Lopes Marinho et al., 2024) have confirmed strong performance in image-related tasks, such as segmentation and grain identification, making them essential for the automated analysis of microstructural images. GBT models (Masood Chaudry et al., 2021) excelled in predicting capabilities for structured tabular data when experimental data is scarce but well-organized. The rise of graph-based models, such as HGGAT (Shu et al., 2022), marks a significant advancement in materials informatics. The microstructural intricacies were represented through images or scalar values, as well as relational structures (e.g., grain networks). These models provide a more physics-informed perspective on ML, thereby enhancing generalization and interpretability. Advanced ML models, such as CNNs, GBTs, and graph networks, are adept at handling complex tasks involving image segmentation, pattern recognition, and structure-property relationships. These ML models pave the way for research for automated, explainable, and high-throughput materials.

Figure 6a depicts the most commonly used input parameters of ML models for predicting the mechanical behavior and microstructural features of Al- and Mg-based alloys. Hardness is the most frequently referenced parameter, highlighted in works such as (Jagadeesh and Setti, 2023; Shaban et al., 2023; El-Garaihy et al., 2024; Jain et al., 2024). Since hardness is a critical measure of material strength and durability, it is a logical focus for ML-driven prediction and optimization. The composition was also mentioned (Masood Chaudry et al., 2021; Jain et al., 2024; Zhu Y. et al., 2025) in alloy design and performance enhancement research. Stress and strain for assessing material deformation and flow behavior were prevalent in studies (Murugesan et al., 2023; Bai et al., 2024) Li 2024. Furthermore, grain size, a vital microstructural aspect affecting properties like strength, fatigue, and corrosion resistance, was emphasized in research such as (Jaafreh et al., 2022; El-Garaihy et al., 2024).

Figure 6b presents an overview of the ML models that consistently demonstrate high predictive performance in the literature reviewed. ANN is the most commonly used and top-performing model, cited in over ten studies, including (Bai et al., 2024; Yu et al., 2024; Zhu Y. et al., 2025). ANNs are particularly praised for their adaptability and capacity to model non-linear relationships between input parameters and mechanical outputs. The ML model, Backpropagation ANN (BP-ANN), has also received attention for its effectiveness in modeling flow stress and microstructural evolution, as noted in studies such as (Murugesan et al., 2023; Bai et al., 2024; Li X. et al., 2024). Gradient Boosted Trees (GBT) and SVR, mentioned in (Masood Chaudry et al., 2021; Shaban et al., 2023; Wang Q. et al., 2023; Mishra, 2024a) were recognized for their robustness and interpretability in predicting alloy properties. More complex models, CNNs and GANs, have been effectively utilized in image-based microstructural analysis and data augmentation (Byun et al., 2024; Lopes Marinho et al., 2024), showing impressive generalization capabilities. Furthermore, recurrent models such as LSTM and GRU outperformed traditional ANNs in sequence-based or simulation-derived datasets, as established in (Mokhtari et al., 2024). These trends reflect an increasing preference for DL models with memory or spatial learning functions in complex materials design challenges. A schematic representation of a CNN framework trained to predict tensile strength from electron backscatter diffraction (EBSD) or microstructure images is presented in Supplementary Figure SM1.

“AI in welding and advanced processing techniques,” cluster four studies presented in Supplementary Tables SM10. It emphasizes the role of AI and ML in areas such as materials processing, performance prediction, hydrogen storage, and imaging. Mg-based alloys, including AZ31B, ZK30, WE43, and Mg-Gd alloys, have received considerable attention due to their advantageous properties for both structural and biomedical applications.

Jiang Q. et al. (2024) and Athul et al. (2025) examined Mg hydride systems for hydrogen storage, utilizing density functional theory (DFT) alongside ML to foresee desorption characteristics and storage capacities. Regarding AI/ML methodologies, the studies explore diverse approaches. Conventional algorithms, such as RF and ANN, are employed in (Wiese et al., 2023; Bahari-Sambran et al., 2024), Maqbool et al. (2024), and to predict high-temperature flow stress, corrosion behavior, and mechanical properties, respectively. DL techniques like U-Net and DenseNet are used in (Díaz-Romero et al., 2023; Menze et al., 2024) for image segmentation in micro-CT imaging of bioresorbable scaffolds and sorting aluminum scrap using fused RGB, depth, and LIBS data. Evidence of hybrid models and optimizations is also present: Maqbool et al. (2024) applied virtual sample generation through particle swarm optimization (PSO) to improve ML accuracy, while Yadav and Khurana integrated an artificial neural network (ANN) with a genetic algorithm for optimizing friction stir welding.

Applications are equally diverse and (Yadav and Khurana, 2022; Mishra, 2024a; B et al., 2025) focused on welding and friction stir processing of Mg alloys and predicted tensile strength, wear, and corrosion outcomes based on tool parameters. Österreicher et al. (2024) and Wiese et al. (2023) employed ML models to predict grain structure and anisotropic behavior during extrusion processes. In biomedical domains, (Menze et al., 2024), combined synchrotron-based micro-CT and DL to analyze degradation morphology in coronary implants. Even natural language processing (NLP) found its place (Mishra, 2024b) applied five summarization algorithms to extract key knowledge from FSW literature, evaluating them with the ROUGE metric. The ML framework for predicting post-treatment mechanical properties from thermal parameters, namely, solutionizing temperature, aging time, and cooling rate is depicted in Supplementary Figure SM2.

Cluster five focuses on biomaterials, microstructural optimization, and the incorporation of rare earth (RE) elements in Mg alloys, with an emphasis on both biomedical and structural applications. The reviewed studies span a range of alloy systems (e.g., AZ31B, AZ61, WE43, ZE62, ZK60) and applied a diverse suite of ML and DL techniques, including ANN, CNN, SVM, GP, GPR, KNN, Extra Trees, and AIMD simulations, Supplementary Tables SM11.

Key studies (Li et al., 2021; Raguraman et al., 2024; Ma et al., 2025) highlighted biodegradable implant alloys, with ML models accurately predicting corrosion, fracture, and degradation behavior critical for orthopedic applications. Other works such as (Chen et al., 2020; Zhang et al., 2020; Suh et al., 2022) emphasized process optimization (e.g., ECAP, SPD, aging), improved mechanical properties such as yield strength (YS), ultimate tensile strength (UTS), and elongation.

Additive manufacturing studies (Liu et al., 2020; Yanase et al., 2022) used ML to optimize parameters for densification and microstructure refinement. Farooq et al. (2025) investigated combustion behavior and fire resistance, essential for aerospace alloys.

ML-assisted optimization outperformed traditional experimentation by reducing trials and enhanced predictive accuracy. Chen et al. (2020) achieved a 27% increase in strength using just four experiments guided by ML, while Jain et al. (2025) demonstrated high predictive capability (R² = 97.2%) using ensemble methods. Furthermore, rare earth additions (e.g., Y, Gd, La, Ce) are consistently linked to microstructure refinement and property enhancement, as seen in (Otieno et al., 2024; Wu H. et al., 2025). The SHAP summary plot illustrates the relative contributions of individual features toward predicting the biodegradation rates of magnesium-based implants, as depicted in Supplementary Figure SM3.

A wide array of research centered on AI/ML-based corrosion modelling was conducted (Supplementary Tables SM12). The primary focus is on Mg alloys and related materials, including aluminum, zinc, and high-entropy alloys (HEAs). These studies employed various ML algorithms, including ANN, RF, SVR, XGBoost, DT, GA, and hybrid models, to predict corrosion rates, mechanical performance, hydrogen evolution, grain size, stress, fatigue life, and phase development.

Many studies focused on biodegradable Mg alloys (WE43, AZ80, ZK30, AZ31, ZX10) to enhance their in vitro and in vivo corrosion behavior, mechanical strength, and biomedical applicability. Some research investigated their coating properties, hydrogen storage potential, and the impact of alloy composition and processing parameters on degradation and overall performance. Dargusch et al. (2024) and Maqbool et al. (2024) employed CT image-based ML and virtual sample generation to model hydrogen production and corrosion induced by friction stir processing, respectively.

Alongside experimental data, various studies utilized physics-based simulations such as FEM, DFT, and phase field modeling, enhanced by ML for greater predictive accuracy. Techniques such as active learning, SHAP analysis, and multi-objective optimization were employed to enhance feature selection and facilitate the interpretation of model outcomes, thereby bridging the gap between data science and materials engineering. In addition to Mg alloys, research on Al-Mg-Si, Zn-based biodegradable alloys, LHEAs, and Al–Li aerospace alloys illustrates the broad applicability of ML in studying corrosion and property trade-offs across materials. Some works, such as (Zhu et al., 2024; Liang et al., 2024), also feature Bayesian learning and explainable AI, showcasing sophisticated hybrid frameworks for designing and forecasting the performance of smart materials. An LSTM-based time-series model for predicting the evolution of grain size in magnesium alloys under dynamic thermal conditions is illustrated in Supplementary Figure SM4.

Cluster seven studies in Supplementary Tables SM13 highlight the increasing influence of ML and data-driven optimization in materials design, considering Mg alloys. Many of these works utilized ML algorithms, including XGBoost, SVR, neural networks, RFs, and ensemble models, to predict and enhance mechanical, electrochemical, and thermal properties.

In Mg alloys, multiple studies (El-Sanabary et al., 2024; Ling et al., 2024; Mi et al., 2024; Liu X. et al., 2025) investigated aspects such as anode performance, corrosion resistance, and mechanical behavior through ML-guided design. Ling et al. (2024) and Liu X. et al. (2025) integrated ML with multi-objective optimization strategies, resulting in high-performance alloy compositions, such as Mg-Ca and Mg-Al-Zn-Cu. Mi et al. (2024) employed Bayesian active learning to develop wrought alloys with outstanding tensile properties, whereas El-Sanabary et al. (2024) improved ECAP processing parameters for wear resistance by combining ML and DOE- GA approaches. Çolak (2025) and Cheng et al. (2025) expanded ML’s application to high-entropy alloys and bulk metallic glass (BMG), respectively, using ANNs and stacking ensemble learning alongside SHAP analysis for enhanced model interpretability. These studies prioritized model accuracy and interpretability in forecasting complex properties such as discharge capacity and glass-forming ability.

In addition to Mg alloys, (Li H. et al., 2024; Dong et al., 2023), applied feature selection and descriptor-based regression to design Al–Li and hydrogen storage alloys, respectively, illustrating how ML supports compositional screening. Dong et al. (2023) utilized SHAP values to inform element selection and forecasted high-capacity Mg-Sm and Mg-Ni-Sm alloys.

Several investigations emphasized hybrid methodologies. For example, Gupta et al. (2021) and Jafari et al. (2015) merged fuzzy logic, Taguchi methods, and genetic algorithms to optimize manufacturing processes like pressure die casting and L-bending, while (Song et al., 2019) combined backpropagation neural networks with the gray wolf optimization algorithm (GWOA) to enhance weld strength in dissimilar metal joints.

Finally, Shimanek et al. (2022) and Jafari et al. (2015) bridged atomic-scale descriptors and macro-scale properties. (Shimanek et al., 2022). proposed atomic features for ML modeling of shear strength, while Jafari et al. (2015) combined multi-objective evolutionary algorithms with ML to predict age-hardening behavior in Al alloys, showing that preprocessing significantly improves model accuracy. Supplementary Figure SM5 illustrates a data-driven optimization framework combining experimental data, simulation outputs, and literature sources with ML and optimization algorithms.

The transformative potential of AI, ML, and DL in Mg alloy research is most evident in their capacity to uncover previously inaccessible structure property relationships. For example, CNNs have been effectively trained on EBSD maps to predict tensile strength and fracture toughness, capturing textural and grain boundary features beyond the reach of conventional statistical methodologies. This obviates the necessity for manual feature engineering, facilitating the extraction of insights from high-dimensional microstructural datasets.

Similarly, unsupervised learning methods and representation learning frameworks, such as autoencoders, demonstrate potential for distilling complex microstructural images into latent vectors that correlate with mechanical and corrosion performance. RF and SVM regression models have also been applied in biomedical contexts to relate alloy composition and processing parameters to biodegradation rates, thereby offering actionable clinical guidance for the design of magnesium-based implants.

Unlike linear modelling, AI techniques can capture non-linear, multi-parametric interactions, including the synergistic effects of alloying elements (Zn, Ca, Mn) on ductility and corrosion resistance. Modelling such intricate interactions through first-principles or empirical equations presents considerable challenges.

AI and ML models play an increasingly pivotal role in accelerating computationally intensive simulations. ML potentials trained on high-level ab initio data, such as DFT, can reproduce PES features with significantly reduced computational costs. For example,:

• Deep neural network potentials (e.g., DeepMD, SchNet) can achieve speedups over DFT by factors ranging from 10³ to 10⁵ while maintaining chemical accuracy (He et al., 2023).

Key benchmarking metrics for evaluating the efficiency of AI-driven simulation methods relative to classical algorithms include:

• Wall-time per configuration,

The importance of model interpretability is contingent upon the application domain. In PES prediction and atomic-scale simulations, predictive accuracy and scalability typically take precedence over interpretability. Conversely, in biomedical applications particularly where Mg alloys are deployed as transient implants high interpretability is essential to ensure safety, trust, and regulatory compliance. Illustrative approaches include:

• SHAP for understanding feature contributions in property predictions Zhang H. et al. (2024).

Future research should guide model selection through a contextual trade-off framework balancing interpretability, predictive accuracy, and application-specific risk considerations.

The reviewed literature encompasses a wide range of AI models, including ANNs, CNNs, SVMs, RFs, Gradient Boosting, LSTM, GNNs, and Transformer-based models. While this diversity may reflect exploratory research and the pursuit of novelty, there are substantive methodological reasons underlying model selection:

• Regression tasks with limited datasets are better addressed using ANNs and SVMs.

This methodological diversity is justified along the axes of data modality, scale, and task complexity. Nevertheless, future research should strive for systematic benchmarking and avoid indiscriminate “algorithm hopping” in the absence of performance comparisons.

Despite the promising capabilities of AI in Mg alloy research, several technical and practical challenges hinder its broader adoption:

• Data Scarcity: Many studies rely on small or synthetic datasets, lacking real-world variability.

Addressing these limitations necessitates:

• Development and dissemination of open-source datasets and standardized benchmarks,

Collectively, these strategies can ensure that AI, ML, and DL tools in magnesium alloy research evolve from promising experimental techniques into robust, industry-relevant solutions.

To transition from academic modeling to industrial application, this review identifies several actionable research directions. Supplementary Tables SM14 summarizes recommended AI approaches, real-world deployment challenges, and proposed technical solutions for various Mg alloy applications, including automotive, biomedical, high-temperature, and corrosion-resistant systems.

In addition to established methodologies in domains such as property prediction and PES modeling, AI and ML techniques demonstrate significant potential in addressing various facets of Mg alloys within emerging research areas. Surrogate models derived from ab initio thermodynamic data offer the capability to enhance free energy calculations, which are critical for accurately determining phase stability and transformations. DL frameworks such as GNNs and Variational Autoencoders (VAEs) have gained prominence for their capacity to learn Gibbs free energy surfaces at a substantially reduced computational cost, thereby facilitating more rapid construction of phase diagrams and enabling efficient thermodynamic modeling. Furthermore, diffusion processes, vacancy dynamics, and solid–liquid interfacial energies traditionally computationally demanding to simulate using conventional techniques are increasingly becoming accessible through AI-driven methods. These advancements hold the potential to fundamentally transform the data-driven exploration and optimization of multicomponent alloy systems.

The incorporation of AI, ML, and DL into Mg alloy research has yielded significant progress. Nevertheless, various ongoing challenges still impede broad adoption, scalability, and practical application. Table 2 presents a structured synthesis of challenges and proposed solutions in recent AI/ML/DL-driven studies on magnesium alloy design. It categorizes key issues such as model accuracy, data limitations, property trade-offs, and process complexity. It aligns them with mitigation strategies, including generative models, active learning, hybrid optimization, and explainable AI. This comprehensive overview highlights current gaps and methodological advancements across structural, biomedical, and energy-related Mg alloy applications. Figure 8 illustrates the significant challenges associated with applying AI, ML, and DL techniques in Mg alloy research and identifies corresponding potential solutions.

Surrogate modeling is a key enabler to minimize the computational expense of high-fidelity simulations and experimental campaigns in materials science. Of the several surrogate models available, Gaussian Process Regression (GPR) stands out because it can offer uncertainty quantification in addition to predictions, and hence can be used in applications such as alloy composition design, fatigue life prediction, and microstructure-property mapping. Likewise, neural network surrogates, intense feedforward networks and convolutional surrogates, are being more and more employed to emulate costly simulations (DFT, FEM) for fast predictions in high-dimensional design spaces (Misener and Biegler, 2023).

Aside from single-fidelity surrogates, multi-fidelity modelling integrating low-fidelity physics-based models with high-fidelity experimental or computational data using ML methods has become a powerful paradigm. These models take advantage of the speed of low-fidelity physics-based simulations and correct errors using machine learning trained on high-fidelity data. Blending a phase-field model with a neural surrogate learned from experimentally observed grain evolution data will optimize heat-treatment schedules or alloying strategy without sacrificing physical coherence. This layered modeling strategy improves prediction accuracy and optimization efficiency, and it is consequently very well-adapted for real-time process control and alloy design (Zhang L. et al., 2022). Incorporating such frameworks into optimization approaches (e.g., in mechanical modeling or corrosion prediction) can greatly promote inverse design, uncertainty-based decision-making, and adaptive experimentation in Mg alloy research.

Recent advancements in Mg alloy research reveal several key trends shaped by the integration of AI, ML, and DL techniques:

• Transition from Empirical to Data-Driven Modeling: ML models, most notably artificial neural networks (ANNs), backpropagation neural networks (BPNNs), and ensemble methods, are increasingly preferred for predicting intricate mechanical behaviors, such as flow stress, fatigue life, and elongation. Microstructural attributes, such as grain size, crystallographic orientation, and phase distribution, are being increasingly utilized in these models.

Despite promising progress, several critical research and methodological gaps persist:

• Data Availability and Standardization: The lack of large-scale, standardized datasets limits training, benchmarking, and cross-validation. Most data is either siloed or proprietary and limits reproducibility and generalization.

To accelerate innovation and practical implementation of AI/ML/DL in Mg alloy research, the following directions are recommended:

This review employs a robust multi-database validation approach by utilizing both Scopus and Web of Science, thereby enhancing the reliability, reproducibility, and generalizability of the findings. While Scopus was selected as the primary source for in-depth cluster analysis due to its broad coverage in materials science and computational modeling, cross-validation with WoS confirmed a high degree of thematic consistency. Minor differences in author and journal indexing were observed, primarily due to database-specific inclusion policies, which may result in slight variations in coverage.

Although the review focused on English-language journal articles, potentially excluding some regional studies, the inclusion criteria ensured data quality and compatibility with bibliometric tools. Future studies could further expand the scope by incorporating additional databases such as Dimensions or PubMed and by integrating alternative metrics, including patent data and conference proceedings, to capture a more comprehensive view of the research landscape.

This review employed a co-occurrence-based clustering methodology for thematic mapping using VOSviewer; however, alternative clustering paradigms may offer supplementary insights. For example, application-oriented clustering, such as corrosion prediction, mechanical property modeling, or microstructural evolution, could enhance synthesis by aligning findings with specific problem domains. Similarly, method-centric groupings based on approaches such as ANN, CNN, GNN, or ensemble models may facilitate the identification of algorithmic trends and existing gaps within the field. Future review analyses could be refined through the adoption of hybrid clustering frameworks that integrate topic, technique, and data modality dimensions, thereby yielding more granular and actionable insights to inform and direct research development within targeted domains.