Source research

Review of Neural Network Modeling of Shape Memory Alloys

Rodayna Hmede, Frédéric Chapelle, Yuri Lapusta · 2022 · 62 citationsRead the paper

Shape memory materials are smart materials that stand out because of several remarkable properties, including their shape memory effect. Shape memory alloys (SMAs) are largely used members of this family and have been innovatively employed in various fields, such as sensors, actuators, robotics, aerospace, civil engineering, and medicine. Many conventional, unconventional, experimental, and numerical methods have been used to study the properties of SMAs, their models, and their different applications. These materials exhibit nonlinear behavior. This fact complicates the use of traditional methods, such as the finite element method, and increases the computing time necessary to adequately model their different possible shapes and usages. Therefore, a promising solution is to develop new methodological approaches based on artificial intelligence (AI) that aims at efficient computation time and accurate results. AI has recently demonstrated some success in efficiently modeling SMA features with machine- and deep-learning methods. Notably, artificial neural networks (ANNs), a subsection of deep learning, have been applied to characterize SMAs. The present review highlights the importance of AI in SMA modeling and introduces the deep connection between ANNs and SMAs in the medical, robotic, engineering, and automation fields. After summarizing the general characteristics of ANNs and SMAs, we analyze various ANN types used for modeling the properties of SMAs according to their shapes, e.g., a wire as an actuator, a wire with a spring bias, wire systems, magnetic and porous materials, bars and rings, and reinforced concrete beams. The description focuses on the techniques used for NN architectures and learning.

What the full text shows

Working prototype

LSTM NN with a regression layer modeled antagonistic SMA systems with low error and fast prediction. — RMSE < 2%, training time < 3 min, prediction time in ms (Antagonistic SMA system (two identical wires attached at a midpoint))
LM algorithm-based NN achieved high precision in controlling Nitinol wire position. — Global minimum error of 2–3 µm (Nitinol wire (diameter 0.15 mm, length 75 mm) using 3000 epochs)
GRNN with VIKOR-Fuzzy logic optimized electrical discharge machining of Ni-Ti SMA specimens. — ±5% error (54 datasets (60% training, 20% validation, 20% testing))
HFLANN (FLANN modified by PSO and NARX) used as a predictor controller for SMA wire and steel spring. — Error ≤ 0.02 mm (Compared to sinusoidal and multistep predictors)
PID NN using backpropagation characterized hysteresis loops for SMA and steel spring. — 0.0073 mm error (major loop); 0.0101 mm error (major and minor loops) (Magnetic approach control)
Multilayer NN using electrical resistance (ER) feedback to determine SMA displacement. — Approximate error of 7% (Ni-Ti single wire (length 228.6 mm, diameter 0.381 mm), 3-layer structure with 8 hidden neurons)

Setup: Review of multiple studies involving Nitinol (Ni-Ti) and Flexinol wires, varying diameters (0.15 mm to 0.381 mm), and different mechanical configurations (single wire, spring-biased, antagonistic systems).
How: The paper reviews various ANN architectures (Feedforward, NARX, GRNN, Jordan-Elman, LSTM) and training algorithms (Backpropagation, LM, PSO) used to model the nonlinear behavior of SMAs. It categorizes these models based on SMA geometry (wires, rings, bars, porous materials) and application types.
Limitations: Many studies do not justify the choice of training or learning algorithms. · Hysteresis effects cause higher errors in some models (e.g., 7% error in ER-based displacement). · Lack of adequacy studies between specific NN architectures and diverse SMA applications.

Summarized by Gemma 4 from the open-access full text · 7 Oct 2026. Read it