Thermodynamics-constrained generative adversarial network for synergistic strength-ductility inverse design of rare-earth magnesium alloys

Data-driven methodologies are reshaping magnesium alloy research, yet resolving the critical “strength-ductility trade-off” in high-dimensional composition spaces remains hindered by several unsolved bottlenecks in existing data-driven models for rare-earth magnesium alloy design, such as data sparsity in high-performance regions, and lack of thermodynamic feasibility and physical interpretability. To address these issues, this study proposes a novel Thermodynamics-Constrained Generative Adversarial Network (TC-GAN) framework, which integrates physical metallurgy rules with generative AI via a synergistic “Data Augmentation-Property Prediction-Inverse Design” pathway to realize end-to-end inverse design of manufacturable alloys. The key novelties and quantitative contributions are threefold: (i) A distribution-aware data augmentation module (PIDAM) is developed to reconstruct high-dimensional sparse data manifolds, expanding the original dataset from 860 to 2060 thermodynamically feasible high-fidelity samples; (ii) A sequential attention-based TabNet deep ensemble predictor (TabNetDP) is constructed, achieving high-precision prediction of mechanical properties for multi-component magnesium alloys, with a test-set R² > 0.95 and rigorous uncertainty quantification; (iii) For the first time, Hume-Rothery rules and phase equilibrium constraints are embedded into the generative adversarial network via the thermodynamics-constrained generator (MOTCG), establishing a “physical fence” mechanism that boosts the thermodynamic feasibility rate of generated compositions to 98.2%. Applied to the Mg-Gd-Y-Zn-Zr system, TC-GAN designs significantly transcend the existing Pareto front: the optimal alloy achieves an ultimate tensile strength of 485 MPa and an elongation of 12.5%, outperforming the commercial WE43 alloy by 18.3% in strength and 47.1% in ductility. SHAP analysis and CALPHAD calculations verify the underlying synergistic strengthening mechanism, establishing a universal physics-data fusion paradigm for high-performance multi-component alloy design.