Machine learning-assisted reconstruction of process-dependent metastable phase diagrams in magnetron-sputtered nitride thin films

Metastable phase formation critically determines the microstructure, phase stability, and functional performance of magnetron-sputtered nitride thin films. However, constructing process-dependent metastable phase diagrams remains challenging because experimental mapping is labor-intensive, while thermodynamic calculations alone cannot fully describe kinetically constrained deposition processes. Here, a two-step machine learning-assisted sequential dual transfer learning (SDTL) framework is proposed for process-dependent prediction of metastable phase formation from limited experimental data. The framework combines SDTL-based phase classification with boundary-temperature (T c ) regression for continuous reconstruction of metastable phase diagrams. Transferable elemental and process-related knowledge are sequentially integrated to classify metastable phase regions in TiAlN and VAlN films, after which the predicted phase boundaries are converted into T c data for descriptor-based regression. The SDTL classifier achieves a phase-classification accuracy of 0.956, while the T c regression model attains a test-set R 2 of 0.996. The proposed strategy shows excellent agreement with experimentally reported metastable phase diagrams and accurately reconstructs process-dependent phase diagrams over an extended temperature range of 50–2000 °C. Feature analysis identifies mean molar mass and mean atomic radius as the dominant descriptors governing boundary-temperature evolution. This work establishes an efficient strategy for machine learning-assisted construction of process-dependent metastable phase diagrams and provides a practical route for accelerating the design and understanding of advanced nitride thin films.