An organic electrochemical transistor for multi-modal sensing, memory and processing

Conventional AI hardware separates sensing, processing and memory, incurring large energy and latency overheads due to frequent data movement and analogue–digital conversions. Biological nervous systems are far more energy–area efficient because sensing, processing and memory are co-located, enabling advanced behaviors such as conditioned reflexes. Engineering artificial systems that fuse these functions is difficult due to the heterogeneity of sensors, synapses and neurons, which complicates fabrication, integration density and conductance matching at small scales. Prior in-sensor computing devices based on 2D materials offer sensing and processing but lack non-volatile memory; conversely, phase-change memories and redox memristors enable in-memory computation but lack sensing capability. Organic electrochemical transistors (OECTs) operate at low power in wet environments and have shown either sensing or analogue memory, but multi-modal sensing akin to polymodal biological receptors and reconfigurable volatile/non-volatile behavior in a single homogeneous device remain challenging. Achieving both volatile (for sensing) and non-volatile (for memory) operation is fundamentally difficult because they require contradictory ion kinetics. This work addresses these challenges by designing a single OECT that reconfigures between a volatile receptor and a non-volatile synapse, enabling homogeneous hardware that fuses sensing, memory and processing.

Heterogeneous module integrations have enabled specific artificial olfaction, tactile and gesture systems, but retain separated sensing and processing modules, limiting scalability and efficiency. Devices for in-sensor computing using 2D materials have demonstrated promising multi-modal perception but lack true non-volatile memory for processing. Phase-change materials and RRAM-based memristors excel at analogue in-memory computing but cannot sense external modalities. Prior OECT efforts have achieved high-performance chemical and electrophysiological sensing and, separately, non-volatile behavior by enforcing open-circuit ion kinetics using heterogeneous device stacks, sometimes reaching hundreds of analogue states with limited retention. However, integrated multi-modal sensing together with robust non-volatile memory and processing within a single, homogeneous OECT platform has remained elusive. Additionally, increasing crystallinity to suppress ion motion improves non-volatility but can reduce volumetric capacitance and ion mobility, degrading sensing sensitivity, highlighting a trade-off that must be addressed by architectural and microstructural design.

Device design and materials: A vertical traverse OECT (v-OECT) architecture was developed using an ion gel (1-ethyl-3-methylimidazolium bis(trifluoromethylsulfonyl)imide [EMIM][TFSI]: PVDF-HFP) or aqueous electrolyte gate. The organic mixed ionic–electronic conductor (OMIEC) channel is PTBT-p, processed to form a crystalline–amorphous composite microstructure. Channel geometry: channel length (film thickness) 40–80 nm; channel thickness (depth) ~100 µm, yielding a high depth/length ratio (d/L) ≈ 2,000. This geometry provides high gain in volatile mode and small potential gradients across the channel depth to help retain doped ions for non-volatility. Crystallization control: Films were annealed at 100–200 °C to tune crystallinity and crystallite size (~30 nm by cryo-EM and GIWAXS). Operando UV–vis and synchrotron GIWAXS were used to distinguish doping of amorphous (low-gate potential, LGP) and crystalline (high-gate potential, HGP) regions, correlating to different volumetric capacitances. Gate electrode process: Polarizable Au gates enable reversible redox of [EMIM]+, preventing counterion compensation and preserving trapped anions in the channel for non-volatile behavior; non-polarizable Ag/AgCl gates lead to volatile response due to counterion compensation. Mechanism probes: Operando UV–vis spectroscopy tracked 0–1 absorption peak suppression and polaron formation under bias; GIWAXS monitored lamellar stacking expansion (from 1.39 to 1.53 nm) indicating anion embedding in crystalline glycol side chains under HGP; XPS depth profiling (F 1s from [TFSI]-) quantified anion distribution pre/post bias and with channel thickness. Electrical characterization: Steady-state and transient volatile characteristics were measured with Keithley SMUs and DMM; non-volatile write pulses supplied by a function generator with read currents measured to compute GDS; sub-µs pulsing assessed via oscilloscope across a load resistor. Physiological sensing: ECG recorded on a human volunteer with gel-assisted Ag/AgCl electrodes interfaced to the cv-OECT gate and grounded source; plant ion variations (Mimosa pudica, Venus flytrap) sensed via ion-gel-coated Ag/AgCl contacts to leaves/stems. STDP implementation: Homogeneous 1T1R synapse formed from two identical cv-OECTs (T volatile, R non-volatile). Pre-/post-spike timing controlled by SMUs; conductance changes read with an additional SMU. Microstructural characterization: Cryo-EM prepared via plunge freezing and imaged at 200 kV; GIWAXS at ALS beamline 7.3.3 with controlled cross-sectional ionic access; operando UV–vis/NIR with fiber-coupled spectrometers; XPS with monochromatic Al Kα source; cyclic voltammetry and EIS with Au or Ag/AgCl working electrodes in EMIMTFSI. Simulations: ANN and SNN training on MNIST with experimentally extracted device non-idealities (nonlinearity, noise, variations). STDP learning rules fitted to exponential functions from measurements. Reservoir computing (RC) for ECG diagnosis simulated using experimentally calibrated volatile conductance dynamics dG/dt = k i(t) + n G(t) (k=10, n=2), mapping ECG segments to reservoir outputs, followed by an ANN readout.

Dual-mode operation: The cv-OECT (crystalline v-OECT annealed at 200 °C) reconfigures between a volatile receptor (LGP) and a non-volatile synapse (HGP) within the same device. Mechanism: Two-stage doping observed by operando UV–vis (amorphous then crystalline regions), with an anion trapping energy barrier ~0.8 eV (Au gate). GIWAXS shows lamellar spacing expands 1.39→1.53 nm under HGP, consistent with anion embedding among ordered glycol side chains; thick channels preserve trapped ions to ensure non-volatility. Volatile receptor performance: Normalized peak transconductance g_m/V_DS = 27 mS V−1; on/off ratio up to 5×10^5 (and as high as 8×10^6 via geometry); subthreshold swing ~65 mV dec−1 near the thermal limit (59.6 mV dec−1), outperforming p-OECTs (121 mV dec−1) and reported v-EGOFETs (90.5 mV dec−1). Fast switching with τ_on ≈ 6.67 ms (10 µm^2) and 0.82 ms (400 µm^2); stable 10^6 on/off and speed over 30 min cycling in air. Operation in aqueous 0.1 M NaCl without degradation. Multi-modal sensing: Plant ion dynamics recorded upon light and touch (Mimosa pudica, Venus flytrap) without external amplifiers; ECG captured using flexible cv-OECT at <1 µW power (vs. typical PEDOT:PSS OECTs >500 µW). Additional modalities: gustation and temperature sensing with normalized responses of 19.0% per decade (ionic concentration) and 3.2% per °C; optoelectronic features suitable for artificial vision. Non-volatile synapse performance: Non-volatility achieved for |V_G| > 0.8 V (ion-gel-gated), with as-cast devices remaining volatile. Hysteretic transfer curves yield a memory window of 2.1 V. State retention >10,000 s across six analogue levels in air with low drift (γ ≈ 0.003–0.006). High-precision analogue updates: 1,024 (10-bit) conductance states over a wide dynamic range during LTP with low switching stochasticity. Under voltage control, nonlinearity μ_p/μ_u = 0.20/1.63 and (ΔG_PS/σ)^2 = 179 for 50-state programming; under current control, (ΔG_PS/σ)^2 = 290 and −0.49% cycle-to-cycle variation over 2,000 events in 50 cycles. Sub-µs programming feasible: 20×20 µm^2 device supports 200 ns write pulses with 800 ns write–read delay; 100×100 nm^2 scaling projects near-ns switching. STDP: Homogeneous 1T1R synapse exhibits STDP with a time constant ~60 ms and accurate analogue tuning without heterogeneous integration. Array and tasks: A monolithic 9×2 cv-OECT array (9 Ts, 9 Rs) implements supervised STDP; simulated MNIST classification achieves ~89% (SNN) and ~91% (ANN) with cv-OECTs, surpassing 6-bit heterogeneous RRAM SNN (~83%) and ANN (~87%). Fused sensing–processing: Conditioned reflex demonstrated with learning rate modulated by ionic environment (via Nernst potential). Reservoir computing for ECG: A homogeneous 12×1 cv-OECT receptor array plus a 156×5 cv-OECT-based ANN readout achieves simulated 100% accuracy after ~700 epochs on five ECG categories (PTB-XL), using cv-OECT volatile dynamics as the reservoir.

The presented cv-OECT resolves the long-standing tension between volatile sensing and non-volatile memory in OECTs by combining a deep vertical channel and controlled crystalline–amorphous microstructure with a polarizable gate process. This enables selective ionic doping: amorphous-region doping under low bias for fast, reversible receptor behavior, and crystalline-region trapping under higher bias for robust non-volatile synaptic states. The high d/L ratio flattens internal fields, reducing back-drift of trapped ions and supporting long retention without heterogeneous elements. As volatile receptors, devices provide near-thermal-limit subthreshold swings, record-high on/off ratios for OMIEC transistors, millisecond dynamics, and multi-modal sensing at ultra-low power, enabling direct acquisition of biological signals (plants, ECG) without auxiliary amplifiers. As non-volatile synapses, the devices deliver 10-bit analogue programmability, low write stochasticity, wide dynamic range and long retention, supporting accurate in-memory computing and bio-plausible learning (STDP) in a homogeneous 1T1R scheme. System-level demonstrations—including conditioned reflex and RC-based ECG diagnosis—show that homogeneous integration of identical devices can fuse sensing and processing, reduce network size via RC pre-processing, and improve classification performance over heterogeneous RRAM baselines when device non-idealities are accounted for. These results highlight a practical path toward compact, energy-efficient edge AI hardware that mimics biological organization by co-locating perception, memory and computation.

An organic electrochemical transistor with a vertical traverse architecture and engineered crystalline–amorphous channel enables reconfigurable operation as both a volatile multi-modal receptor and a non-volatile analogue synapse. By controlling device geometry, channel microstructure and gate electrode process, the device achieves near-thermal-limit switching, record-high on/off ratios, fast millisecond dynamics, 10-bit conductance states with low noise, and >10,000 s retention. Homogeneous arrays implement accurate STDP learning, conditioned reflex behavior and, when combined with reservoir computing, real-time ECG classification in simulations using experimentally calibrated models. This approach advances the development of compact, low-power edge AI systems that fuse sensing, memory and processing within a single device technology. Future work should address large-scale array uniformity and yield, further device miniaturization for sub-ns operation, and full hardware demonstrations of the integrated sensing–computing pipelines.

A primary limitation noted is poor device uniformity and yield in large OMIEC thin-film arrays due to defects (pinholes, thickness variations), hindering practical large-scale OECT-based ANN/SNN implementations. The demonstrated network performances for MNIST and ECG tasks are simulation-based using experimentally calibrated device models rather than full hardware demonstrations. Non-volatile operation relies on polarizable Au gates to prevent counterion compensation; non-polarizable Ag/AgCl gates yield only volatile behavior, constraining materials/stack choices. The hardware demonstrations use relatively small arrays (e.g., 9×2), and ECG classification employed a limited dataset (50 samples across five classes) within the RC framework.