Volatile ZrO₂ Antiferroelectric Tunnel Junctions for Rapid, Energy-Efficient Physical Reservoir Computing

Physical reservoir computing requires nonlinear response, fading memory, and rich transient state diversity, yet conventional nonvolatile memories often rely on explicit reset operations or long relaxation times. ZrO₂-based two-terminal antiferroelectric tunnel junctions (AFTJs) exploit the field-induced tetragonal-to-orthorhombic transition and spontaneous back-switching of antiferroelectric ZrO₂. This intrinsic self-relaxation provides reset-free fading memory in the sub-ms regime. An amorphous In-Ga-Zn oxide interlayer enlarges the dynamic range, and stoichiometric control identifies the 2:2:1-ZrO₂ AFTJ as the optimal composition, delivering an I_on/I_off of ∼890, a peak nonlinearity factor of ∼48.4, and a paired-pulse facilitation index of 1.79. The enhanced memory margin and nonlinear dynamics support 16 transient current states and yield 90.4% accuracy in Modified National Institute of Standards and Technology classification with a fourfold reduction in spatiotemporal dimensionality. Temporal information processing is further assessed using an experimentally calibrated circuit-level reservoir model, enabling waveform classification, one-step-ahead Hénon-map prediction (normalized root-mean-square error [NRMSE] = 0.01489), and forecasting of a noisy real-world semiconductor index time series (NRMSE = 0.12263). The fabricated 40 000 µm² AFTJ shows a unit latency of ∼2 µs and energy consumption below 480 pJ per operation. Analytical area scaling projects show that a 100 µm² device could achieve ∼192 ns latency and ∼115 fJ per operation.