Computationally Efficient Torque Estimation in SMA Microactuation Systems: A Neural Network Surrogate Trained on Parametric FEM Data – Rundong Jia

Rundong Jia
WW8, FAU

21. Juni 2025, 17:00
WW8, Fürth

The design of complex micro-origami systems actuated by bi-directional Shape Memory Alloy (SMA) components is a significant challenge, often hindered by the high computational cost of traditional Finite Element Method (FEM) simulations for parametric design space exploration. This thesis presents a neural network (NN) surrogate model which predicts the non-linear torque-angle characteristics of SMA self-bending microactuators with significantly lower computational cost.

A comprehensive parametric dataset was first generated using a validated FEM framework in Abaqus, which employed a UMAT subroutine to capture the thermo-mechanical behavior of NiTi actuators. The study systematically varied key geometric parameters (bending radius, width, thickness) and operational conditions (maximum bending angle for passive loading-release, temperature for active actuation). Subsequently, a feedforward neural network was trained on this dataset to map the input parameters to the resulting torque curves for both passive (antagonist) and active (protagonist) modes.

The results demonstrate that the trained NN achieves excellent predictive accuracy while preserving good robustness. Specific case studies further verify that the surrogate model replicates well the non-linear torque-angle curves for both the martensite loading-release and the austenite actuation processes.

The validated neural network serves as a surrogate model, enabling rapid and accurate performance prediction. Crucially, by intersecting the predicted torque curves of the antagonist and protagonist, the equilibrium point of the actuated bi-directional system can be determined, allowing for the direct prediction of the balance angle for any given parameter configuration without new FEM simulations. This work provides a powerful framework for accelerating the design, system-level optimization, and rapid exploration of the vast parameter space for advanced, bi-directional SMA-based micro-actuation systems.

Keywords: Active Origami, Shape Memory Alloy (SMA), Bi-directional Actuator, Finite Element Method (FEM), Machine Learning (ML), Neural Network (NN)