Wet Etching-Based WO₃ Patterning for High-Performance Neuromorphic Electrochemical Transistors

Abstract

WO₃-based electrochemical transistors (ECTs) are recognized as candidates for three-terminal memristors due to their high on–off ratio, long retention time, and rapid switching speed. However, their patterned fabrication often relies on complex vacuum systems or extreme processing conditions, hindering cost-effective scalability. Here, we developed a novel wet etching technique integrated with sol–gel-derived WO₃ channels, enabling ambient-air fabrication of Nafion-WO₃ ECTs. The wet-etched devices achieve an on–off ratio of ~10⁵, surpassing unetched and dry-etched counterparts by orders of magnitude. Furthermore, they exhibit exceptional paired-pulse facilitation and long-term stability, maintaining 12 distinct conductance states for 10³ s, and an on–off ratio of ~10² over 25 read–write cycles. XPS result shows higher W⁵⁺ content and M-O-H bond proportion for wet-etched devices, revealing an optimized interface, with enhanced H⁺ injection efficiency. The simulated artificial neural network using this wet-etched ECT shows ~97% recognition accuracy for handwritten numerals. This approach offers a novel patterning strategy for developing cost-effective, high-performance neuromorphic devices.

1. Introduction

Neuromorphic computing, leveraging physical neural networks, has emerged as a next-generation paradigm to enable machine learning applications such as natural language processing and computer vision [1,2]. While traditional von Neumann architectures can simulate neural networks, their inherent separation of memory and processing units restricts data transfer speeds and hampers efficient large-scale data processing, particularly for complex neuromorphic algorithms [3]. In contrast, biological brains integrate computation and storage, achieving high efficiency and low power consumption—a feature inspiring efforts to overcome von Neumann bottlenecks. To mimic such biological synergy, researchers have developed diverse synaptic devices, including resistive random-access memory (ReRAM) [4,5], phase-change memory (PCM) [6,7], ferroelectric random-access memory (FeRAM) [8,9], and electrochemical random-access memory (ECRAM) [10,11].

Electrochemical transistors (ECTs), a key class of electrochemical random-access memory (ECRAM), function as three-terminal memristors governed by electrochemical ion doping. By modulating the gate voltage or current, ions in the electrolyte layer are controlled, enabling precise tuning of channel conductance. ECTs offer distinct advantages, including decoupled read/write operations and multi-state programmability [12]. Common doping ions—Li⁺ [13,14], Na⁺ [15], and H⁺ [16,17]—vary in performance. Protons (H⁺), with minimal ionic radius and mass, enable rapid transport and reduced structural degradation during doping. Nafion, a proton-conducting polymer electrolyte (>10² S cm⁻¹ ionic conductivity), can enhance ECT response speed and durability due to its stability under harsh conditions [10,18]. For channel materials, transition metal oxides (TMOs) such as MoO₃ [19], Nb₂O₅ [20], TiO₂ [21], and WO₃ [22] are widely explored. WO₃, a wide-bandgap semiconductor (2.8–3.2 eV) [23], exhibits a layered structure and active d-orbitals that facilitate proton insertion at [WO₆] octahedral sites under bias. This process injects electrons into the conduction band, significantly boosting channel conductivity and enabling high conduction on–off ratios [24,25].

Despite its widespread use in high-performance memory devices [26], WO₃ lacks a simple, low-cost method for patterned fabrication, which limits the fabrication of large-area arrays for the application of artificial neural networks (ANN). Current approaches include shadow mask-assisted RF sputtering to confine deposition areas [18], reactive RF magnetron sputtering combined with photolithography and SF₆-based reactive ion etching (RIE) [27], high-temperature (>100 °C) thermal atomic layer etching (ALE) using sequential boron trichloride (BCl₃) and hydrogen fluoride (HF) treatments [28], and direct hydrothermal synthesis under high-temperature and atmospheric-pressure conditions with p-nitrobenzoic acid additives [29]. These methods require complex vacuum systems or extreme processing environments, posing challenges for safe, scalable, and economical ECT production. Therefore, exploring a new green and safe etching process to replace the traditional dry etching process with high hazard and power consumption is the purpose of this work.

In this study, we present a wet etching technique for WO₃ patterning that integrates seamlessly with conventional photolithography, enabling ambient-air fabrication of channel structures. By combining this method with sol–gel-derived WO₃ and a Nafion electrolyte, we demonstrate an ECT with exceptional storage performance, including high on–off ratios and prolonged retention times. By using experimental data of updatable conductance, we simulated an ANN based on the lookup table method with high accuracy in recognizing images. In Section 3.1, we compare the performance of wet etching, dry etching, and no etching on the WO₃ film. In Section 3.2, we present the performance of wet etching WO₃ devices on long-term and short-term memory. In Section 3.3, we present the implementation of the simulated ANN. This work may provide the solution for the patterning and integration of tungsten-based oxide on circuits in future work.

2. Materials and Methods

Tungsten hexachloride (WCl₆, Aladdin, Shanghai, China) was dissolved in 99% absolute alcohol (Aladdin, Shanghai, China), and 0.1 mol/L precursor solution of WCl₆ was prepared. The precursor solution of Nafion electrolyte was synthesized by mixing Nafion117 (Macklin, Shanghai, China) and 99% isopropanol (Aladdin, Shanghai, China) according to a volume ratio of 1:1. The WCl₆ precursor solution was spin-coated on p++Si/SiO₂ at 4000 rpm/30 s, and then annealed at 200 °C for 10 min and 400 °C for 1 h in air, thus obtaining the 30 nm WO₃ thin film. Photolithography was employed to define channel patterns, followed by wet or dry etching.

Wet etching: Samples were developed using an alkaline positive photoresist developer (RZX-3038, Suzhou, China), primarily composed of tetramethylammonium hydroxide (TMAH). After development, the substrate was immersed in TMAH for 2 h under ambient conditions and rinsed with deionized water. The electricity power consumption is estimated at about 0.6 kW h.

Dry etching: Reactive ion etching (RIE, E100, Tailong Electronics, Beijing, China) was performed using SF₆ and O₂ (100 sccm each) at 100 W PolePower for 35 s, with a HighValvePOS of 30. The electricity power consumption is estimated at about 7.5 kW h for the total operation time (~1.5 h) of the equipment. Note that SF₆ is an asphyxiant and greenhouse gas, which needs strict management, and its leakage may have an impact on the environment.

Following etching, residual photoresist was removed by sequential ultrasonic cleaning in acetone and ethanol (10 min each). A 60 nm aluminum (Al) source/drain electrode was deposited via thermal evaporation. The Nafion electrolyte precursor was spin-coated onto the sample at 2000 rpm for 30 s and annealed at 100 °C for 10 min. This process was repeated four times to achieve a 290 nm-thick electrolyte layer. Finally, a 100 nm Al gate electrode was thermally evaporated. Metal masks defined the patterns for source, drain, and gate electrodes, with channel dimensions of L = 200 μm and W = 1000 μm. X-ray photoelectron spectroscopy (XPS, Thermo K-Alpha, Guangzhou, China) was employed to analyze elemental composition changes during etching electrochemical doping. An atomic force microscope (AFM, Dimension Icon, Guangzhou, China) was used to analyze the WO₃ surface morphology of ECTs without etching and with dry etching and wet etching. The thickness of Nafion and WO₃ were both measured by AFM.

3. Results

3.1. Comparison of Different Etching Methods

Figure 1a illustrates the fabrication process of the Nafion–WO₃ ECT, integrating sol–gel synthesis, photolithography, etching, and thermal evaporation. The device adopts a top-gate top-contact structure, with WO₃ as the channel layer and Nafion as the proton-conducting electrolyte. Figure 1b,c presents optical micrographs of patterned WO₃ channels etched via SF₆-based dry etching and TMAH wet etching, respectively. Both methods produced well-defined edges with negligible residual WO₃ outside the target regions, confirming the precision and compatibility of the etching processes. Figure 1d shows the AFM images of the WO₃ surface of the ECTs without etching and with dry etching and wet etching. Note that, compared with the pristine WO₃ surface without etching, the surface of the etched WO₃ is different from that of the pristine state because of the additional process. The root mean squares (RMS) of the roughness are 10.6 nm, 16.4 nm, and 8.45 nm for the non-etching, dry-etching, and wet-etching films, respectively. It is found that the film with the dry etching process is rougher than that with wet etching. This is likely because the ion bombardment caused damage to the film underneath the photoresist. The thickness of the WO₃ film is about 30 nm without significant change after dry etching and wet etching, confirmed by the AFM.

Figure 2 presents the transfer and output characteristics of ECTs fabricated without etching and with different etching methods at V_D = 3 V. All transfer curves exhibit counterclockwise hysteresis, indicating non-volatile conductance enhancement during the first voltage sweep, which persists in subsequent scans. The switching dynamic range ∆I_D(on/off), defined as the ratio of I_D(on) to I_D(off) at V_G = −2 V, reached 2.775 × 10⁵ for wet-etched devices—over 1000× and 2× higher than unetched (23.44) and dry-etched (1.153 × 10⁵) devices, respectively. Similarly, output curves demonstrated non-volatile I_D modulation with increasing V_G. At low V_G (<1 V), all curves almost overlapped, reflecting a stable low-conductance state. For V_G ≥ 3 V, I_D surged irreversibly, confirming a switch to the high-conductance state. The saturation dynamic range ∆I_D(sat), calculated as I_D(sat) at V_G = 5 V versus −1 V (V_D = 1 V), further highlighted the superiority of wet etching (6.103 × 10⁴), outperforming dry etching (6.097 × 10³) and unetched devices (75.14) by orders of magnitude. These results demonstrate that patterned WO₃ ECTs achieve significantly broader operational dynamic ranges.

The enhanced on–off current ratio in patterned WO₃ ECTs, compared to unpatterned devices, arises from suppressed leakage currents. Unpatterned channels permit multiple conductive pathways between source and drain electrodes, whereas patterned geometries isolate the active channel, drastically reducing off-state currents by orders of magnitude. The non-volatile conductance modulation stems from proton (H⁺) dynamics. Under positive V_G, sulfonic acid groups (-SO₃H) in Nafion release protons that migrate to the Nafion/WO₃ interface and electrochemically dope into the channel. Under negative V_G, protons in WO₃ are extracted. The above process can be described via the reaction [18]:

$$x{\mathrm{H}}^{+}+x{\mathrm{e}}^{-}+{\mathrm{W}\mathrm{O}}_3\rightleftharpoons {\mathrm{H}}_x{\mathrm{W}\mathrm{O}}_3$$

(1)

This doping injects electrons into WO₃’s conduction band, elevating channel conductance. The protons remain trapped in WO₃ under zero bias, stabilizing the high-conductance state. Conversely, negative V_G extracts protons, reverting the channel to a low-conductance state. To describe the adjustable conductance by the injected hydrogen, the conductance can be described as $G=G_0+e{\mu }_{\mathrm{n}}{∆c}_{\mathrm{e}}Wd/L$, where $e$, ${\mu }_{\mathrm{n}}$, $L$, $W$, $d$, $c_{\mathrm{e}}$, and $G_0$ are the electron charge, the electron mobility, channel length, the channel width, channel thickness, electron concentration induced by the injected hydrogen in Reaction (1), and the initial conductance, respectively.

As demonstrated in Figure 2, gate voltage can effectively modulate the I_D through proton injection dynamics. To emulate synaptic long-term potentiation/depression (LTP/LTD), we applied sequential positive/negative current pulses to the gate, tuning the channel conductance (G = I_D/V_DS, V_DS = 1 V). Figure 3a–c shows the LTP/LTD characteristics, where G_min and G_max denote the initial and post-positive pulse conductance, respectively. The on–off ratio (G_max/G_min) of unetched devices reached only 22 under 30 pulses (2 μA, 200 ms), while dry- and wet-etched devices achieved 2827 and 9429, respectively, with fewer and narrower pulses (15 pulses, 2 μA, 50 ms) (Figure 3d). Notably, wet-etched ECTs attained an on–off ratio near 10⁵ under identical conditions (2 μA, 50 ms), surpassing dry-etched counterparts by ~3.3× (9429 vs. 2827). This stark contrast underscores two key findings: (1) patterning WO₃ channels drastically enhances conductance modulation efficiency, even with shorter pulses; (2) the TMAH wet etching outperforms conventional SF₆ dry etching, enabling higher on–off ratios (>10⁵) and expanded settable conductance states for robust neuromorphic computing.

To evaluate the symmetry of conductance states during LTP/LTD switching, we calculated the asymmetry ratio (AR) using [30,31,32]:

$$AR={\displaystyle \frac{\max \left|G_{\mathrm{P}}\left(n\right)-G_{\mathrm{D}}\left(n\right)\right|}{G\mathrm{m}\mathrm{a}\mathrm{x}-G\mathrm{m}\mathrm{i}\mathrm{n}}}$$

(2)

where G_P(n) and G_D(n) represent the conductance at the nth pulse during potentiation and depression, respectively. AR approaching zero indicates ideal symmetry. As shown in Figure 3e, patterned WO₃ ECTs exhibited lower average AR and standard deviation than unpatterned devices, confirming improved state distribution symmetry and consistency [33]. Notably, wet-etched devices achieved marginally higher AR (0.69) compared to their dry-etched counterparts (0.64). The reason for this phenomenon may be that RIE leads to the amorphization of WO₃, which makes the ion migration path disordered. Compared with the crystal structure retained by wet etching, the energy barrier distribution of ion migration in amorphous materials is more uniform, which leads to the narrowing of the conductance change difference of LTP/LTD.

To investigate the influence of the etching process on H⁺ injection, XPS analysis was performed on WO₃ channels before and after electrochemical H⁺ injection (Figure 4). Pre-injection spectra of all devices exhibited characteristics with only W⁶⁺ 4f_7/2 (at 35.22 eV) and 4f_5/2 (at 37.37 eV) peaks, confirming the pristine WO₃ phase [34]. Deconvolution of O1s spectra revealed three components: metal–oxygen bonds (M-O, at 529.94 eV), oxygen vacancies (O_v, at 530.88 eV), and hydroxyl groups (M-O-H, at 532.58 eV) [35]. The minor M-O-H contribution likely originated from surface-adsorbed oxygen species [36]. After H⁺ injection (10 μA, 10 min), new W⁵⁺ 4f_7/2 (at 34.16 eV) and 4f_5/2 (at 36.31 eV) peaks emerged, accompanied by an increase in M-O-H content, confirming proton intercalation. The new peak at 534.84 eV, caused by the residue of Nafion on the surface of channel, will not be included in the percentage changes of W and O elements in the three ECTs. The difference in XPS images of the three ECTs can be explained by the quantitative analysis of the peak area ratio change in Figure 4a–l. As displayed in Figure 4m,n, the area ratios between the W element peak and O element peak under each etching method are converted into percentages and their relative change is shown.

Quantitative analysis of XPS peak areas (Figure 4m,n) revealed distinct trends in W and O composition during H⁺ injection. Post-injection, the W⁵⁺ content increased by 37.07%, 44.53%, and 46.29% for unetched, dry-etched, and wet-etched devices, respectively. The W:O ratios are 1:2.81, 1:2.78, and 1:2.77 for the channel without etching, with dry etching, and with wet etching, respectively. Concurrently, hydroxyl oxygen (M-O-H) content rose by 17.81%, 25.10%, and 25.74% for unetched, dry-etched, and wet-etched devices, respectively, while the W⁶⁺, M-O, and O_v contributions decreased proportionally. These shifts confirm the reduction of W⁶⁺ to W⁵⁺ during H⁺ intercalation, consistent with Reaction (1). The stoichiometric parameter x in H_xWO₃—calculated from W⁵⁺ fractions—reached 0.37 (unetched), 0.45 (dry-etched), and 0.46 (wet-etched), demonstrating enhanced H⁺ injection in patterned channels. Notably, wet-etched devices exhibited the highest ΔW⁵⁺ (+46.29%) and ΔM-O-H (+25.74%), outperforming their dry-etched counterparts. This superiority likely stems from TMAH wet etching’s milder reaction kinetics compared to SF₆ RIE, which minimizes structural defects and preserves the WO₃ lattice integrity, thereby facilitating deeper H⁺ injection. The improved injection efficiency explains the superior on–off ratios (>10⁵) of wet-etched ECTs.

3.2. Long-Term and Short-Term Memory in WO₃ ECTs

Figure 5a,b shows the LTP/LTD characteristics of wet-etched WO₃ ECTs under varying gate pulse widths and numbers, achieving stable on–off ratios over 10³ with an average AR of 0.63 (

Table 1. The on–off ratio and asymmetry ratio of ECT with wet etching under different gate pulse widths or numbers.

Pulse Width/Number	Gmax/Gmin	AR	$\overline{\mathit{A}\mathit{R}}$
50 ms	259.0	0.67
100 ms	1731.1	0.68
200 ms	3192.4	0.89
			0.63
15p + 15d	259.0	0.67
30p + 30d	859.2	0.58
50p + 50d	3652.0	0.31

).

To emulate synaptic short-term plasticity, paired-pulse facilitation (PPF) was evaluated. A pair of pulses from presynaptic neurons will stimulate postsynaptic neurons to generate two excitatory postsynaptic currents (EPSCs), and the second EPSC is larger than the first one [37,38]. PPF arises from transient H⁺ adsorption/desorption at the WO₃ surface, distinct from the bulk doping responsible for LTP/LTD. Under a positive gate pulse, H⁺ from Nafion adsorbs onto WO₃, forming surface O-H bonds, providing additional carriers in the channel and temporarily increasing channel conductivity I_EPSC. Upon pulse removal, internal electric fields drive partial proton desorption, causing I_EPSC decay. When a second pulse is applied before complete desorption (∆t < relaxation time), residual adsorbed protons amplify subsequent I_EPSC (Figure 5c,d) [16].

The PPF index (I₂/I₁) increased from 1.01 to 3.38 as ∆t decreased from 800 ms to 50 ms, confirming enhanced short-term memory at shorter intervals [39]. The PPF curve is fitted by a biexponential function [40,41]:

$$\mathrm{P}\mathrm{P}\mathrm{F} \mathrm{i}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x}=1+A_1\mathrm{e}\mathrm{x}\mathrm{p}\left(-\mathsf{\Delta }t/{\tau }_1\right)+A_2\mathrm{e}\mathrm{x}\mathrm{p}\left(-\mathsf{\Delta }t/{\tau }_2\right)$$

(3)

where τ₁ and τ₂ correspond to the characteristic relaxation time of H⁺ after the first and second pulse stimulation [42]. The fitting yielded relaxation times of τ₁ = 43.3 ms and τ₂ = 218.2 ms. The longer τ₂ reflects sustained H⁺ retention after sequential pulses, enabling short-term cumulative conductance enhancement.

Figure 5e,f demonstrates the stability and endurance of wet-etched ECTs. Twelve distinct conductance states (2 nS–350 μS) were programmed, and all states remained non-overlapping after 10³ s of operation in ambient air, confirming robust state retention. Conductance variation within each state was quantified using the coefficient of variation (c_v) [43]:

$$c_{\mathrm{v}}=\sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{m}}{\left(G_{\mathrm{i}}-\overline{G}\right)}^2/\left(\mathrm{m}-1\right)}/\overline{G}$$

(4)

where $G_{\mathrm{i}}$ is the conductance value sampled at different times, $\overline{G}$ is the mean conductance of $G_{\mathrm{i}}$, and m = 201 (5 s sampling interval over 10³ s). All states exhibited ultra-low c_v < 0.05, indicating minimal conductance drift. Furthermore, the device retained an on–off ratio ~10² after 25 read–write cycles, demonstrating its cycling reliability. The on–off ratio and the number of read–write cycles of ECT are two important parameters in neural morphology calculation. A high on–off ratio can accommodate more storage states, so that there is a larger fault-tolerant interval between the storage states. A sufficient number of cycles can ensure that the conductance drift of the device is negligible when updating conductance, and the conductance can be written accurately.

3.3. Simulated ANN for Image Recognition

To prove the potential of wet-etched ECTs for application in ANNs, a multilayer perceptron (MLP) neural network was constructed by using the Pytorch 1.7.1 platform. The image recognition for the Modified National Institute of Standards and Technology (MNIST) data set was simulated. The normalized LTP/LTD conductance data (Figure 6a, blue and magenta dots) of ECT were used in the weight updating process of the ANN, as shown in Figure 6a. As a comparison, an ideal datapoint (Figure 6a, gray dots) is generated from the computer as the case with best linearity.

Figure 6b represents the MLP neural network model consisting of 784 input neurons, 64 × 3 hidden neurons, and 10 output neurons, respectively. The training and recognition process of the ANN was carried out in a training data set containing 60,000 images and a separate test data set containing 10,000 images, respectively. As for the weight updating method, based on Adam optimizer, we rewrite a gradient optimizer of discrete pulse coding, which realizes the pulse quantization of weight updating through a bio-inspired LTP and LTD look-up table (LUT). On the basis of retaining the adaptive learning rate characteristics of Adam optimizer, this optimizer replaces the continuous gradient update value with discrete pulse amplitude. Specifically, consistent with the traditional Adam optimizer, it first calculates the first moment (momentum) m and the second moment (variance) v of the gradient, and then calculates the original weight update amount Δw; on this basis, the continuous update amount Δw is realized by double look-up table (LTP/LTD-LUT). Then, we multiply it by the learning rate to obtain the final weight update.

As seen in Figure 6c, wet-etched ECT displayed ~97.04% image recognition accuracy after 20 epochs, while the accuracy of the ideal device was ~97.33%. The corresponding average loss decreased from 0.650 to 0.061 after 20 epochs for the experimental device and decreased from 0.659 to 0.058 for the ideal device. The accuracy for the experimental data is close to that for the ideal case. The reason is likely the training gradient in the early stages of the MLP. Consequently, the weights trained in the original code primarily utilize the first half of the LTP/LTD data. This subset is highly linear and sufficiently dense to cover the normalized conductance range of 0–0.4, thereby improving convergence speed and accuracy. These results demonstrate that wet-etched ECTs can realize high-accuracy image recognition of handwritten numerals, as well as perform multilevel data storage and neuromorphic computing [44,45].

The above algorithm and simulation can be extended in platforms, such as MATLAB R2022b and SPICE (PySpice1.5), that can be combined with the Python 3.8 interface. For hardware implementation, this algorithm can be conducted by a hybrid computing architecture with a field programmable gate array (FPGA) and ECT array. The FPGA, as a digital controller, executes the gradient calculation, LUT query, and other logical operations. As an analog computing unit, the ECT array encodes input pulse, stores the weight values, and realizes multiplication and addition operations [46,47,48].

As a comparison, we summarized the performance of the existing advanced ECTs and the results of this work in

Table 2. The comparison between the existing advanced ECTs and the results of this work.

Channel	Electrolyte	Ion Type	On-Off Ratio	Conductance Retention Time	Recognition Accuracy	Ref.
MoO3	H2SO4	H+	~2	50 s	-	[16]
Nb2O5	LixSiO2	Li+	~20	103 s	-	[20]
VO2	CsClO4	H+	~2	4 × 103 s	98%	[49]
In2O3	Al2O3	H+	~3	120 s	85%	[50]
WO3	ZrO2	H+	~20	103 s	93%	[27]
IZO	Nafion	H+	~2.5	103 s	84%	[45]
WO3	Nafion	H+	~107	-	-	[18]
WO3	Nafion	H+	~105	103 s	97%	Our work

. The ECT in this work shows balance performance on the on–off ratio, retention time, and recognition accuracy in the simulated neural network.

4. Conclusions

We developed a novel wet etching process for WO₃ patterning, enabling the fabrication of Nafion–WO₃ ECTs with sol–gel-derived WO₃ channels. The wet-etched ECTs achieved an on–off ratio of ~10⁵, outperforming unetched and dry-etched devices by orders of magnitude. These devices also exhibited excellent paired-pulse facilitation (PPF) and long-term stability, maintaining 12 distinct conductance states for 10³ s and retaining an on–off ratio of ~10² over 25 read–write cycles. XPS analysis confirmed that wet etching enhanced H⁺ injection efficiency, with higher W⁵⁺ and M-O-H content compared to dry etching, leading to more thorough electrochemical reactions and superior device performance. The ANN simulation using this wet-etched ECT showed ~97% image recognition accuracy for handwritten numerals. The average loss of wet-etched ECT decreased from 0.65 to 0.06 after 20 epochs. This work demonstrates that wet etching is a promising approach for scalable, high-performance neuromorphic devices.