Machine learning–guided design of shape memory composite: balancing piezoelectricity and mechanical recovery in PVDF/SMPU/BTO

PVDF-based multifunctional composites capable of simultaneously delivering piezoelectric energy harvesting and shape memory behavior are highly desirable for next generation flexible electronics, soft robotics and wearable system. In this study a data – driven design approach combining Taguchi L27 design of experiments and machine learning (ML) was employed to optimize Polyvinylidene fluoride (PVDF) /Shape memory polyurethane (SMPU)/ Barium Titanate (BTO) nanocomposite. Structural characterization confirmed that BTO promotes β-phase nucleation and enhance crystallinity, while SMPU improves flexibility and elastic recover. SMPU-rich composites exhibited excellent shape recovery (> 90%), whereas PVDF/BTO-rich formulation delivered high piezoelectric output voltage (up to 12 V), revealing a trade-off between actuation and energy harvesting. An optimized composition (85 wt% PVDF, 5 wt% SMPU, 10 wt% BTO) achieved a balance tensile strength of 21 MPa with multifunctional functional performance. ML models were developed to establish structure-property relationships yielding R 2 = 0.81 for shape recovery and R 2 = 0.75 for output voltage using leave-one-out-cross-validation (LOOCV). Gradient boosting showed higher apparent accuracy for shape recovery (R 2 = 0.96), while ridge regression provided more stable voltage predictions. Feature importance analysis identified SMPU as dominant factor for shape recovery, whereas PVDF and β-phase content governed piezoelectric response. This study demonstrates an integrated experimental –ML framework for optimizing multifunctional polymer composites for advanced flexible device application.