An informatics approach to transformation temperatures of NiTi-based shape memory alloys

Abstract

The martensitic transformation serves as the basis for applications of shape memory alloys (SMAs). The ability to make rapid and accurate predictions of the transformation temperature of SMAs is therefore of much practical importance. In this study, we demonstrate that a statistical learning approach using three features or material descriptors related to the chemical bonding and atomic radii of the elements in the alloys, provides a means to predict transformation temperatures. Together with an adaptive design framework, we show that iteratively learning and improving the statistical model can accelerate the search for SMAs with targeted transformation temperatures. The possible mechanisms underlying the dependence of the transformation temperature on these features is discussed based on a Landau-type phenomenological model.

High-Dimensional Materials and Process Optimization Using Data-Driven Experimental Design with Well-Calibrated Uncertainty Estimates

“…Ueno et al [10] presented a Bayesian optimization framework which they applied to the case of determining the atomic structure of crystalline interfaces. Xue et al [11] investigated the use of SL for discovering shape memory alloys with high transformation temperatures. They used a simple polynomial regression on three material parameters to drive their predictions.…”

Active learning in materials science with emphasis on adaptive sampling using uncertainties for targeted design

“…However, few focus on aspects of making the optimal next decisions for synthesis and characterization by experiments or calculations. [27][28][29][30][31][32][33][34][35][36] The predictions from machine learning are not necessarily optimal. Aspects related to multiscale modeling and constitutive response at the engineering design scale are discussed in the book by McDowell et al 37 In contrast, the approaches we will discuss are based on an active learning or adaptive design paradigm, whereby the predictions from a surrogate model, which can be an inference model or a physics-based or reduced order model (ROM), serve as the input to define a utility or acquisition function, the optimal of which dictates the next experiment or calculation to be performed.…”

“…Systematic design strategies aid in guiding or recommending iteratively the optimal compounds for synthesis. The strategies discussed above have been applied to alloys 27,29 and ceramic materials. 30,35,36 Example 1: piezoelectrics with large electrostrains.…”

Accelerated Search for BaTiO3‐Based Ceramics with Large Energy Storage at Low Fields Using Machine Learning and Experimental Design

“…[19][20][21][22] The strategy has been quite successful in finding alloys and ceramics with enhanced properties. [11][12][13][14]23,24] Even though Bayesian based algorithms using utility functions, such as expected improvement to maximize or minimize the objective or property, have proved especially successful in finding compounds with desired properties, the search space is often too large and leads to excessive exploration. This is a reflection of a multidimensional search space, and what is needed is to isolate regions in this space containing a relatively large number of extrema so that the probability of finding a compound with the desired targeted response is much higher than in the original search space.…”