Dynamic flexible optical sensing based on liquid integrated circuits

Abstract

Vision supplies more than 80% of the information, and innovations in artificial vision can catalyze a foundational shift for technologies such as embodied intelligence. Dynamic flexible optical sensing (DFOS), where the sensor itself modulates curvature on demand, offers curvature-matched optical acquisition, focus switching without moving optics, and multiplane light-field acquisition, yet repetitive and reversible deformation is still challenging for recent electronic arrays. Here, we propose a DFOS system based on liquid integrated circuits (LICs), integrating photosensitive ionic liquid (PIL) and stretchable liquid metal electrodes to achieve active curvature modulation for geometry-adaptive optical sensing. The PIL exhibits selective color perception, reversible photo-thermo-electric responses, and robust stability. This all-liquid architecture allows for large-amplitude hydraulically controlled curvature modulation (>31.9°) with a strain rate of ∼1.62% per second and exceptional durability (600 reversible cycles) without mechanical fatigue. With the assistance of machine learning, the system demonstrates multimodal signal acquisition (concave, planar, and convex configurations) for color, pattern, and distance recognition with high accuracy through hydraulically controlled curvature modulation.

INTRODUCTION

More than 80% of the information we perceive is acquired through visual systems (1, 2), which also serve as the primary sensory systems in technologies such as embodied intelligence (3–5). Historically, the development of artificial visual systems has predominantly focused on lens-based imaging systems. For more than seven decades, mainstream efforts have largely been directed toward the development of lens systems, specifically tailored to accommodate static planar light sensors (Fig. 1A) (6). However, planar sensors intrinsically suffer from optical aberrations like field curvature (Petzval surface) and vignetting, necessitating complex, bulky, and expensive multilens assemblies for correction (7–10). Compared with the notable advances in lens technology, sensor-side optimizations to overcome these physical limitations have remained relatively underdeveloped.

Fig. 1. Concept of DFOS and properties of PIL and the pure IL.

(A) Schematic diagram of the multiple lens imaging and the DFOS. (B) Schematic diagram of LICs. (C) Optical images of the PIL containing 5 mg ml⁻¹ of FeCl₃, CoCl₂, and CuCl₂, respectively. (D) UV-Vis spectra of the PIL and the pure IL. (E) Thermal responses of the PIL and the pure IL over temperature changes. Error bars correspond to SD (n = 5). (F) Electrical responses of the PIL and the pure IL under the irradiation of light with different N values in the RGB model, respectively. Error bars correspond to SD (n = 5). (G) Electrical responses of the PIL and the pure IL after 1-min irradiation of red (N_r = 190), green (N_g = 190), and blue (N_b = 190) light, respectively. Error bars correspond to SD (n = 5). For the PIL evaluated in Fig. 1, the metal salt concentration was set to 5 mg ml⁻¹.

Current studies have achieved bio-inspired curved sensing (e.g., human or compound eye structures) through structural designs using rigid semiconductors like monocrystalline silicon and germanium (11–16), demonstrating wide field-of-view, deep depth-of-field, and low-aberration sensing capabilities (3, 4, 17). A few pioneering works of focus-tunable imagers have attempted to realize dynamic flexible optical sensing (DFOS) (18–20), which endows optical sensors with on-demand curvature modulation. DFOS is highly desirable as it not only reduces field curvature aberrations to reduce distortion and vignetting (fig. S2) but also achieves focus switching without lens movement through curvature-modulated equivalent optical path changes (see text S1 and fig. S1 in the Supplementary Materials). Besides, DFOS could enable simultaneous multifocal plane image capture through deformation-driven optical adaptation.

Although these efforts have achieved certain stretchability through geometric engineering, they remain constrained by the intrinsic mechanical limitations of solid-state optoelectronic materials and inherent sensing principles. Challenges include narrow deformation ranges, compromised mechanical durability, and the absence of color recognition capabilities. Traditional stretchable optoelectronics are prone to mechanical fatigue and structural resonance under dynamic conditions (21, 22). Furthermore, since optoelectronic materials inherently lack color discrimination capability, conventional planar optical sensors rely on Bayer filters, which are inherently difficult to integrate into deformable optical systems due to their rigid architecture.

To address these challenges, we propose a paradigm shift from solid-state geometric engineering to liquid-based intrinsic stretchability. Liquids have unparalleled ductility, allowing them to stretch and contract freely without the limitations imposed by solid-state materials (23, 24). This unique property enables liquid integrated circuits (LICs) to readily accommodate repetitive and reversible curvature modulations while maintaining robust signal fidelity. Moreover, by endowing the liquid sensing material with specific absorption properties, color discrimination can be achieved without external filters, overcoming the bottleneck of traditional color imagers.

Here, we present a fully integrated DFOS system based on LICs composed of photosensitive ionic liquid (PIL) and stretchable liquid metal electrodes (Fig. 1, B and C), enabling hydraulically driven adaptive optical control. Crucially, unlike solid-state sensors restricted by fracture limits, the inherent fluidity of our system enables unprecedented large-amplitude (modulation angle of >31.9°) and highly reversible morphological reconfigurations between concave, planar, and convex states, maintaining performance stability of more than 600 cycles without mechanical fatigue. Benefiting from this liquid architecture, the system exhibits exceptional environmental robustness, maintaining drift-free baselines even under continuous mechanical vibration and outdoor fluctuations. With the assistance of convolutional neural networks (CNNs) for signal decoupling, the system achieves high accuracy in multimodal recognition (color, pattern, and distance).

RESULTS

Properties of PIL

Ionic liquids (ILs) are materials that remain liquid at room temperature, entirely composed of cations and anions, exhibiting super-thermosensitive properties with pronounced temperature-dependent conductivity (25, 26). 1-Butyl-3-methylimidazolium dicyanamide ([BMIm][N(CN)₂]) was selected because its anions ([N(CN)₂]⁻) form supramolecular interactions with metal ions (27). By screening ultraviolet-visible (UV-Vis) absorption spectra of IL solutions containing different metal salts (Fig. 1D and fig. S3), we observed that unlike the pure IL, which exhibits negligible absorption across the visible spectrum, the introduction of metal ions markedly modulates the optical properties. FeCl₃, CoCl₂, and CuCl₂ were selected to prepare PIL (denoted as IL-Fe³⁺, IL-Co²⁺, and IL-Cu²⁺) due to their selective absorption of blue, green, and red light, respectively (Fig. 1C). Furthermore, the absorption intensity of these PIL exhibits a strong concentration dependence, progressively strengthening with higher metal salt concentrations in accordance with the Beer-Lambert law (fig. S4). Notably, these PILs with intrinsic color perception eliminate the need for RGB-based filters, substantially reducing light attenuation and avoiding complex device designs (28, 29).

Upon light absorption, a photo-thermo-electric process is initiated in PIL, involving sequential photothermal and heat-to-electricity conversions (fig. S5) (30, 31). The exceptional photothermal performance of PIL originates from fluorescence quenching, which will be detailed later. Regarding thermal response sensitivity, as shown in Fig. 1E, the temperature-dependent electrical responses ($∆G/G_0$, Eq. 1) of PIL and pure IL were measured. Their thermal responsiveness stems from heat-induced weakening of supramolecular interactions, which elevates free ion concentration and enhances conductivity. Furthermore, the incorporation of metal ions amplifies their thermal responsiveness due to the introduction of the coordination between metal ions and ILs. Cyclic testing between 25° and 50°C (fig. S6) demonstrates repeatable and reversible responses essential for sustained sensing performance, which indirectly confirms the dynamic nature of supramolecular interactions. In addition, thermogravimetric analysis (TGA) reveals PIL stability below 200°C without decomposition or volatilization (fig. S7), ensuring broad operational temperature ranges.

Systematic investigations have been conducted on the photosensitive capabilities of PIL. As illustrated in Fig. 1F, when exposed to blue, green, and red light, respectively, IL-Fe³⁺, IL-Co²⁺, and IL-Cu²⁺ exhibit enhanced electrical responses with increasing N values in the RGB model (N_r, N_g, and N_b), a phenomenon attributable to greater light intensity generating more heat (fig. S8). In contrast, the pure IL displays negligible electrical responses under identical illumination conditions. This control experiment confirms that the incorporation of metal ions is essential not only for modulating absorption wavelengths but also for substantially boosting the photothermal conversion efficiency. Besides, distinctive electrical responses emerged when exposing PIL to light stimuli with different colors but equivalent N values (N_r = N_g = N_b = 190), validating color-discrimination potential (Fig. 1G). It is noteworthy that the magnitude of the electrical response is not exclusively determined by the initial optical absorption intensity. As thoroughly discussed in text S2, the comprehensive light-to-electrical signal transduction in PIL is governed by a synergistic coupling of three distinct processes: wavelength-dependent light absorption ($\mathrm{\alpha }$), photothermal conversion efficiency ($\mathrm{\eta }$), and temperature-dependent ionic thermal sensitivity ($\mathrm{\delta }$). Consequently, the overall photoresponse ($∆G/G_0$) relies on the intricate interplay of heat generation, dissipation, and the specific temperature sensitivity of the PIL, rather than optical absorption alone. Furthermore, a matrix-based mathematical decoupling model confirms the system’s fundamental capability to resolve and recognize specific colors even under mixed broad-spectrum illumination (fig. S9). To assess the long-term stability of PIL, the electrical responses exhibited excellent stability throughout cyclic illumination testing (fig. S10) with randomly assigned light parameters (N_r = 168, N_g = 189, and N_b = 120). The photoresponsive characteristics remained invariant through 5-day prolonged testing (fig. S11), establishing foundational reliability for device integration.

Mechanistic analysis of interactions and fluorescence quenching

The reversible and stable photo-thermo-electric response of PIL stems from dynamic supramolecular interactions (Fig. 2A) (32), which were verified through x-ray photoelectron spectroscopy (XPS) analysis. As shown in Fig. 2B, the N 1s signals of the IL are split into three peaks at 397.94, 399.21, and 401.57 eV, corresponding to cyano-N, cyano-linked-N, and imidazolium-N, respectively. Those peaks associated with cyano-N and cyano-linked-N exhibit notable positive shifts (33). This observation suggests that those N atoms are donating electrons under metal ion coordination, driven by strong electron-withdrawing effects. Notably, the peaks related to cyano-N display larger positive shifts compared to those of cyano-linked-N, indicating that cyano-N may play a more dominant role in this process. In terms of metal ions, taking IL-Fe³⁺ as an example, the two main peaks at 710.74 and 723.94 eV are assigned to Fe 2p_3/2 and 2p_1/2 orbitals. Compared with the characteristic peaks of the FeCl₃ monomer, the Fe peaks in IL-Fe³⁺ shift toward lower binding energies (Fig. 2C). This shift suggests that the Fe³⁺ ions in the IL acquire more electrons, indicating stronger coordination interactions. The results confirm the coordination between Fe³⁺ ions and the anions ([N(CN)₂]^⁻), and it can be inferred that the coordination of Fe³⁺ ions with [N(CN)₂]^⁻ is stronger than that with Cl⁻. This conclusion is further supported by the negative shifts observed in the Cl element peaks (Fig. 2D). This behavior is attributed to weakened Fe-Cl interactions due to competitive coordination from [N(CN)₂]⁻. As Fe³⁺ preferentially coordinates with nitrogen atoms, the electrostatic interaction between Fe³⁺ and Cl⁻ is reduced, resulting in decreased polarization and a more localized electron density around Cl⁻, which leads to lower binding energies. Therefore, the concurrent negative shifts of Fe³⁺ and Cl⁻ binding energies arise from coordination-induced electronic redistribution and ligand competition effects (fig. S12).

Fig. 2. Interaction and fluorescence quenching analysis in the PIL.

(A) Schematic illustration of supramolecular interactions between ILs and metal ions. (B) XPS spectra of N 1 s of the PIL (metal salt concentration: 100 mg mL⁻¹) and the pure IL. (C and D) XPS spectra of Fe 2p (C) and Cl 2p (D) of FeCl₃ and the IL-Fe³⁺. (E) The energy-distance scan plot for [N(CN)₂]⁻ and metal ions; below shows the isosurface map of the interaction region indicator (IRI) of three systems at the state of minimum energy. Sign(λ₂)ρ is mapped on the isosurfaces according to the coloring method below the figure. The isovalue of IRI is chosen to be 1.0. (F) PL emission spectra of the PIL (metal salt concentration: 5 mg ml⁻¹) and the pure IL. (G) Optical image of the PIL and the pure IL under UV light (365 nm) irradiation. (H) Time-resolved PL (TRPL) profiles of IL-Co²⁺ and the pure IL. (I) Schematic illustration of electron transfer from the PIL to metal ions induced nonradiative loss for the excited state E₁.

Identical trends are detected in metal element and Cl peaks of IL-Co²⁺ and IL-Cu²⁺ (fig. S13), with detailed binding energy data provided in table S1. The table also reveals that cyano-N binding energy undergoes the largest shift upon metal ion introduction, evidencing stronger intermolecular interactions.

To further validate these interactions, density functional theory (DFT) calculations were conducted. A relaxed scan was performed to identify the optimal configurations by systematically varying the metal ion to cyano-N distances, which revealed minimum energy states at 1.91 Å for Fe³⁺, 2.07 Å for Co²⁺, and 1.83 Å for Cu²⁺ (Fig. 2E). The three figures below display corresponding interaction region indicator (IRI) isosurface plots at optimized configurations, graphically visualizing the interactions (34, 35). Distinct blue isosurfaces between metal ions and cyano-N demonstrate strong supramolecular interactions (electron density ≥ 0.04 a.u.).

According to the above results, it can be concluded that upon the addition of metal salts, coordination interactions between the anions ([N(CN)₂]⁻) and metal ions are established, with cyano-N playing a predominant role in these interactions. The strength of these interactions is greater than that of the coordination interactions between Cl⁻ and metal ions. However, these interactions can still be classified as supramolecular in nature (36), as they are markedly weakened upon heating.

Such supramolecular interactions not only enhance thermal sensitivity and adjust the absorption peaks in the visible region but also substantially improve photothermal performance. Figure 2F displays the photoluminescence (PL) emission spectra of PIL, revealing that the introduction of metal ions markedly reduces the fluorescence of the IL. Specifically, Fe³⁺ and Cu²⁺ completely quench the fluorescence (Fig. 2G). Further spectral investigations demonstrate that the PL lifetime decays from 3.67 to 2.39 ns upon the introduction of Co²⁺ (Fig. 2H). This suggests that the quenching mechanism is not an absolute static quenching process but rather involves dynamic coordination interactions (37). Previous reports have demonstrated that the fluorescence intensity of imidazolium-based ILs can be quenched by coordination interactions between ILs and metal ions (38, 39). Specifically, Fe³⁺, Co²⁺, and Cu²⁺ are paramagnetic ions with unfilled d-shells (40). Therefore, they may quench the fluorescence of ILs through ligand-to-metal charge transfer (i.e., nonradiative process K_nr; Fig. 2I), ultimately enhancing the photothermal performance of PIL. To confirm that this photophysical modulation is primarily governed by the specific metal cations rather than their counter anions, control experiments using CuBr₂ and CuSO₄ were conducted. These systems exhibited essentially identical visible absorption and fluorescence quenching behaviors to those of the chloride-based system (fig. S14). Furthermore, all these systems displayed comparable temperature-dependent electrical responses.

To further elucidate the fate of the absorbed optical energy, thermal decoupling experiments were conducted by immersing the PIL device in stirred water to rapidly dissipate generated heat. Under these conditions, the photoresponse was almost completely suppressed (fig. S15), indicating that photothermal conversion dominates the conductivity modulation process. Furthermore, Fourier transform infrared (FTIR) spectra collected before and after illumination showed no detectable structural changes (fig. S16), excluding photochemical transformation as a major pathway.

Fabrication and characterization of the LICs

LICs overcome the inherent limitations of solid materials in structural designs for curved surface sensing, particularly their insufficient stretchability and mechanical robustness, although developing fully liquid systems remains challenging. In our system, the liquid nature of PIL inherently provides flexibility, while the use of room temperature liquid metal eutectic gallium-indium alloy (EGaIn) ensures high conductivity and stretchability of electrodes. To stably support these liquid components, spandex fabric is selected as the structural substrate due to its minimal mechanical hysteresis and superior elasticity (fig. S17), ensuring reliable cyclic curvature modulation.

As shown in Fig. 3A, the LICs are fabricated through three simple steps: (i) coating EGaIn on the upper and lower surfaces of spandex fabric, (ii) modification of spandex via patterned surface plasma treatment, and (iii) adsorbing PIL into the spandex. The polyethylene terephthalate (PET) shadow mask in fig. S18 enables the fabrication of liquid metal electrodes on both surfaces of the spandex through mechanical scraping of EGaIn, where applied force overcomes its cohesion to adhere the metal onto the fabric, forming conductive pathways (fig. S19). The freshly prepared EGaIn circuit exhibits a shiny and metal-like surface, which gradually becomes dull after prestretching due to the formation of a thin oxide layer. Under over 100 stretching cycles at 30% strain (∼1.62% s⁻¹), EGaIn demonstrates high conductivity and stability with only ∼2% resistance increase (0.06 ohm), which is negligible compared to the PIL (Fig. 3B). This exceptional performance arises from EGaIn dynamically deforming with the spandex substrate during stretching. The mechanical strain triggers localized rupture of the native oxide layer, exposing fresh EGaIn to ambient oxygen and promoting self-limiting reoxidation. This process synergistically achieves two critical objectives: strain adaptability through fluidic reorganization and electrical continuity via the oxide-mediated percolation network (41, 42). Specifically, when the spandex returns to 0% strain, mechanical competition between the harder oxide layer and softer spandex induces EGaIn wrinkling, with the folded structure ensuring stable electrical performance under repeated deformation (fig. S20).

Fig. 3. Fabrication and characterization of LICs.

(A) Schematic diagram of the LICs fabrication process. (B) Resistance change of the EGaIn electrodes on spandex fabric as a function of cycling time during stretching at 30% strain (R refers to the resistance after stretching, and R₀ refers to the resistance before stretching, exceeding 100 cycles.); insets show optical images of the electrodes before and after cyclic stretching. (C and D) Scanning electron microscopy images of the spandex before (C) and after (D) air plasma treatment; insets show contact angle images of IL droplets on the respective surfaces. (E) Optical image of a miniaturized photosensitive array. (F) Microscope images of the photosensitive unit. (G) Optical images of the photosensitive array in different mechanical states of stretching and recovering. (H) Schematic diagram of LICs. (I) Optical image of the LICs. (J) Optical image of the LICs on a human eye model to demonstrate its flexibility.

Subsequently, polymethyl methacrylate (PMMA) masks are used for patterned plasma treatment. As shown in Fig. 3 (C and D), the modified spandex exhibits a rougher surface, directly reducing the contact angle from 131.3° to 0°, substantially enhancing hydrophilicity to enable photosensitive array fabrication. To verify circuit continuity after PIL adsorption, a miniaturized array is developed to observe PIL distribution within the spandex (Fig. 3E and fig. S21). As shown in Fig. 3F, despite numerous voids in the spandex, PIL fully infiltrates the fabric via capillary forces, forming complete sensing units along the interwoven warp and weft yarns. The pixel density of the array is primarily determined by the patterning precision of the liquid metal electrodes and the size of the photosensitive units. By using laser-cut acrylic masks for patterned plasma treatment, a minimum pixel diameter of ∼2.2 mm was achieved in this work (fig. S22), while the engraving process exhibited an acceptable structural variation of 6.8% (fig. S23). Further miniaturization toward practically relevant high-density arrays can be realized by adopting advanced microfabrication techniques, such as inkjet printing coupled with in situ UV-cured polymerizable ILs (poly-PIL) to physically arrest fluidic spreading (fig. S24). Furthermore, the intrinsic flexibility of the spandex maintains array integrity under various mechanical states (Fig. 3G). Integrated with PIL adsorption (Fig. 3H and fig. S25), the LICs are successfully constructed, which can bend and stretch with the flexible substrate to adapt to dynamic curved surfaces (Fig. 3, I and J).

DFOS system based on PIL

The DFOS system consists of a spandex layer adsorbed with PIL and EGaIn, sandwiched between two polydimethylsiloxane (PDMS) layers and encapsulated within a hemispherical resin mold (Fig. 4A). Specifically, silicone adhesive was applied to seal the double-layer PDMS structure, eliminating air leakage paths through the porous fabric to ensure effective and reversible curvature modulation (fig. S26). Crucially, these core structural components demonstrate outstanding mechanical robustness. Under rigorous 1000-cycle loading-unloading tests, both the spandex fabric and the PDMS substrate exhibited stable mechanical profiles and retained their original macroscopic morphologies without deformation (fig. S27). To interface the soft liquid metal electrodes with rigid external electronics, a custom-designed printed circuit board featuring a 15-pin D-sub connector was used, ensuring robust electrical contact and high-fidelity signal transmission to a multifunctional switch data acquisition system.

Fig. 4. Demonstration of a DFOS system for color recognition, pattern identification, and distance recognition.

(A) Exploded view illustration of the components of the DFOS system. (B) Tilted view of this system in concave, planar, and convex configurations, actuated by pressure applied to a fluid-filled chamber underneath. (C) Sensing test protocol of a DFOS sensor. (D) Schematic diagram illustrating the procedure of electrical signal recognition at different colors, patterns, and distances, including signal input, model training, and prediction. (E to G) Loss variations during training epochs for color recognition (E), pattern identification (F), and distance recognition (G). (H to J) Confusion matrix of classification results at different colors (H), patterns (I), and distances (J).

A programmable stepper motor–driven syringe is used to precisely regulate internal fluid pressure, translating hydraulic drive into on-demand curvature modulation (fig. S28). Here, three typical modes: concave, planar, and convex surfaces have been used in the following demonstration (Fig. 4B). Benefiting from the synergistic elasticity of PDMS-spandex composites and the flexibility of the liquid metal circuit, the application of positive hydraulic pressure induces substantial deformation of the photosensitive array to form a convex curvature. Conversely, upon pressure removal, the resultant negative pressure couples the elastic rebound of spandex fibers with PDMS contraction, driving the sensing interface to transition from planar to concave configurations (movie S1). The vertex height and radius of curvature exhibit repeatable and systematic changes controlled by hydraulic adjustment. As continuously monitored during operation (fig. S29), this hydraulic actuation operates with outstanding reproducibility and no noticeable hysteresis. Typically, the application of positive pressure constructs a convex sensing interface with a strain rate of ∼1.62% s⁻¹ (fig. S30), achieving a curvature modulation angle of 31.9° (fig. S31). Benefiting from the >24.3% strain capacity, this modulation amplitude surpasses that of state-of-the-art focus-tunable imagers (table S2).

This sensing architecture substantially simplifies optical modules while simultaneously reducing optical aberrations and capturing multiscale optical information. Achieving such large-amplitude dynamic deformation physically precludes the use of conventional rigid color filters, which inevitably suffer from severe structural delamination during actuation (fig. S32A). Moreover, quantitative comparisons confirm that relying on external filters induces marked optical attenuation (fig. S32, B to J). Consequently, our intrinsic filter-free architecture uniquely preserves both mechanical reliability and high signal integrity.

To explore basic operations, we first examine the behavior using a fixed sensing system (Fig. 4C). For the same cross image, representative images collected by sensors in concave, planar, and convex configurations are shown in fig. S33. During the continuous transition across these states, in situ tracking confirms that the liquid sensing units robustly maintain photoresponsive behavior under progressive strain, accurately capturing dynamic focal plane shifts (fig. S34). Notably, the DFOS approach enables acquisition of different optical path information while fixing the sensor by adjusting the sensing interface configuration, achieving simultaneous multifocal plane image capture. This dynamic focal plane scanning overcomes the critical limitation of conventional static planar sensors, which are fundamentally incapable of distinguishing objects at different distances that form identically sized two-dimensional (2D) projections (fig. S35).

Building on this optical architecture, multimodal optical sensing was conducted to further demonstrate the potential of dynamic curvature modulation by evaluating three parameters: patterns, colors, and distances. Electrical responses of 32 sensing units under illumination were scanned and processed through a CNN model (Fig. 4D) to computationally decouple the inherent electrical cross-talk (sneak-path currents) within the passive array (fig. S36). The macroscopic system exhibits a response time of ∼44 s and a recovery time of ∼105 s (fig. S37), a kinetic behavior governed by the high thermal mass of the macroscale substrate. However, this is geometrically scalable; a miniaturized pixel achieved a reduced response time of 0.5 s and a recovery time of 0.6 s (fig. S38). For color recognition, signals acquired under three curvatures served as machine learning inputs. The training process converged after 60 epochs, achieving exceptionally high classification accuracy, with similar variation trends observed in both training and testing, demonstrating effective model optimization (Fig. 4E). The confusion matrix in Fig. 4H confirmed high classification precision for each color. Similarly, both pattern and distance recognition accuracies approached near-perfect levels (Fig. 4, F and G), with their confusion matrices provided in Fig. 4 (I and J), respectively. The model demonstrated high prediction accuracy across diverse conditions in receiver operating characteristic (ROC) curve analysis [area under the curves (AUCs) = 1; fig. S39]. Fivefold cross-validation (fig. S40) and independent random seed tests (fig. S41) further confirmed the robustness of these results. An expanded distance recognition task with nine discrete distances also achieved 100% classification accuracy (fig. S42), supporting the reliability of the model under controlled experimental conditions and the future potential of this proof-of-concept platform.

Beyond recognition accuracy, the fully liquid system exhibited exceptional mechanical resilience and stability. Under continuous mechanical vibration, the intrinsic fluidity of the LICs effectively decoupled the electrical signals from mechanical noise, ensuring a drift-free baseline (fig. S43). Crucially, during this dynamic operation, the sensing unit maintained an ultralow energy consumption of ∼2.50 μW. The system also demonstrated excellent fatigue resistance; the photoresponse intensity remained invariant throughout 600 cycles of reversible curvature modulation (convex to planar) (fig. S44). Even under severe mechanical damage simulated by 300 repeated hammer impacts, the photoresponse remained stable without structural failure (fig. S45). Furthermore, excellent device-to-device reproducibility was verified using five independently fabricated DFOS devices, which exhibited consistent performance under identical illumination conditions (fig. S46).

To evaluate the system’s operational viability under challenging real-world scenarios, comprehensive environmental assessments were conducted. In outdoor field tests, despite exposure to uncontrolled variables such as wind and temperature fluctuations, the device successfully captured focused sunlight signals with high fidelity (fig. S47, A to E). The sensor array also reliably processed spatially distributed, spectrally mixed natural illumination without relying on any external focusing elements (fig. S47, F to H). Because the sensing mechanism relies on photothermal conversion, fluctuations in ambient temperature can inevitably couple with these optical signals. To resolve this, we implemented a quadratic temperature compensation algorithm that dynamically normalizes the nonlinear enhancement of ion mobility at elevated temperatures (fig. S48). This algorithmic decoupling successfully isolates the pure photoresponse from ambient thermal disturbances, ensuring consistent optical recognition across variable temperature environments. Ultimately, the DFOS system effectively acquires multidimensional visual information through curvature-modulated light paths. While the current demonstration focuses on discrete distance classification, our supplementary distance-response calibration clearly establishes a reproducible, continuous monotonic relationship (fig. S49). This confirms the fundamental feasibility of evolving the present architecture toward quantitative, continuous 3D spatial sensing.

DISCUSSION

In summary, we have developed a geometry-adaptive DFOS system that integrates PIL with stretchable liquid metal electrodes on spandex fabric, enabling reversible curvature modulation for adaptive optical sensing. The PIL, based on [BMIm][N(CN)₂] coordinated with Fe³⁺, Co²⁺, and Cu²⁺, exhibits selective color perception and stable photo-thermo-electric conversion, attributed to dynamic supramolecular interactions between transition metal ions and [N(CN)₂]⁻. XPS and DFT analyses confirm the nature of these interactions, while fluorescence quenching provides evidence for efficient photothermal conversion. The system also demonstrates excellent durability, as verified by repeated cyclic deformation and thermal stability tests.

Combined with stretchable LICs, the system enables multimodal (concave, planar, and convex configurations) sensing encompassing color, pattern, and distance recognition and maintains high classification accuracy. The fully liquid architecture establishes a fundamentally different optoelectronic paradigm in which optical functionality is dynamically reconfigured through geometric deformation of the sensing system itself. This lens-free, sensor-centric strategy effectively mitigates field-curvature mismatch and enables adaptive focal plane modulation as well as simultaneous signal acquisition across multiple depths, providing capabilities that are difficult to achieve using conventional rigid imaging platforms. Although the current prototype does not yet match the pixel density of commercial silicon sensors, its unique combination of mechanical adaptability, filter-free multimodal sensing, and compatibility with soft materials highlights the tremendous potential of liquid optoelectronic systems. Future developments, including higher-density integration and polymer-stabilized PIL sensing units, may further enhance the scalability, robustness, and long-term operational stability of this architecture.

MATERIALS AND METHODS

Chemicals and reagents

[BMIm][N(CN)₂] was purchased from Lanzhou Greenchem ILS, LIPC, CAS (Lanzhou, China). Calcium chloride (CaCl₂) and palladium chloride (PdCl₂) were obtained from Shanghai Macklin Biochemical Co. Ltd. Cobalt chloride (CoCl₂), potassium chloride (KCl), and manganese chloride (MnCl₂) were provided by Shanghai Aladdin Biochemical Technology Co. Ltd. Iron(III) chloride (FeCl₃), copper(II) chloride (CuCl₂), and iron(II) chloride (FeCl₂) were provided by chloride Meryer (Shanghai) Biochemical Technology Co. Ltd. Copper(I) chloride (CuCl) was purchased from Beijing Yili Fine Chemicals Co. Ltd. Aluminum chloride (AlCl₃) was purchased from Sinopharm Chemical Reagent Co. Ltd. Copper(II) bromide (CuBr₂), copper(II) sulfate (CuSO₄), acrylamide, and photo-initiator 1173 were purchased from Anhui Zesheng Technology Co. Ltd. Liquid metal EGaIn was obtained from Dongguan Wochang Metal Products Co. Ltd., and the melting point was 12°C. The PDMS (Sylgard 184) was provided by Corning Co. Ltd. All chemicals were used as received without further purification. PET film, PMMA, and electrospun styrene-ethylene-butylene-styrene fibers were commercially available.

Instrument and characterization

The UV-Vis spectra of coordination IL solutions were measured on a SHIMADZU UV-3600 spectrophotometer. The resistance responses and current changes were collected by using a CHI760E electrochemical workstation (Shanghai CH Instrument Co. Ltd.). The TGA was performed on a TA Q50 instrument (TA Instruments Inc.). The temperature was incrementally raised from room temperature to 600°C at a rate of 10°C min⁻¹ under a nitrogen atmosphere. The color light was provided by a CB-2065 digital projector [Epson (China) Co. Ltd.]. The XPS measurements were performed on an Axis Ultra imaging photo-electron spectrometer (Kratos Analytical Ltd.) using a monochromatized Al Kα radiation and the C 1s peak at 284.8 eV as the internal standard. The steady-state PL emission spectra, and time-resolved PL (TRPL) were taken using an Edinburgh FLS 980 spectrometer. Specifically, for the TRPL measurement, the excitation light pulse was provided using a picosecond diode laser at a 405-nm wavelength. The instrument used to stretch the spandex fabric was Instron 2367 apparatus with a strain rate of ∼1.62% s⁻¹. The FTIR spectra were obtained through a CaF₂ optical window, with a Bruker TENSOR 27. One thousand–cycle loading-unloading fatigue tests were conducted on the spandex fabric and PDMS substrates at a maximum strain of 30% using the aforementioned Instron 2367 apparatus. The air plasma treatment was performed using a plasma cleaner (PDC002-HP, Harrick Plasma) for 10 min. The microstructures of the spandex before and after air plasma treatment were observed using a scanning electron microscope (Hitachi SU8010, Japan). The contact angle was measured by a contact angle system (DSA30, KRÜSS), which calculated an average value from three data at different positions by using an IL droplet. An inverted fluorescent microscope (DMi8, Leica) was used to capture the bright-field images of the photosensitive unit. The PET shadow masks and acrylic stencils were processed using a commercial laser cutting system (Tianlang Laser Technology Co. Ltd., China). The internal hydraulic pressure was monitored during the cyclic curvature modulation using a digital pressure gauge (MIK-Y290, Asmik, China). Multichannel measurements of the DFOS system were conducted via a CHI 760E paired with a CHI 684 multiplexer and a multifunctional switch data acquisition (Keysight, 34980A). During the continuous mechanical vibration test, the current was continuously recorded using the amperometric i-t curve technique at an applied potential of 0.5 V using a CHI660F electrochemical workstation. The high-density sensing array was printed with an inkjet printer (MicroFab Inkjet, Ruidu, Shanghai, China).

Fabrication and testing of photosensitive chips

Hollow cylindrical cavities were fabricated in PDMS (base resin and catalyst in 10:1 ratio by mass) via mechanical punching. Glass slides and ring-shaped PDMS were bonded via air plasma treatment (3 min), followed by PIL injection and gold electrode integration to realize the photosensitive chips. A projector served as the light source, with color and intensity controlled by RGB input values (N_r, N_g, and N_b) ranging from 0 to 255, corresponding to red, green, and blue light intensities respectively.

Temperature-dependent electrical responses: The response values ($∆G/G_0$, in percentage) were calculated using the following equation

$$\begin{array}{c}∆G/G_0(\%)=\left(\frac{\mathrm{\sigma }-{\mathrm{\sigma }}_0}{{\mathrm{\sigma }}_0}\right)\times 100\%\hfill \\ =\left(\frac{R_0-R}{R}\right)\times 100\%=\left(\frac{I-I_0}{I_0}\right)\times 100\%\hfill \end{array}$$ (1)

where ${\mathrm{\sigma }}_0$, $I_0$, and $R_0$ are the initial ionic conductivity, current, and resistance, respectively. $\mathrm{\sigma }$, $I$, and $R$ are the real-time ionic conductivity, current, and resistance under external thermal or light stimulation (43). By using this normalized fractional metric, the device performance is effectively characterized independent of the applied external bias.

Composition and testing of the DFOS system

The DFOS sensing array consists of three PILs, which were formulated by individually dissolving FeCl₃, CoCl₂, and CuCl₂ into [BMIm][N(CN)₂] at a concentration of 100 mg ml⁻¹. During the evaluation, the real-time resistance or current signals of the PIL units were continuously acquired via the multichannel measurement, and the corresponding photoresponses were calculated using Eq. 1.

DFT calculations

The DFT calculations were carried out with the Gaussian 09W software package to optimize the structure of molecules. Geometrical optimization adopted the pbe1pbe method with 6-31G(d) basis sets (for C, N, and H) and SDD basis sets (for Fe³⁺, Co²⁺, and Cu²⁺). The electronic density analysis and IRI analysis were performed using Multiwfn version 3.8 (dev) and VMD 1.9.3 software.

IRI analysis

IRI is a graphical method for analyzing chemical bonds and noncovalent interactions in chemical systems. IRI is simply defined as follows: $\text{IRI} \left(r\right)=\mid \nabla \mathrm{\rho }\left(r\right)\mid /{\left[\mathrm{\rho }\left(r\right)\right]}^{1.1}$, where $\mathrm{\rho }$ is the electron density and $\mid \nabla \mathrm{\rho }\mid$ denotes its gradient norm. Essentially, IRI is the gradient norm of electron density weighted by scaled electron density. The sign(λ₂)ρ function is projected onto IRI isosurfaces through color mapping to differentiate interaction characteristics. According to the chromatic scheme in Fig. 2E, predominantly red isosurface coloration indicates steric repulsion, while distinctly blue regions signify strong attractive interactions (electron density ≥ 0.04 a.u.). Detailed definitions of IRI can be found in the original methodology paper (34).

Optical modeling

The ray optic simulations were performed with the open-source program Ray Optics Simulation (https://phydemo.app/ray-optics/simulator/).

Statistical analysis and reproducibility

Independent measurement repeats are explicitly specified in the respective figure captions and text. For material-level characterizations (e.g., thermal and photoresponses in Fig. 1), data were obtained from n = 5 independent measurements. Regarding the comprehensive system-level evaluations of the 32-pixel macroscopic array, specifically including the mechanical impact resistance test, the 600-cycle curvature modulation stability test, the two outdoor field demonstrations, the continuous mechanical vibration test, and the color/distance/pattern recognition tasks in Fig. 4, one fully integrated DFOS device was independently fabricated and used for each specific testing scenario to evaluate extreme single-device endurance. It is important to emphasize that while a single device was used for the recognition tasks in Fig. 4, the training and testing datasets were rigorously acquired through five independent measurement repeats under each of the three curvature configurations. For the three-class tasks (color, pattern, and distance) in CNN, all classification tasks were implemented in PyTorch. The network architecture consisted of two convolutional blocks: the first block comprised a 2D convolution (1 input channel, 3 output channels, kernel size 2, stride 1, and padding 1) followed by a scaled exponential linear unit (SELU) activation and max pooling (kernel size 4 and stride 4); the second block applied an analogous convolution (3 input channels, 1 output channel, kernel size 2, stride 1, and padding 1), SELU activation, and max pooling (kernel size 4 and stride 4). The feature maps were flattened and fed into two fully connected layers (with SELU activation). Network weights were initialized using Kaiming normal initialization, and biases were set to 0.01. The model was trained using stochastic gradient descent. The loss function was cross-entropy loss. Each dataset comprised a batch size of 15, with input tensors of shape (15, 400, and 96), corresponding to 576,000 data points per task. The train/test split was performed using torch.utils.data.random_split, with 80% allocated to training and 20% to testing. Reproducibility was ensured by fixing the random seed via torch.manual_seed(seed = 1) before splitting. Model performance was evaluated by classification accuracy, confusion matrices, and ROC curves with AUC metrics.

Supplementary Materials

The PDF file includes:

Supplementary Text

Figs. S1 to S49

Tables S1 and S2

Legend for movie S1

Other Supplementary Material for this manuscript includes the following:

Movie S1