Design of PMMA-Cotton Composite Textile with Tunable Properties via a Physics-Aware Bidirectional Neural Network...

We present a vacuum-assisted Polymethyl methacrylate (PMMA) impregnation process for cotton textiles, coupled with a physics-aware bidirectional artificial neural network (ANN) framework, to both predict and tune the natural fiber composite response from a compliant and flexible to a stiff and strong behavior. Cotton fabric samples were impregnated with acetone-borne PMMA baths, ranging from 0 to 5 wt.% polymer concentration. After drying, the PMMA formed conformal fiber coatings and inter-fiber bridges, with optimal load transfer observed at approximately 0.5-1.0 wt.%. Mechanical properties, including the elastic modulus, tensile strength, ductility, and toughness, were measured alongside Differential Scanning Calorimetry (DSC), Glass Transition Temperature (T_g), Change in heat capacity at constant pressure (ΔCp), gravimetry, and morphology tests. Rule-of-mixtures, porosity, and thermal constraints were embedded as regularization within the ANN loss functions to improve the physical consistency of the training. The forward and inverse models achieved sub-percent prediction errors with narrow bootstrap confidence intervals. It was found that removing physics regularization notably increases forward model error (by fivefold), as well as the inverse model error by one order of magnitude.