Accelerated Cu-Ti alloy design under data scarcity through fine-tuned transfer learning

• Layer-wise unfreezing transfer learning models achieve R 2 > 0.93 for Cu-Ti properties with only 40 samples. • Bayesian optimization identifies optimal CTFC alloy (Cu-4Ti-0.3Fe-0.2Cr) within three active learning iterations. • The CTFC alloy exhibits 1107 MPa UTS, 20.8 % IACS EC by dual dislocation/precipitate strengthening. This study presents a novel machine learning framework using transfer learning (TL) to overcome data scarcity in designing high-performance Cu-Ti alloys with balanced high ultimate tensile strength (UTS) and electrical conductivity (EC). We leveraged a pre-trained model built on a large dataset of multi-system copper alloys and transferred this knowledge to the target Cu-Ti system with only 40 samples. Two TL strategies here were developed: first-layer freezing (TL1) and layer-wise unfreezing (TL2). For the Cu-Ti dataset, TL1 froze only the first hidden layer during training, while TL2 employed a gradual layer-wise unfreezing method. The TL2 models for Cu-Ti alloys achieve superior prediction accuracy (UTS model: R 2 = 0.96, EC model: R 2 = 0.93). Such TL models significantly outperform conventional machine learning models with R 2 < 0.77. The optimized TL2 model guides a Bayesian optimization-driven inverse design with active learning, rapidly identifying an optimal alloy—Cu-4Ti-0.3Fe-0.2Cr (CTFC)—within three iterations. Experimental validation confirms CTFC’s exceptional properties: UTS = 1107 MPa (1.8 % prediction error) and EC = 20.8 % IACS (2.8 % prediction error). Microstructural analysis reveals that the synergistic precipitation of nanoprecipitates β′-Cu 4 Ti, Fe 2 Ti, and Cr enhances both strength and electrical conductivity. This machine learning framework suggests an efficient pathway for designing high-performance materials under constrained data conditions.