From microstructure to instability: a multimodal data-driven workflow for leakage-aware screening of memristive devices

Stress-induced microstructural and interfacial changes are widely implicated in the electrical instability of memristive devices, yet they are rarely quantified within a single leakage-aware predictive framework. Here we present a proof-of-concept small-data study that integrates registered pre- and post-stress SEM differencing, DC current–voltage hysteresis descriptors, and illumination-sensitive electrochemical impedance spectroscopy features to predict a device-level electrical instability proxy. All analyses are performed using strict device- and batch-aware grouping to prevent cross-device leakage and to evaluate cross-batch generalization. Multimodal fusion improves predictive performance relative to single-modality baselines under device-grouped validation, whereas leave-one-batch-out analysis reveals heterogeneous transfer across fabrication batches. Grouped permutation testing and bootstrap uncertainty estimates support a modest average cross-batch benefit while highlighting substantial fabrication-dependent domain shift. Feature attribution identifies complementary predictive signals across modalities: SEM texture and Fourier descriptors summarize stress-associated image changes, DC metrics capture switching asymmetry, and impedance phase features reflect illumination-sensitive interfacial response. These attributions are interpreted as physically motivated, hypothesis-generating indicators rather than direct evidence of unique causal mechanisms. Together, these results establish a reproducible workflow for instability-oriented screening and batch-level process assessment in memristive materials, while defining a small-data benchmark for future validation against direct reliability endpoints.