Unveiling Search-Robust One-Shot Learning Operation in FeTFT-Based Ternary Content-Addressable Memory

With the rapid growth of data-centric artificial intelligence (AI), ternary content-addressable memory (TCAM) has emerged as a promising platform for parallel in-memory search. Here, we systematically investigate how the gate-dielectric material and structural design of metal-ferroelectric-metal-insulator-semiconductor (MFMIS) ferroelectric thin-film transistors (FeTFTs) affect TCAM search operation and long-term robustness. By comparing ZrO₂ (11 nm), ZrO₂ (8 nm), and HfO₂ (8 nm) gate dielectrics, we decouple the effects of material substitution, thickness scaling, and equivalent-oxide-thickness (EOT)-matched behavior. All devices show proper initial TCAM operation, with ZrO₂ 8 nm exhibiting the highest initial sensing performance owing to its larger read current. Under repeated search cycling, however, ZrO₂ 11 nm proves the most robust, while HfO₂ 8 nm suffers the most severe degradation-governed by competition between electron trapping and donor-like state creation rather than conventional ferroelectric fatigue. One-shot classification simulations using experimentally extracted bit error rate (BER) trends confirm that device-level degradation directly translates into accuracy loss, with ZrO₂ 11 nm retaining the highest accuracy. Increasing the memcapacitor area ratio enlarges the memory window but degrades sensing response due to added parasitic capacitance and RC delay. These findings establish gate-dielectric and interface design guidelines for search-robust FeTFT-based TCAM.