Machine Learning for Shape Memory Alloy Property Optimization

Abstract

One of the obstacles to the deployment of shape memory alloys (SMAs) in solid-state actuation is the low efficiency and functional instability due to the transformation thermal hysteresis and large temperature ranges during martensitic phase transformation. Numerous studies have been conducted in an effort to minimize the thermal hysteresis and transformation temperature range of SMAs through ternary and quaternary alloying of known binary alloy systems, such as NiTi, and considerable success has been achieved. However, and crucially, the alloys discovered so far have failed to maintain a narrow hysteresis under applied stress. In the present study, an AI-enabled materials discovery framework was successfully used to identify both SMA chemistries and the associated thermo-mechanical processing steps that result in narrow transformation hysteresis and transformation range under an applied stress. The major elements of the proposed workflow are described in detail and its materials-agnostic character makes it widely applicable to other alloy discovery challenges. Following this framework, a large, high-quality SMA dataset is developed for use in data-enabled alloy design. The dataset is then used to train machine learning models and candidate alloy predictions are selected based on expected improvement. Using this framework, and without relying on subsequent experimental exploratory analysis, an SMA composition, i.e. Ni32Ti47Cu21 (at. %), was predicted and confirmed to have the narrowest thermal hysteresis and transformation range under stress achieved thus far for a NiTi-based SMA. Furthermore, the alloy was shown to exhibit excellent cyclic stability and actuation strain. Additionally, a new descriptor is created with high feature importance and is used in training new machine learning models which are experimentally validated for extrapolative prediction accuracy with quaternary alloy predictions in a fractional-factorial experimental design. The methodology and the dataset introduced here can be extended to design novel SMAs with other target functions.