Bipolar physical reservoir computing for intrinsic processing of signed temporal signals

Physical reservoir computing (RC) offers a highly energy-efficient paradigm for processing spatiotemporal signals. However, most existing physical RC implementations rely on electronic devices with inherently unipolar responses, fundamentally constraining their ability to encode and process bipolar signals that are ubiquitous in real-world data, such as financial volatility, directional mechanical motion, and event streams from dynamic vision sensors (DVS). Here, we report a bipolar physical RC system enabled by ZnO-interlayered metal-oxide thin-film transistors that generate intrinsic bipolar responses within a single device. By explicitly incorporating full sign information, the information encoding efficiency is enhanced by a factor of 1.58, accompanied by a substantial reduction in signal ambiguity. Leveraging this capability, the bipolar RC boosts experimental accuracy from 89.9% to 96.9% in DVS-based event recognition compared with unipolar counterparts. Moreover, it exhibits superior performance and energy efficiency in challenging human action classification benchmarks. These results underscore the critical importance of polarity integration and establish a viable pathway toward fully exploiting the computational potential of physical reservoir computing.