Hybrid optical-digital polymorphic computing via volumetric scattering complexity for hardware-defined context switching

Conventional optical computing architectures often rely on static physical layers or bulky optical systems, limiting their practicality for compact, cost-sensitive edge applications. We introduce a reconfigurable hybrid opto-electronic processing architecture in which a monolithic device performs voltage-programmable feature encoding. Specifically, we realize a physically reconfigurable optical encoder using an electrically tunable liquid crystal-polymer composite (LCPC). By exploiting the volumetric reorientation of liquid crystal domains via a single scalar voltage control, we instantiate unique random scattering kernels that map the same input to statistically distinct output speckle fields. In a hybrid opto-electronic prototype, we demonstrate that this single optical frontend executes multiple, statistically independent encoding operations, enabling hardware-defined context switching. A unified neural network recovers task labels with ∼90 % accuracy under matched voltages, while crosstalk under mismatched voltages is effectively suppressed, demonstrating native physical functional specificity. The final output of this hybrid system is a class decision (semantic label) rather than a reconstructed image. The output patterns exhibit near-maximal entropy (∼7.5/8 bits ) and resilience under coarse spatial sampling due to the holographic broadcasting nature of the scattering, enabling privacy-preserving, bandwidth-efficient processing suitable for resource-constrained nodes. We further simulate a diffractive metasurface that optically implements the inverse transformation, demonstrating a route toward an active-passive hybrid pipeline with power-efficient inference. This architecture outlines a route to versatile, low-latency, and physics-native edge computing with potential in real-time photonic co-processors and autonomous systems.