A Physics-Regularized Neural Surrogate Framework for Printed Memristors

Printed memristors offer a pathway to low-cost, flexible neuromorphic hardware, but their variability and filamentary switching pose challenges for fast, accurate device modeling. We introduce a Physics-Regularized Neural Surrogate (PRNS) for printed memristors, combining data-driven learning with physics-based regularization terms (masked Ohmic/Space-Charge Limited Current (SCLC) constraints and Ordinary Differential Equation (ODE) consistency). While classical Physics-Informed Neural Networks (PINNs) enforce governing equations through differentiable Partial Differential Equation (PDE)/ODE residuals, here physical relations act as regularization rather than hard constraints. The model embeds Ohmic conduction, SCLC, and interfacial transport equations into network training while explicitly handling device-to-device variability via parameter perturbation and noise injection. We validate the approach on a literature-calibrated synthetic dataset representative of flexographically printed Ag/PMMA:PVA/ITO devices, achieving relative Root Mean Square Error (RRMSE) of 0.0633 versus 0.2730 for the Voltage-ThrEshold Adaptive Memristor (VTEAM) model (4.31× improvement). Cross-validation yields RRMSE 0.115 ± 0.062, with stability under 10% added noise. Extended validation across digitized I-V curves from three printed memristor technologies (inkjet-printed IGZO, aerosol-jet MoS₂, paper-based MoS₂/graphene) demonstrates competitive or superior accuracy versus VTEAM, Yakopcic, Stanford-PKU, and MMS models. Supplementary experiments (15 tests) confirm comprehensive reliability modeling including temperature-dependent retention (activation energy Eₐ = 0.379 eV), endurance degradation (660-cycle lifetime), and electro-thermal coupling effects. Circuit integration via one-transistor-one-resistor (1T1R) memory cell simulation shows 29% lower write energy (94.2 pJ vs. 133.6 pJ) with gradual analog switching. The formulation supports SPICE-compatible lookup table (LUT)/behavioral export; all code and datasets are released openly for reproducibility.