Designable van der Waals Crystal for Artificial Neuronal Cell Mimicking Authors: Jinhyoung Lee, Gunhyoung Kim, Dongho Lee, Seowoo Son, Hyunho Seok, Sihoon Son, Hyunbin Choi, Geonwook Kim, Geumji Back, Hyunkyu Kim, Chaerin Park, Junmin Ahn, Seongyun Je, Chaeyoung Im, Junil Cho, Magdalena Grzeszczyk, Seongho Kim, Eunseo Go, Hyunwoo Shim, Donghwan Choi, Muyoung Kim, Hyoeng‐U. Kim, Won‐Jun Jang, Taesung Kim Published: 2026 ABSTRACT ABSTRACTOptogenetics employs light to regulate neuronal activity with exceptional spatiotemporal precision, thereby enabling the direct modulation of learning and memory processes in the human brain. This capability to externally control neuronal signaling via optical stimuli has provided profound insights into brain function and established a versatile strategy for engineering bio‐inspired information processing. Herein, a designable van der Waals (vdW) crystal has been demonstrated for device‐scale neuronal cell mimicking. The structural similarity between ion‐channel in biological membranes and layered vdW lattices is realized with nano‐crystallization. As the sequential transition from carrier transport dominance to ion transport dominance reveals the dynamic control over synaptic weight updates. Optoelectric synaptic plasticity in designable vdW crystal (long‐term potentiation and depression, paired‐pulse facilitation, and a tunable short‐term to long‐term memory transition) were conclusively observed and correlates with photo‐induced carrier trapping and ionic migration. Furthermore, learning‐forgetting‐relearning cycles achieve 34.7% increased retention efficiency compared to bulk ReSe2. Functional demonstrations in edge detection and CIFAR‐10 image recognition confirm the synaptic plasticity into system‐level neuromorphic computation, with a recognition accuracy of 96.24%. In conclusion, we envision that designable vdW artificial crystal will provide the versatile advances for 3D stackable neuromorphic vision architectures. Introduction In the era of artificial intelligence (AI), the ever‐increasing demand for real‐time data processing comprehensively requires the development of neuromorphic vision systems capable of image sensing and optogenetic‐inspired information processing within a single unit. Neuromorphic vision systems are a set of neuronal devices designed to mimic the visual perception and processing of human visual systems [1], which hold significant promise in fields, including the machine vision, autonomous vehicles, and the Internet of Things. As the human visual system including the retina and visual cortex [2], overwhelms the conventional vision system with high integration density and low‐energy consumption. The human retina has been configured as a hierarchical biostructure, enabling the visual sensing and pre‐processing, including the photo‐reception, feature extraction, temporal contrast enhancement, and denoising [3]. Owing to this structural configuration of retina, high‐level visual information processing in the visual cortex can effectively enhance the data transmission efficiency between sensor and processing units [4], while simplifying the hardware design complexity with a higher vertical 3D integration level [5]. Under the light illumination at neurons and synapses, light‐sensitive ion channels can be activated or deactivated in receptors, proteins, or ligands. Such light‐induced bionic action enables the modulation of memory functions in the human brain [6]. Moreover, light illumination enables the manipulation of the biological behaviors of neurons and synapses with a higher spatiotemporal resolution than conventional electrical bias [7]. These optical controllability of neuronal activity in real brain can be homogeneously applied to the artificial optogenetic‐inspired intelligent systems [8], including neuromorphic and in‐memory computing system. To realize the device‐scale visual cortex [9], neuromorphic visual systems with in‐memory computing architecture have been proposed by integrating the image sensors with optoelectronic resistive random‐access memory [10] that replicates the human retina for advanced sensing and pre‐processing functions. The core operation components of optogenetics‐inspired intelligent systems are an optoelectronic synapse, operating with internal conductance variation with incident light signals. The internal synaptic weight and its corresponding conductance state are sequentially updated with internal photo‐carrier dynamics [11]. Optical synaptic behavior generally derived from the persistent photoconductivity effect [12], or a photo‐induced conductance variation, which persists for a long period after the optical excitation termination. Such persistent photoconductivity effect can be attained with spatial confinement of photo‐induced carrier dynamics [13], deriving from the heterogeneous bandgap configuration. Such heterogeneous bandgap configuration generates the band bending and its corresponding potential barriers at interfacial contact between heterogenous bandgap, thereby accumulating the interfacial photo‐carriers [14]. An optically‐generated photo‐carriers increment [15] induces the internal potential variation, thereby enabling the channel conductance controllability (synaptic weight). Recently, van der Waals (vdW) materials and its heterostructure have been reported as promising candidates to realize the optoelectrical synapse [16] owing to its optical performance [17], mechanical flexibility [18], and atomic‐scale thickness [19], overcoming the conventional limitations of optical materials and applications [20]. Nonetheless, direct synthesis of designable vdW crystal for artificial neuronal cell mimicking has not commercially been achieved, attributed to the three major limitations: (i) the difficulty of achieving precise control over vdW grain control and intercalation [21], (ii) the accumulation of polymeric residues and mechanical warpage at vdW interfaces [22], and (iii) the large‐area scalability of vdW crystallinity [23]. These systematic limitations have motivated us to synthesize the designable vdW crystal and its corresponding functional optoelectrical synapse application. To overcome the intrinsic limitations of conventional vdW stacking and to realize brain‐inspired optoelectronic synapses, a designable vdW crystal has been presented with vdW nano‐crystallization with consistent grain boundary. As the grain boundary atomically confine the ionic migration pathways, thereby establishing deterministic control over synaptic weight updates. This structural similarity derives optoelectronic synaptic functionalities analogous to those of biological neurons, including multi‐level conductance modulation, long‐term potentiation/depression (LTP/LTD), paired‐pulse facilitation (PPF), and a tunable short‐term to long‐term memory (STM‐LTM) transition, governed by the interplay of photo‐induced carrier trapping and defect‐mediated ionic migration. Beyond elementary synaptic plasticity, the system reproduces higher‐order cognitive processes such as learning‐forgetting‐relearning cycles, demonstrating 34.7% enhanced retention efficiency compared to bulk ReSe2. Furthermore, functional demonstrations in image edge detection and CIFAR‐10 recognition results in a recognition accuracy of 96.24%. In conclusion, the designable vdW crystal will offer the versatile advances for complementary metal‐oxide‐semiconductor (CMOS) compatible, energy‐efficient, and 3D integrated neuromorphic vision architectures. Results and Discussion Designable vdW Artificial Crystals for Device‐Scale Neuronal Cell Operation As schematically illustrated in Figure 1a, Ca2 + and Na+ ions migrate across the neuronal membrane via optically gated ion channels [24], establishing the fundamental principle of visual signal transduction and memory formation. In biological neurons, the influx of Na+ ions through voltage‐ and light‐sensitive channels initiates depolarization and action potential generation [25], while Ca2 + entry at the presynaptic terminal regulates neurotransmitter release and downstream synaptic plasticity. Such spatiotemporally confined ionic transport governs the reconfiguration of local electrochemical gradients [26], modulating signal fidelity and memory consolidation. Conceptually, such biological ionic migration dynamics and ion channel structure in neuronal brain cells has motivated us to replicate the structural configuration and its synaptic functionality. To replicate the neuronal synaptic functionality in device‐scale, bulk vdW ReSe2 has been nano‐crystallized with Ar + H2S plasma, inducing the grain boundary formation and sulfur intercalation. Our previous research revealed the number of vdW layer of the nano‐crystallized vdW materials (Bi2Se3 [27], VSe2 [28, 29]) can be precisely modulated with radio frequency (RF) plasma power variation (0 W∼400 W). The selection of vdW ReSe2 is justified by four intrinsic properties: (1) the vdW architecture provides self‐limiting nano‐crystallization by confining ion‐induced structural modification to the upper layers without disrupting the bulk lattice, (2) the distorted 1T′ triclinic structure (space group P̄1) [30, 31] imposes directional selectivity on S2 − migration along grain boundaries, enabling deterministic synaptic weight control unattainable in isotropic high‐symmetry vdW materials such as 2H‐MoS2, (3) the moderate bandgap (approximately 1.1–1.3 eV) is matched to visible‐light excitation at 515 nm, and (4) the vdW integration capability enables substrate‐agnostic fabrication essential for 3D stackable neuromorphic architectures. Owing to the nano‐crystallization, Raman A 1g peak has been decreased, while E 2g peak was reversibly increased. Such Raman peak redistribution corresponds to the structurally‐reconstructed ReSe2, validating the grain boundary and sulfur intercalation (Figure S1). A cross‐sectional transmission electron microscopy (TEM) image of nano‐crystallized ReSe2 (Figure 1b) reveals an atomically layered configuration, which constructed to replicate the ion‐channel architecture of neuronal membranes. This structural similarity indicates that the nano‐crystalline ReSe2 lattice can function as an optoelectrical ionic channel. The schematic of a neuron‐synapse junction (Figure 1c) illustrates the canonical pathway of action potential transmission, which is conceptually traversed into the operation of the vdW optoelectronic synaptic device. Analogous to Ca2 +/Na+ channels [32], the regulated migration of sulfur atoms through grain boundary confinement dynamically alters the internal potential landscape, enabling precise control of conductance states and artificial synaptic weight adaptation. The device structure (Figure 1d) is consists of pre‐synaptic contact, post‐synaptic contact, and optoelectrical spike at nano‐crystalline ReSe2. Additionally, the sequential dynamics of photo‐excitation and carrier trapping has been illustrated in Figure 1e. After optoelectrical spike, the carrier accumulation is intrinsically coupled to the intra‐/inter‐layer migration of nano‐crystalline ReSe2. Such atomic migration sequentially reconfigures the local potential landscape and defect‐state occupancy, thereby modulating the synaptic conductance and replicating the synaptic weight adjustment in brain cell. Such synaptic regulation enables vdW nano‐crystallization as a robust material platform for emulating synaptic plasticity and memory consolidation, highlighting the potential of the artificial designable vdW crystal to operate with device‐scale brain cell mimicking operations. Designable vdW crystal for mimicking the neuronal light‐sensitive ion channel. (a) Schematic illustration of the neuronal cell operation mechanism, migrating the Ca2+ and Na+ ion through light‐sensitive ion channel. (b) Cross‐sectional TEM image of nano‐crystallized ReSe2 crystal, mimicking the cross‐sectional configuration of light‐sensitive ion channel in neuronal cell membrane. (c) Schematic of a neuron and synapse, operating with transmission of the action potential. After initiation of the external stimulation, neurons transmit signal further to other neurons. (d) Structural configuration of vdW optoelectronic synaptic device for mimicking the brain cell operation, consisting of pre‐synaptic contact, post‐synaptic contact, and optoelectrical spike in nano‐crystallized ReSe2. (e) Sequential carrier dynamics of the photo‐excitation process and the charge‐trap mechanism in nano‐crystallized ReSe2. Sequentially‐Resolved Synaptic Operation Dynamics for Synaptic Weight Updates To reveal the mechanism of the ionic migration behavior in the artificial vdW crystal, the sequential synaptic operation dynamics were systematically investigated. The synthetic pathway of ReSe2 nano‐crystallization (Figure 2a) has been configured with lattice destabilization, sulfur intercalation, and subsequent nano‐crystallization. The corresponding cross‐sectional TEM image (Figure 2b) confirms the emergence of nano‐crystalline grain and sulfur intercalation. The transition from bulk ReSe2 to nano‐crystalline ReSe2 has been further resolved within cross‐sectional STEM imaging (Figure S2). In the first stage (lattice destabilization), Ar+ ion bombardment disrupts the long‐range periodicity of the upper layers while the underlying bulk lattice and intrinsic vdW gap are preserved, as confirmed by the reduction of the Raman A 1g peak intensity and the loss of periodic lattice fringes confined to the upper region. In the second stage (sulfur intercalation), H2S+ ions generated through Penning ionization (Equations (1), (2), (3)) penetrate the destabilized region, yielding sulfide species that occupy Se vacancy sites or intercalate into the expanded vdW gaps. The progressive emergence of Re–S bonding in XPS Re 4f spectra (Figure S3), absent under Ar‐only treatment at equivalent power, provides chemical confirmation, and spatial EDS mapping (Figure S4) validates sulfur localization within the treated region. In the third stage (nano‐crystallization), sulfur saturation generates local strain fields that drive recrystallization into discrete nano‐grains with lateral dimensions of 5–10 nm. To further investigate the nano‐crystallization effect, optical microscopy (OM) and conductive atomic force microscopy (C‐AFM) measurements were conducted, resolving the nanograin formation and its average domain size distribution as 551.55, 2123.72, and 7542.96 nm2 (Figures S5 and S6). To further validate the synaptic operation mechanism and to decouple the grain boundary (GB) contribution from potential bulk defect‐mediated conduction, a designable vdW crystal and its corresponding conductance mapping was resolved within RF plasma power variation (300, 350, and 400 W). As depicted in the C‐AFM images, the increment of RF plasma power results in the GB density enhancement. Furthermore, the spatial conductance maps reveal that the conduction is predominantly localized through the structural ReSe2 GB, providing the direct evidence of local conduction confinement. During Ar + H2S plasma treatment, Penning ionization and ion penetration processes are activated. Specifically, the interaction of Ar gas with energetic electrons generates Ar+ ions (Equation 1), which subsequently transfer charge to H2S molecules to yield H2S+ (Equation 2) [33, 34]. (1) Ar+e−→Ar++2e− (2) Ar++H2S→Ar+H2S+ Sequentially‐resolved synaptic operation mechanism in nano‐crystalline ReSe2. (a) Sequential illustration of synthetic process of ReSe2 nano‐crystallization, consisting of the lattice destabilization, sulfur intercalation, and nano‐crystallization state and (b) its corresponding cross‐sectional TEM image of nano‐crystallized ReSe2. (c) Arrhenius plot of nano‐crystallized ReSe2, exhibiting the sequential shift of phase transition temperature (1) carrier transport dominant, (2) carrier/ion transport co‐existence, (3) ion transport dominant) break point (1000/T, K−1). (d) Linear increment of phase transition temperature (break point, 1000/T, K−1) within RF plasma power (300, 350, 400 W). (e) Sequential increment of ion transport dominance (J ion/J ion+J carrier) within temperature variation (1000/T, K−1). (f) Activation energy of ion transport and carrier transport. (g) Ion transport dominance (J ion/J ion+J carrier) comparison within variation of sulfurization condition. In parallel, sufficient free electrons within the plasma can directly ionize H2S to form H2S+, as shown in Equation (3). Without Ar gas, it is difficult to generate H2S plasma despite the low ionization energy of H2S gas [35]. The Penning ionization mechanism (Equations (1), (2), (3)) generates H2S+ ions that, upon dissociative impact with the ReSe2 surface, yield sulfur species incorporated into chalcogenide lattice sites as S2 −, consistent with vacancy‐mediated substitution reported in other vdW systems [36, 37]. The migration of S2 − ions along grain boundary pathways and the transport of sulfur vacancies (VS) in the opposite direction constitute complementary descriptions of the identical physical displacement within the lattice. (3) e−+H2S+→2e− Without Ar gas assistance, H2S plasma generation is inefficient despite its relatively low ionization energy, highlighting the necessity of gas mixture optimization. For precise control of sulfurization, an Ar + H2S gas mixture ratio of 1:1 (50 sccm for each) was adopted. Consistent with previous reports, modulation of RF power enables controlled ion penetration through the vdW lattice. Sequential TEM analysis demonstrates nano‐crystallization, with nano‐grain formation (5∼10 nm lateral dimension) arising from lattice distortion induced by electrical field‐driven ion bombardment. The designability of synaptic functionality in nano‐crystalline ReSe2 is substantiated through the systematic modulation of RF plasma power (300‐400 W). Arrhenius analysis reveals a linear shift in phase transition temperatures (break point, 1000/T, K−1), indicating three distinct phases as (1) carrier‐dominant regime, (2) carrier/ion coexistence regime, and (3) ion‐dominant transport regime. The internal charge transport mechanisms were resolved within temperature‐dependent Arrhenius analysis (Figure 2c). The Arrhenius plots exhibit a distinct bifurcated behavior, characterized by a phase transition temperature (break point, 1000/T, K−1) that separates carrier transport phase and ion transport phase. Notably, this transition temperature linearly increases as a function of the RF plasma power, demonstrating that the threshold for ion transport is systematically tunable via the sulfurization intensity (Figure 2d). The quantitative shift in the dominant transport is further substantiated by the ion transport dominance ratio (J ion/J ion+J carrier). As the temperature increases, the ratio of ion transport sequentially rises within higher RF powers, yielding an increment of ionic dominance across the measured temperature range (1000/T, K−1) (Figure 2e). Under the presence of H2S plasma (Ar + H2S 400 W), the ion transport dominance surpasses 95.61%, whereas the absence of H2S plasma (Ar 400 W) remains predominantly carrier transport dominance with a ratio of 23.76% (Figure 2g). Such transition is fundamentally based in the selective modification of kinetic barriers. The functional role of H2S plasma in synaptic modulation was quantitatively resolved with LTP/LTD measurements. A comparative synaptic behavior of pristine ReSe2, ReSe2 (Ar plasma 400 W), and ReSe2 (Ar + H2S plasma 400 W) indicates a significant disparity in synaptic behavior. While Ar plasma treatment yields an increase in the maximum post synaptic current to 11.52 nA, the addition of H2S plasma results in a 2.25‐fold enhancement, reaching 25.88 nA (Figure S7). To further demonstrate the reproducibility, a statistical analysis has been conducted within 25 devices. First, the devices exhibit a remarkable yield of 92.0% with high uniformity across all critical metrics. The PPF index shows a mean value of 184.25% with a low coefficient of variation (CV) of 4.03%, indicating consistent short‐term plasticity. Furthermore, the linearity of the weight updates, which is essential for high‐accuracy hardware neural networks, remains consistently high with an average value of 0.92. The device‐to‐device variability was further quantified via the dynamic range, yielding a mean of 1.25 × 104. The energy dissipation per synaptic event was evaluated under typical operating conditions to assess the potential for low‐power computing, featuring a mean energy consumption of 380.12 fJ per pulse. To ensure the integrity of these synaptic weights during operation, a read voltage (Vread) of 0.10 V has been selected. This specific voltage was selected to provide a sufficient signal‐to‐noise ratio for weight retrieval while remaining well below the threshold required for ionic migration or carrier injection. Consequently, the architecture achieves effective read‐write decoupling, ensuring that inference operations do not alter the programmed synaptic states. The devices maintained stable conductance switching over a mean endurance of 2.13 × 105 pulses. Additionally, the retention characteristics confirm that the programmed states are preserved for over 4.01 × 104 s. Activation energy (Ea) analysis reveals that while the carrier activation energy remains relatively stable (0.23 eV), the ion activation energy exhibits a consecutive increment within H2S plasma sulfurization intensity (Figure 2f). The ion activation energies in the present work (approximately 0.3–0.9 eV) are systematically higher than those reported for Na+ transport through voltage‐gated sodium channels (0.1–0.3 eV) (B. Hille, Ion Channels of Excitable Membranes, third ed., Sinauer Associates, Sunderland, MA 2001.) and Ca2 + transport through voltage‐gated calcium channels (0.2–0.5 eV) [38]. This disparity originates from the distinct dielectric environments: biological ion channels operate in an aqueous medium (dielectric constant of approximately 80) that screens Coulombic interactions and reduces migration barriers, whereas sulfide ions in nano‐crystalline ReSe2 migrate through a crystalline lattice (dielectric constant of approximately 7–15) where each ionic hop requires displacement through a covalently bonded construction without solvent‐mediated screening. The elevated barrier confers a functional advantage for neuromorphic operation: room‐temperature thermal energy (approximately 26 meV) is insufficient to drive spontaneous back‐diffusion across barriers of 0.3–0.9 eV, conferring pronounced non‐volatility on programmed conductance states with retention times exceeding 4.01 × 104 s. Biological synapses, operating with lower barriers, require continuous metabolic expenditure to sustain synaptic states against ionic leakage. Optoelectrical Synaptic Operation of Nano‐Crystalline ReSe2 The optoelectronic synaptic characteristics of the nano‐crystalline ReSe2 were systematically examined to evaluate its capability for brain‐inspired synaptic processing. As shown in Figure 3a, the synaptic weight exhibited a stepwise cascade response as the number of optical pulses (5 pulses, 9 pulses, 10 pulses, 15 pulses, and 20 pulses) increased, thereby demonstrating multi‐level conductance modulation analogous to incremental potentiation in biological synapses. This progressive conductance tuning highlights the feasibility of employing optically driven spike inputs to encode synaptic weights in a controllable and reproducible manner. The LTP and LTD characteristics of the device are clearly observed (Figure 3b), confirming the ability of nano‐crystalline ReSe2 to emulate bidirectional synaptic plasticity. Notably, the application of a single optoelectrical spike (pulse width 0.5 s and pulse interval as 1.5 s) under a bias of +5 V/−5 V with increasing laser power density (0∼250 mW/cm2 and wavelength 515 nm) induces a distinct transition from short‐term memory (STM) to long‐term memory (LTM), as depicted in Figure 3c. Comparative LTP/LTD measurements on destabilized, S‐intercalated, and nano‐crystallized ReSe2 devices under identical pulse conditions (±5 V, 5 cycles, 200 total pulses, 515 nm) confirm the structural necessity of nano‐crystallization (Figure S8). The nano‐crystallized device exhibited the highest peak post‐synaptic current of 3.55 nA, exceeding the S‐intercalated (2.56 nA) and destabilized (1.09 nA) counterparts, and was the only phase to demonstrate symmetric potentiation and depression approximating ideal analog weight updates. The performance disparity is attributable to the quality of grain boundary ionic transport in each phase. In the destabilized phase, interfacial boundaries separate disordered regions with broadly distributed migration barrier heights, resulting in stochastic ionic transport and severely limited dynamic range. The S‐intercalated phase shows intermediate performance, where sulfur disrupts interlayer coupling and produces inconsistent migration pathways. Only the nano‐crystallized phase provides lattice‐coherent grain boundaries with uniform barrier heights for deterministic S2 − transport, as corroborated by the distinct carrier/ion transport bifurcation in the Arrhenius analysis (Figure 2c) and localized conduction in C‐AFM mapping. In this process, short‐lived conductance states generated by transient carrier excitation are stabilized into long‐lasting conductance states through charge trapping and ionic migration across grain boundaries. This transition mechanism mimics the biological consolidation of STM into LTM, wherein repeated or sufficiently strong stimuli trigger durable synaptic modifications. The efficiency of dynamic synaptic response was further assessed by paired‐pulse facilitation (PPF). The PPF index of nano‐crystalline ReSe2 markedly exceeded that of bulk ReSe2 (Figure 3d), evidencing the critical contribution of nanoscale crystallization and grain boundary engineering to enhanced synaptic plasticity. Such an enhancement arises from accelerated ionic transport and more effective defect‐mediated carrier trapping, both of which amplify the temporal correlation between successive stimuli. To evaluate the long‐term operational stability of the nano‐crystallized ReSe2 optoelectronic synaptic device, LTP/LTD cycling measurements were performed on the identical device over a period of 15 days under ambient storage conditions (room temperature, without encapsulation). As presented in Figure S9, the device exhibited a peak post‐synaptic current of 4.53 nA after 1 day, which decreased gradually to 4.31 nA (day 5), 4.17 nA (day 10), and 3.89 nA (day 15), corresponding to a retention of approximately 85.9% of the initial performance. No abrupt degradation or sudden current collapse was observed throughout the measurement period, and the LTP/LTD curve profiles remained qualitatively consistent, preserving the gradual potentiation and complete depression characteristics that are essential for analog synaptic weight modulation. The absence of distorted potentiation behavior or incomplete depression across all time points indicates that oxide formation does not penetrate the grain boundary pathways responsible for S2 − ionic migration, as such penetration would induce additional scattering centers that disrupt the deterministic energy landscape governing conductance modulation. The functional integrity of the potentiation and depression profiles thereby confirms that the S2 − ionic migration pathways at grain boundaries are not compromised by ambient exposure. The gradual decline in peak current, at a rate of approximately 1% per day, is consistent with self‐limiting surface oxidation confined to the uppermost layers of the heterostructure [39]. The sequential evolution of post‐synaptic current (Figure 3e) can be delineated into four stages (1) photo‐induced carriers, (2) modulation of band barriers at grain boundary sites, (3) partial recombination of photo‐carriers via radiative pathways, and (4) charge accumulation within localized defect states. These processes collectively underpin the persistent photo‐conductivity effect, which serves as the physical basis of synaptic weight retention. The corresponding band diagram (Figure 3f) illustrates the defect‐state excitation and subsequent modulation of conduction pathways, establishing a direct link between atomic‐scale carrier dynamics and memory behavior. The optoelectronic synaptic response in nano‐crystalline ReSe2 arises from carrier‐mediated short‐term plasticity and ion‐mediated long‐term memory. The non‐volatile component originates from field‐driven S2 − migration along grain boundary channels, which reconfigures the local potential landscape to produce persistent conductance modulation. The Arrhenius analysis confirms this: ion transport dominance reaches 95.61% under Ar + H2S 400 W, with a 2.25‐fold enhancement in maximum post‐synaptic current. The grain‐boundary ion channels thus gate sulfide ion transport analogously to how Ca2 +/Na+ channels gate ionic fluxes in biological synapses. Tip‐enhanced photoluminescent (PL) and electroluminescent (EL) measurements reveal that nano‐crystalline ReSe2 exhibits distinct spectral broadening and peak separation compared to bulk ReSe2, indicative of defect‐mediated carrier localization and enhanced excitonic coupling. Under progressive optoelectrical stimulation, the modulation of PL and EL responses directly correlates with the transition from short‐term to long‐term memory states, confirming that persistent photoconductivity and ion migration at grain boundaries provide the physical basis for durable synaptic weight retention. These observations validate that the spectroscopic evolution of emission characteristics serves as direct evidence of STM‐LTM conversion in vdW synaptic devices (Figures S10–S12). Hence, these observations confirm that optoelectrical spike‐driven synaptic plasticity in nano‐crystalline ReSe2 arises from the synergistic interplay of photo‐carrier dynamics and ion migration at grain boundaries. The precise tunability of STM‐LTM transition through optical pulse parameters (bias amplitude, duration, and repetition pulse number) underscores the potential of the designed vdW crystal as a robust platform for brain‐inspired neuromorphic systems. By coupling optically induced charge dynamics with atomically engineered migration pathways, nano‐crystalline ReSe2 demonstrates the capability to reproduce memory formation and consolidation processes characteristic of biological neuronal networks. Optoelectrical synaptic functionality of nano‐crystalline ReSe2. (a) Synaptic cascade of pulse number variation (5, 9, 10, 15, and 20 pulses). (b) LTP/LTD measurements and (c) single optical spike with +5 V / −5 V pulse configuration with laser power density (0∼250 mW/cm2), converting the short‐term memory to the long‐term memory. (d) PPF index comparison of nano‐crystalline ReSe2 and bulk ReSe2 and (e) its corresponding post synaptic current, sequentially deriving the (1) light stimulation, (2) band barrier modulation, (3) partial recover (radiative recombination), and (4) internal charge accumulation. Optoelectrical pulses are configured with wavelength 515 nm, laser power density (0∼250 mW/cm2), pulse width 0.5 s, and pulse interval as 1.5 s. (f) Band diagram for optically‐induced carrier dynamics, illustrating the defect band excitation and its corresponding photo‐conductance variation. Memory‐Formation Process in Artificial vdW Crystal for Brain Cell Operation The programmable synaptic functionalities of nano‐crystalline ReSe2 were further substantiated through pulse‐driven post synaptic current measurements. LTP/LTD characteristics were measured over 1000 operational cycles, exhibiting a progressive expansion of the dynamic current range within RF plasma power increment (Figure 4a). The reproducibility was validated within the consistent overlap of 25 consecutive cycles, which indicates high cyclic stability and minimal degradation across the various sulfurization conditions (Figure 4b). To further validate the spatial reliability and device‐to‐device variation, 25‐pixel mapping of the post synaptic current was performed within the various sulfurization conditions (Figure 4c–f). This mechanism demonstrates that nano‐crystalline ReSe2 can act as an effective analogue of biological memory consolidation, wherein repetitive or sufficiently intense stimulation extends synaptic modifications from volatile to non‐volatile regimes. A comparative learning‐forgetting‐relearning experiment further highlights this functionality (Figure 4g). During the initial learning phase, nano‐crystalline ReSe2 exhibited a post‐synaptic current of 1.98 nA, significantly surpassing the 1.10 nA observed in bulk ReSe2 under identical stimulation conditions. Following the forgetting phase, where conductance decayed due to carrier recombination and redistribution of sulfur atoms, the relearning process revealed pronounced retention. Specifically, nano‐crystalline ReSe2 required only 23 pulses to restore the same synaptic current level, whereas bulk ReSe2 required 31 pulses. Learning‐forgetting‐relearning process was configured within sequential pulse measurements (first learning as 30 pulses), cognitive decay (forgetting process), and second learning process. Retention efficiency η is defined as the reciprocal of the pulse count required to reinstate the synapse to its potentiated state (1.10 nA). As the nano‐crystalline ReSe2 attained the target current (1.10 nA) within 23 pulses, bulk ReSe2 results in the 31 pulses, representing the 34.7% improvement in retention efficiency η (Nnano−NbulkNbulk). This reduced stimulation threshold evidence superior relearning efficiency, attributable to residual defect‐state occupancy and metastable sulfur configurations that preserve a trace memory of previous excitations. Such behavior parallels biological synapses, where prior activation history reduces the threshold for subsequent plasticity induction through synaptic meta‐plasticity. Furthermore, synaptic performance benchmark has been provided to compare the relative device performance of nano‐crystalline ReSe2, including FeFETs (bulk oxide) [40], FTJs (bulk oxide) [41], RRAMs (bulk oxide) [42], bulk vdW materials [43], and phase‐change materials (PCMs) [44] (Figure S13). While it is acknowledged that synaptic device performance can exhibit significant variations depending on specific material compositions, device dimensions, and measurement protocols, the data points for comparison were selected from representative reports to reflect the typical performance limits of each technology. The evaluation considers eight key synaptic parameters, which widely represents the typical device performance and characteristics (energy efficiency, switching ratio, endurance, linearity, PPF index, dynamic range, retention, and cycling endurance). While conventional synaptic devices often exhibit optimized performance in specific metrics, encountering the trade‐offs in linearity and the PPF index. In comparison, the nano‐crystalline ReSe2 device demonstrates balanced characteristics across all evaluated parameters. Notably, it maintains stable long‐term retention and a broad dynamic range while simultaneously achieving improved linearity in potentiation and depression. Memory‐formation behavior with synaptic plasticity for brain cell mimicking. (a) Synaptic pulse operation and its corresponding post synaptic current behavior (LTP/LTD, 1000 pulse operation). (b) 25 cycle overlap of sulfurization condition variation ((1) Ar 400 W, (2) Ar + H2S 300 W, (3) Ar + H2S 350 W, (4) Ar + H2S 400 W). 25‐pixel mapping for post synaptic current (nA) of (c) pristine ReSe2, (d) ReSe2 (Ar + H2S 300 W), (e) ReSe2 (Ar + H2S 350 W), (f) ReSe2 (Ar + H2S 400 W). (g) Learning‐forgetting‐relearning behavior of comparison with nano‐crystallized ReSe2 and bulk ReSe2. As post synaptic current results in 1.98 nA (nano‐crystallized ReSe2) and 1.10 nA (bulk ReSe2) with first learning process (30 pulse), required pulse number for achieving the same post synaptic current (1.98 and 1.10 nA) distributes as 23 pulses (nano‐crystallized ReSe2) and 31 pulses (bulk ReSe2). CIFAR‐10 Image Recognition Task With Optoelectronic vdW Synaptic Performance The CIFAR‐10 image recognition task was conducted to validate the capability of the artificial vdW synaptic device, thereby bridging fundamental synaptic plasticity with practical cognitive functionality (Figure 5a). As an initial demonstration, the nano‐crystalline ReSe2 synaptic behavior was applied to edge detection of natural images (Figure 5b). Distinct vertical, horizontal, and combined edge features were successfully extracted, evidencing that gradual and reversible modulation of synaptic weights can be directly translated into convolutional operations central to vision‐based computing. This proof‐of‐concept confirms that optoelectronic regulation of synaptic states enables device‐scale feature extraction, a critical requirement for higher‐order image classification. The recognition performance of the device was further evaluated using the CIFAR‐10 dataset (airplane, automobile, bird, cat, deer, dog, frog, horse, ship, and truck). Post synaptic current variation in nano‐crystalline ReSe2 was mapped onto synaptic weights within a convolutional neural network composed of convolutional and fully connected layers. Our designable vdW crystal achieved a recognition accuracy as 96.24%, overcoming the conventional optoelectronic synaptic platforms. The recognition accuracy matrix (Figure 5c) demonstrates consistent classification across all object categories, confirming the device robustness system in handling heterogeneous datasets (Figures S14 and S15). Collectively, the edge detection and CIFAR‐10 recognition tasks establish a direct translation of device‐level synaptic plasticity into system‐level neuromorphic computation. Regarding the experimental synaptic behaviors (long‐term potentiation/depression (LTP/LTD), paired‐pulse facilitation (PPF), and short‐term to long‐term memory (STM‐LTM) transition), the synaptic performance of nano‐crystalline ReSe2 was extracted into a hardware neural network (Figure 5d). The integration of validated synaptic behaviors (LTP/LTD, STM‐LTM transition, and relearning efficiency) with image recognition performance highlights the dual capability of nano‐crystalline ReSe2 to replicate neuronal operation and to extend them toward scalable artificial intelligence applications. The performance of artificial neuronal cell as recognition accuracy (96.24%) defines a significant benchmark for vdW optoelectronic synapses, enabling the brain‐inspired synaptic functionality with CMOS compatibility, low energy consumption, and real‐time adaptive learning. CIFAR‐10 dataset recognition task with designable vdW crystal. (a) CIFAR‐10 dataset recognition task with nano‐crystalline ReSe2 synaptic device. (b) Edge detection of source image (Chicago river walk), configured with vertical edge, horizontal edge, and combined edge. (c) Recognition accuracy matrix for the CIFAR‐10 dataset, deriving from the I‐V characteristics form nano‐crystalline ReSe2, resulting in the 96.24% recognition accuracy. Configured with CIFAR‐10 pre‐dataset (airplane, automobile, bird, cat, deer, dog, frog, horse, ship and truck), convolutional layer, and fully connected layer, (d) hardware neural network and its corresponding circuit diagram (synaptic devices) for fully connected layer, which directly mimics the brain cell operation in device‐scale. Conclusion A designable vdW artificial crystal based on nano‐crystalline ReSe2 has been established as a materials platform for device‐scale neuronal cell mimicking. Plasma‐assisted nano‐crystallization with sulfur intercalation generates grain‐boundary networks that act as atomically confined pathways for ionic motion. Temperature‐dependent Arrhenius analysis and C‐AFM conductance mapping resolve the sequential transition from carrier‐dominant to ion‐dominant transport, with field‐driven S2 − migration along grain boundary pathways providing deterministic control of the final ion distribution. This atomic‐scale reconfiguration modulates defect occupancy and local potential landscapes, yielding programmable conductance states that operate as synthetic synaptic weights. Designable vdW artificial crystal exhibits hallmark features of biological synapses, including the multi‐level conductance modulation, long‐term potentiation/depression, paired‐pulse facilitation, and a tunable short‐term to long‐term memory transition governed by the interplay of photocarrier trapping and ion migration. Beyond elementary plasticity, higher‐order cognitive analogues are reproduced. In learning‐forgetting‐relearning protocols, nano‐crystalline ReSe2 attains a larger initial post‐synaptic current (1.98 nA/1.10 nA) and requires fewer pulses for recovery after forgetting (23 pulses / 31 pulses), evidencing enhanced retention and meta‐plasticity arising from metastable ionic configurations. Device‐scale edge detection verifies that reversible weight updates implement convolution‐like operations, while CIFAR‐10 inference achieves 96.24% accuracy using arrays mapped to convolutional and fully connected layers. The combination of broad dynamic range, stable retention, and reproducible updates. In conclusion, we envision that designable vdW artificial crystal enables energy‐efficient, CMOS compatible, and 3D stackable neuromorphic vision architectures in which sensing, memory, and processing converge. Experimental Section/Methods Mechanical Exfoliation and Transfer of vdW ReSe2 Before mechanical exfoliation and dry transfer, a polydimethylsiloxane stamp was attached to a glass cover. vdW ReSe2 were mechanically exfoliated from bulk crystals (HQ Graphene, Netherlands) onto polydimethylsiloxane stamps and then transferred onto the substrate by applying a transfer condition of 70°C. vdW 2D/3D Heterostructures Fabrication for Monolithically‐Stacked Artificial Synapse In vdW 2D/3D heterostructures, the nano‐crystallized vdW lattice functions as the resistive element, while the bulk vdW lattice serves as the selector, enabling monolithically‐stacked artificial synapse. To induce nano‐crystallization of ReSe2, inductively coupled plasma‐enhanced chemical vapor deposition (ICP‐PECVD, AFS‐IC6T, Korea) was employed. Prior to plasma treatment, the chamber was evacuated to a high vacuum (∼10− 5 Torr) to eliminate contaminants and suppress undesired reactions during synthesis. The RF plasma power was maintained at 400 W, with constant gas flow conditions of Ar and H2S (50 sccm each) at 25 mTorr pressure and room temperature, ensuring fabrication of vdW 2D/3D heterostructures. SPM Measurements SPM (NX‐10 AFM, Park Systems, Republic of Korea) measurements were conducted with an Electri‐Multi75G cantilever. Customized SPM device set‐up was installed with Mibot (IMINA, Switzerland) and AFM (NX 10, Park Systems, South Korea) with Tungsten probe (PT‐010N‐15‐B8) with 10 nm tip radius, 15 mm in length), enabling the electrical pulse measurements and I‐V curve sweep (Bias range ‐170 V∼+170 V, 60 Hz, AC ±107 V). A silver paste electrode (Elcoat P‐100, CANS, Japan) was selectively deposited on the sample edge to induce electrical contact. The ElectriMulti75‐G cantilever was calibrated with a tip radius of 25 nm, a length of 225 µm, a height of 17 µm, a width of 28 µm, and a spring constant of 3.3 N/m, resulting in a resonance frequency of 60.8 kHz. C‐AFM measurements with a continuous voltage waveform, which was configured with 8 s for one cycle (3 cycle measurements), and sample bias range as −10 V ∼ +10 V (Vmax as +10 V/ Vmin as −10 V). Current‐voltage characteristics through the sulfurized vdW materials were probed within the CAFM probe (Electri‐Multi75G, tip radius as 25 nm) to operate as the top electrode of the vdW materials, grounded to the Au substrate. Material Characterization XPS measurements (XIS Supra+, Kratos, United Kingdom) were used to characterize vdW materials, with an x‐ray spot size of 400 µm. Peak deconvolution was performed with the profiles aligned using the C 1s peak at 285 eV. The XPS data were calibrated using the CASAXPS software (version 8.1). Optical microscopy (U‐MSSP4, Olympus, Japan) and FE‐SEM (S‐4800, Hitachi, Japan) were used to examine the transferred flakes. For cross‐sectional TEM specimen preparation, a focused ion beam instrument (NX2000, Hitachi Ltd., Japan) was used, employing a Ga+ ion beam (30–5 keV) and a lift‐off process to etch the specimens. TEM (JEM‐2100F, JEOL, Japan) and XRD (Empyrean, Malvern PANalytical, United Kingdom) were used to observe the lattice structure, EDS, and SAED patterns of the sulfurized ReSe2 structures at the atomic scale. Convolutional Neural Network The CNN used for CIFAR‐10 classification was implemented using a standard two‐convolutional‐layer architecture. Input images (3 × 32 × 32 RGB) were processed through a first convolutional layer consisting of 32 kernels (3 × 3), followed by 2 × 2 max pooling. A second convolutional layer with 64 kernels (3 × 3) and an additional pooling stage further reduced the feature maps to 64 × 8 × 8. The resulting 4096‐dimensional vector was passed through two fully connected layers (4096→128 and 128→10). Network training was performed using backpropagation with cross‐entropy loss. After training, the synaptic weights were quantized and mapped to the experimentally measured multilevel conductance states of synaptic devices. Classification accuracy was evaluated by replacing software weights with the corresponding synaptic devices. Analog Convolution Image Processing Method Analog convolution experiments were conducted by mapping pixel intensities to input voltages and encoding kernel weights into the multilevel conductance states of the array. MAC operations were performed through conductance‐weighted current generation (Ohm's law) and current summation along bit lines (Kirchhoff's law), enabling vector matrix multiplication in the analog domain. Edge‐detection tests were carried out on a 6 × 6 pixel input image using 3 × 3 convolution kernels. Because positive and negative weights were required, each kernel was differentially encoded using two adjacent columns array (positive weights in the first column, negative weights in the second). Representative filters (smoothing, sharpening, vertical‐edge, and horizontal‐edge) were implemented through column‐wise conductance programming. Author Contributions J. Lee, Gunhyoung Kim, D. Lee, Seowoo Son, and H. Seok prepared samples and performed experiments. Sihoon Son, S. Kim, E. Go, H. Choi, J. Cho, G. Back, Geonwook Kim, D. Choi, and H. Shim were included in the technical discussions on vdW 2D/3D heterostructures. Hyunkyu Kim, C. Park, J. Ahn, S. Je, and C. Im were included in the additional XPS spectroscopy and TEM/EDS experiments. M. Grzeszczyk was included in Raman spectroscopy measurements. M. Kim, Hyoeng‐U Kim and W. Jang included in the optical synapse mechanism and SPM experiments. J. Lee, Gunhyoung Kim, D. Lee, Seowoo Son, and H. Seok wrote the manuscript with contributions from all the authors. T. Kim designed and supervised the study. All the authors have read and approved the final version of this manuscript. Conflicts of Interest The authors declare no conflicts of interests. Supporting information Supporting file: adma73595‐sup‐0001‐SuppMat.docx.