Thermally Pre‐Formed Reconfigurable Resistive Random‐Access Memory Crossbar Arrays: A Dual‐Mode Platform for Robust Physically Unclonable Functions and In‐Memory Computing

ABSTRACT Intelligent edge hardware requires both high‐entropy physical signatures for security and highly uniform devices for accurate In‐Memory Computing (IMC), yet conventional Resistive Random Access Memory (RRAM) relies on electroforming that irreversibly damages the dielectric and fixes the filament state. A forming‐free RRAM platform is demonstrated in which Thermally Pre‐formed Filaments (TPFs) are generated by solid‐state oxygen scavenging between an oxygen‐deficient TiO x reservoir and an ultrathin Al 2 O 3 switching layer. Thermodynamically driven, reversible ionic migration forms percolation‐threshold conduction networks that produce a device‐unique distribution of initial resistance states, enabling Physically Unclonable Functions (PUFs). A reverse‐bias reset fully erases the conductive network and restores a pristine‐like high‐resistance state, allowing the same array to be reconfigured into a uniform analog synaptic platform without additional forming. In a 24 × 24 crossbar array, the PUF responses pass the National Institute of Standards and Technology (NIST) SP 800–22 test suite, and the reconfigured array shows high‐fidelity multilevel conductance programming for neuromorphic IMC. This reversible transition from stochastic security primitives to uniform analog computing establishes a damage‐free route toward reconfigurable secure Artificial Intelligence (AI) hardware.