Design of high-temperature NiTiCuHf shape memory alloys with minimum thermal hysteresis using Bayesian optimization

Chemical composition and thermal processing parameters are used in a first-of-their-kind machine learning (ML) and batch Bayesian optimization (BBO) approach in an iterative fashion in the quaternary NiTiCuHf high-temperature shape memory alloy (HTSMA) composition space to minimize thermal hysteresis in a desired transformation temperature range. The first of three iterations exploited an existing SMA database of lower complexity alloys (binary and ternary) attempting to optimize quaternary NiCuTiHf chemistry and thermal processing for the given constraint and the objective. Alloy synthesis and characterization revealed that the initial ML model displays high error levels between the predicted and experimental values, indicating the need for high-fidelity data in the complex quaternary alloy design space for optimization. The second iteration used this conclusion to explore an expanded design space through tuning Gaussian process (GP) hyperparameters. Utilization of active learning enabled the enlargement of data present in the high-complexity space during the iterative process, improving model accuracy. The third iteration discovered NiTiCuHf HTSMAs with the lowest reported martensitic transformation thermal hysteresis with transformation temperatures between 250 °C and 350 °C to date without precious metals. The effects of optimized secondary heat treatments on the martensitic transformation characteristics were explored and compared to those achieved after the initial homogenization heat treatments to demonstrate the ability of the BBO framework to create optimal alloys with controlled chemistry and thermal processing. In Ni-rich compositions of the designed alloys, the secondary heat treatments suggested by the BBO framework resulted in significant increases in transformation temperatures, suggesting the formation of Ni-rich precipitates.