• A physics-informed neural network (PINN) framework is proposed for interpretable modeling of the thermo-induced deformation behavior of LCE artificial muscles. • A thermodynamic model incorporating the coupling between liquid crystal free energy and rubber elastic energy is established and embedded into the PINN as physical constraints. • A novel electrothermally actuated LCE artificial muscle with integrated microscale electrothermal structures is designed and fabricated via a friction-based process. • Cycle-level K-fold cross-validation demonstrates that the proposed PINN achieves higher prediction accuracy and stability than conventional data-driven LSTM models under limited data conditions. Thermally-driven artificial muscles are widely investigated for soft robotic and flexible mechanical systems due to their large deformation, lightweight structure, and low power consumption. However, strong nonlinearity between driving temperature and deformation response makes reliable and interpretable modeling challenging, especially under limited experimental data. To address this challenge, a physics-informed neural network (PINN) is proposed to model the thermally induced deformation of liquid crystal elastomer (LCE) artificial muscles. A thermodynamic model considering the coupling between liquid crystalline free energy and rubber elastic energy is incorporated into the network training as physical constraints, enabling interpretable prediction of thermal actuation behavior. A novel electrothermally driven LCE artificial muscle is fabricated as the experimental platform by forming microscale rough structures on the LCE surface via a friction-based process and infilling multiwalled carbon nanotubes (MWCNT) to construct an integrated electrothermal layer, achieving controllable Joule-heating-induced deformation. Experimental results show that the proposed PINN model accurately predicts the thermo-induced deformation of LCE artificial muscles while maintaining physical interpretability. The model achieves an R 2 of 95.01% in K-fold cross-validation and 93.52% in validation experiments with different samples and driving conditions. This work provides an effective framework for the modeling and control of electrothermally driven LCE artificial muscles.
Zhou et al. (Sun,) studied this question.