Microelectrothermoforming (μETF): one-step versatile 3D shaping of flexible microelectronics for enhanced neural interfaces

Introduction

High proximity of electrodes to neural cells plays a critical role in achieving efficient neural interfaces for various recording and stimulation applications^1,2,3,4. Improved signal quality in neural recordings can result from intimate contact between the electrode surface and targeted tissues^1,3,5,6, whereas in neural stimulation, lower stimulation thresholds and higher spatial resolution can be achieved through reduced current spreading^2,4,7,8. Thus far, improved electrode–cell proximity has typically been achieved using microelectromechanical system (MEMS) technologies, either by directly structuring rigid materials into needle-like shapes, such as Utah arrays, or by building various three-dimensional (3D) microstructures on top of thin film-based microelectrode arrays (MEAs), such as pillars^{1,4,5,9,10,11}, mushrooms^2,7,12, hemispheres¹³, sheaths¹⁴, and arrowheads¹⁵. However, despite their effectiveness, rigid 3D electrodes may exhibit mechanical mismatch with soft neural tissues. Furthermore, the creation of 3D microstructures on top of thin-film MEAs using conventional MEMS technologies requires additional fabrication steps, such as photolithography, vacuum deposition, electroplating, and wet and dry etching^16,17,18. These steps increase the complexity of the fabrication process, the degree of which usually scales with the complexity of the 3D structures. Moreover, it is challenging to fabricate 3D features with different shapes and heights within a single MEA, owing to the inherent 2D nature of the traditional microfabrication process.

To address these problems, we propose a novel “microelectrothermoforming (μETF)” process for producing versatile 3D microstructures on a polymer-based MEA through a simple one-step thermal pressing of conventional planar arrays. This method leverages a well-established thermoforming technique used in the plastic industry, in which a thermoplastic sheet is heated above the glass transition temperature (T_g) and pressed against a metal mold to create 3D structures in a simple manner¹⁹. The microelectronic adaptation of thermoforming in this study enables the creation of microscopically protruding and recessed versatile 3D structures with embedded electrical functionalities, facilitating the development of tailored structures for optimized electrode–cell interfaces. Thermoforming and microthermoforming for biomedical applications have been explored by Truckenmüller et al.¹⁹ and in subsequent studies, in which heated thermoplastic sheets were pneumatically or mechanically pressed against perforated metal molds to generate 3D shapes, such as microwell platforms^20,21, cell culture chips^22,23, and microchannels^24,25. However, the methods demonstrated in those studies primarily formed bulk polymer sheets that lacked active components such as electrical contacts and interconnections. Additionally, they required specialized tooling to exert pneumatic pressure, and their formable 3D structures were commonly limited to hemispherical shapes.

The proposed μETF process effectively addresses the challenges associated with fabrication complexity and structural restrictions in known 3D microstructures for enhanced neural interfaces. Microelectrothermoforming involves the one-step thermal pressing of a conventionally prepared thin-film planar MEA against a mold carrying the desired 3D structures. Therefore, two distinct advantages can be achieved compared to existing MEMS-based 3D forming: (1) process simplicity and (2) shape versatility. The proposed approach simplifies 3D fabrication by adding only one pressing step to the fabrication of a conventional planar thin-film MEA. Moreover, the use of 3D printing technology for preparing molds enables the formation of virtually any 3D microstructures, allowing heterogeneous structures of diverse shapes and heights to be created within a single electrode array via the same one-step μETF process. This simplicity and flexibility in forming 3D microstructures can expand design possibilities, facilitating the development of optimized electrode–neuron interfaces that align with the anatomical and neurophysiological features of a targeted nervous system.

Liquid crystal polymer (LCP) film was used as a thermoplastic substrate in this study, primarily owing to its mechanical strength, chemical inertness, and biocompatibility. Additionally, the low water absorption rate of LCP can contribute to the long-term reliability of chronically implanted devices²⁶. However, μETF can be utilized for other thermoplastics commonly used in biomedical applications.

This study introduces a one-step μETF process for generating versatile protruding and recessed 3D structures on LCP-based MEAs (LCP MEA) with mechanical considerations for preserving the electrical properties of MEA. As a proof-of-concept, the 3D MEA was optimized for retinal stimulation, the benefits of which were assessed via both computational analysis and ex vivo experiments in a mouse model. A variety of 3D structures, such as wells, domes, walls, and triangles, were constructed on LCP MEAs, demonstrating the potential utility of the μETF in a wide range of biomedical applications.

Results

μETF for simple and versatile 3D structuring

Representative benefits of 3D MEA are illustrated in Fig. 1a. The protruding structures enable closer proximity to target cells, establishing more localized neural interfaces for both neural recording and stimulation^27,28,29,30. This is in contrast to conventional planar MEAs, in which the electrode surfaces inherently lie below the top surface^31,32,33. As outlined in Fig. 1b, the proposed one-step μETF process locally deforms a planar MEA into a 3D MEA with microscopic protruding and/or recessed structures. A planar 25-channel LCP MEA prepared via a conventional microfabrication process³² (see Supplementary Fig. 1) is thermally pressed (>T_g) against a 3D-printed mold within a set of metal jigs and elastomer layers for alignment (see Supplementary Figs. 2, 3 for details), replicating the protruding and/or recessed 3D microstructures of the mold to the MEA, as shown in Fig. 1c. The electrode opening was performed after μETF step to prevent the opening area from being affected by thermoforming step, ensuring the consistent electrochemical properties of MEA.

Microelectrothermoforming (μETF): one-step versatile 3D shaping of flexible microelectronics for enhanced neural interfaces — **Fig. 1: One-step μETF of LCP MEA to create microscopic protruding and recessed 3D structures for enhanced neural interfaces.**

Additionally, the microelectrothermoformed 3D MEA can be grossly deformed into nonplanar shapes to conform to surrounding tissues, such as eye curvature for retinal electrodes, through a similar “macro” electrothermoforming step (mETF, see Supplementary Fig. 4). Using a 3D mold with 80-μm-height pillars (Supplementary Fig. 5), the schematics (top row) and photographs (bottom row) in Fig. 1d show the evolution of a (i) 40-μm-thick planar LCP MEA to a (ii) protruding or (iii) recessed μETF MEA, and subsequently to a (iii) μETF + mETF MEA, achieving both high proximity to target cells and conformability to target tissues. The recessed structures in (iii) can be easily generated using the identical processes and tools to those for the protrusions, simply by flipping the planar array upside down in the fixture, offering a convenient approach to create well-like structures for highly localized electrode–cell environments²⁹. Scanning electron microscopy (SEM) images of individual electrode sites before and after μETF are presented in Fig. 1e, with their cross-sectional profiles shown in Fig. 1f. Each channel site with a diameter of 200 μm was selectively elevated or lowered by a height of 80 μm, which represents an optimized height for subretinal electrode arrays as a proof-of-concept application of μETF (details in subsequent sections). The optical surface profiles of the protruding (top) and recessed (bottom) channel sites are shown in Fig. 1g. The μETF produced 3D structures that replicated the original mold structures, exhibiting a slightly widened base diameter (~150%) and sloped sidewalls (~70°). The consistency of protrusion across different electrode locations is demonstrated in Supplementary Fig. 6. The electrode diameter as small as 100 μm was also successfully thermoformed using the same μETF configuration (Supplementary Fig. 7), which was utilized for following ex vivo experiments.

Similarly, versatile protruding and recessed structures of diverse heights and shapes can be created without adding fabrication complexities. Figure 2a–f present 25-channel LCP MEAs formed with varying protrusion/recessed heights from 80 to 200 μm using a 3D mold (Supplementary Fig. 8). Their SEM images (Fig. 2a, b), cross-sectional images (Fig. 2c, d), and optical 3D profiles (Fig. 2e, f) confirm the faithful replication of the 3D mold structures onto the LCP MEAs. Furthermore, the versatility of the microthermoforming process was demonstrated with various microstructures, including polygons (triangles, rectangles, and hexagrams), ovals (sunken, plateau, and walled), domes, and S-shaped walls, as shown in Fig. 2g, h; their SEM images and optical profiles are shown in Fig. 2I, j, respectively. All distinct structures were created via one-step μETF using corresponding 3D molds shown in Supplementary Fig. 8. Such versatility and consequent design flexibility can be leveraged to create optimized 3D structures tailored for various in vivo and ex vivo neural interfacing applications.

**Fig. 2: Versatile 3D structures with diverse shapes and heights created by the one-step μETF process.**

Electromechanical considerations for μETF

Despite the simplicity and versatility of the μETF process, it is important to consider the tensile strain exerted on the thin gold layers, which may lead to cracks or disconnections. Therefore, the μETF process was optimized with respect to the 3D mold design and pattern geometries.

Figure 3a presents a finite element analysis (FEA) of the mechanical strain induced within the embedded thin gold layer during an 80-μm-height μETF. Based on the strain distribution for the top, bottom, and neutral planes within a gold layer (Fig. 3b), the highest tensile strain (~12%) is expected at the top and bottom corners of the sidewall (insets in Fig. 3a). On the other hand, the top protruding area remained relatively stress-free because of the pillar-shaped mold with a flat top surface. The resulting plateau-like structure of the μETF electrode ensured minimal damage to the circular electrode area, which were designed to be 10 μm smaller than the mold top diameter, as outlined on the transversal strain distribution in Fig. 3c. This is in contrast to forming a 3D structure using a dome-shaped mold, in which electrode sites are subjected to highest tensile strain, leading to significant cracks on the gold electrode surface after μETF (Supplementary Fig. 9).

**Fig. 3: Electromechanical considerations of MEA for 80-μm-height μETF.**

The interconnection lines traversing the region under the greatest strain can be configured with serpentine patterns for enhanced tolerance against elongation (Fig. 3c). However, the μETF poses unique challenges that interconnections undergo permanent deformation on softened polymer substrates at >T_g, a condition distinct from the reversible, solid-state serpentine interconnections typically used in stretchable electronics. Also, unlike conventional serpentines that exploit global out-of-plane buckling^34,35,36, the serpentines in μETF is constrained “in-plane” as it is tightly pressed between jig and mold during thermoforming. Additionally, the spatially localized deformation restricts the serpentine layout in increasing radius r, arc angle θ, and number of arc units m, which are known to enhance stretchability^36,37,38,39.

Given these conditions, sub-micrometer thin-film gold layers are not considered viable, due to their fragility upon substrate-softening and susceptibility to global crack-formation upon stretching⁴⁰. Instead, the gold patterns were mechanically strengthened by electroplating up to a thickness t = 4 μm, enabling freestanding, ribbon-like behavior during deformation within softened substrates.

Next, the serpentine parameters (w and θ) were optimized through FEA under an in-plane constraint at ε_applied = 15% (>ε_max in Fig. 3c), with the maximum radius and length available in the designed 80-μm-height, 25-channel MEA (r = 20 μm and m = 2). As shown in Fig. 3d, e, serpentines with smaller w and larger θ exhibited reduced maximum strain (ε_max), consistent with literature for non-buckling serpentines^38,41. Also, thicker lines exhibited smaller ε_max (Supplementary Fig. 10). Based on these FEA results, the optimum geometries were determined to be w = 10 μm and θ = 180˚, as it provided significant strain relief compared to straight lines, both in terms of ε_max and overall strain distribution. A 5 μm width and 240˚ angle were considered suboptimal, as they could compromise microfabrication yield and require a large inter-channel spacing, respectively.

Although the expected ε_max in the optimized in-plane serpentines exceeded the plastic deformation strain of gold (~0.3%), this is considered relatively acceptable for permanently deformed lines in μETF as long as electrical conductivity is preserved, unlike typical serpentines in stretchable electronics that strictly require reversible mechanical properties. To confirm this, the integrity of the FEA-based interconnection design was experimentally validated using patterns with varying shapes (straight and serpentine) and thicknesses (t = 300 nm to 4 μm) in 80-μm-height μETF. The representative SEM images in Fig. 3f demonstrate that damages to the gold layers were caused by two distinct mechanisms, as quantified by yield analysis after μETF in Fig. 3g. The “electrode cracks” occurred mostly in the thin gold patterns (t = 300 nm and 2 μm) with relatively lower mechanical strength, resulting in a distributed crack-formation throughout the gold patterns. On the other hand, circular electrodes with higher gold thickness (4 μm) remained intact, while focused mechanical stress caused a single spot of line disconnection, which corresponds to the location of the highest stress estimated in Fig. 3c, d. The FEA-optimized serpentine with w = 10, t = 4 μm, and θ = 180˚ (III) ensured 100% tolerance (Fig. 3g) after 80-μm-height μETF. Thicker patterns than 4 μm did not provide mechanical enhancement (Supplementary Fig. 11).

These results imply that the in-plane serpentine provides sufficient structural benefit to ensure reliable thermoforming of gold lines without disconnection under an in-plane restriction of μETF. The serpentine parameters may be individually optimized for variable μETF structures depending on the applications. For example, introducing a linear segment between arcs can enhance the tolerance for greater elongation^37,38.

Electrochemical and mechanical analysis of μETF MEA

The intactness of the neural interfaces during μETF was confirmed via electrochemical analyses, including electrochemical impedance spectroscopy (EIS) and cyclic voltammetry (CV) of MEAs before and after μETF. Additionally, the iridium oxide (IrOx) layer was electrodeposited on top of the gold channels before and after deformation to demonstrate the compatibility of the μETF process with nanoporous surface functionalization techniques widely used for enhancing the charge transfer capabilities of neural interfaces^42,43. As shown in Fig. 4a, b, the impedance magnitude and phase at 1 kHz showed no significant changes after μETF for both gold and EIROF electrodes (gold: 32.3 ± 10.8 to 27.3 ± 14.1 kΩ, IrOx: 4.1 ± 0.5 to 3.7 ± 0.4 kΩ; see Supplementary Fig. 12a, b for EIS spectra). Figure 4c presents the cathodic charge storage capacities (CSC_C), which also confirmed that μETF did not result in any significant degradation (gold: 0.24 ± 0.02 to 0.18 ± 0.02 mC/cm², IrOx: 38.8 ± 8.2 to 35.1 ± 11.7 mC/cm²; see Supplementary Fig. 12c for CV curves).

**Fig. 4: Electrochemical and mechanical analysis of MEA for 80-μm-height μETF.**

The mechanical resilience of μETF protruding structures against compression was evaluated to assess the potential deformation of the μETF LCP structures under physiological or surgical conditions. As shown in Fig. 4d, the μETF MEA with a height of 80 μm were subjected to increasing compression using a motorized force meter. The magnified plot in Fig. 4e suggests that the 3D electrodes are compressed by no greater than 2.2 μm under the normal range of pressures experienced in physiological conditions, including (i) intraocular pressure (1.3–2.8 kPa⁴⁴) and (ii) intracranial pressure (0.9–2 kPa⁴⁵), as described in Supplementary Fig. 13. Even after the μETF MEA was completely flattened under a load pressure of 150 kPa, the original height and mechanical properties were recovered after the force is released, as demonstrated by the three cycles (Fig. 4d) and 100 cycles (Supplementary Fig. 14) of full compression and relaxation. Repeated vertical compressions did not affect the electrochemical properties (EIS) of protruding MEA during and after 50 cycles of pressing at 25 kPa, substantially higher than the pressures under physiological conditions, as shown in Fig. 4f and Supplementary Fig. 15. Moreover, the structure of the protruding MEA remained unchanged after prolonged compressions at 25 kPa maintained for 1 h (Supplementary Fig. 16). The μETF MEA also demonstrated mechanical robustness against lateral stress encountered during implantation, through 10 cycles of forced sliding through a 2% agarose tissue phantom, which has substantially higher modulus (~300 kPa)⁴⁶ than common 0.6% brain phantom (~1 kPa)^47,48, as shown in Supplementary Fig. 17.

These results suggest that the proposed μETF with proper designs enables the reliable production of 3D structures without compromising the physical and electrochemical properties of neural interfaces.

Benefits of μETF MEA for subretinal stimulation in FEA study

The benefits of μETF MEA with protruding structures were evaluated for subretinal stimulation of blind patients through FEA in a human retinal model^49,50. A subretinal electrode is implanted under the retina, facing toward the bipolar cells and retinal ganglion cells (RGCs) (see Supplementary Fig. 18 and Table 1 for the retinal model). Given the distance between the electrode surface and targeted bipolar cells in the inner nuclear layer (INL), subretinal stimulation is expected to benefit from protruding 3D structures, which would create more focused current distribution at the target cells. Improvement in neural interfaces enabled by 3D subretinal electrodes was quantified in terms of stimulation threshold (I_th), dynamic range (DR), and spatial resolution, as shown in Fig. 5.

**Fig. 5: FEA study for estimating benefits of μETF MEA for subretinal stimulation.**

The I_th and DR were evaluated based on the E-field distribution at INL generated by current injection from a 3 × 3 subretinal electrode array with varying protruding heights from 0 to 100 μm. The I_th is defined as the current intensity required to induce an electric field (E-field) exceeding 3000 V/m at the plane of the INL to activate bipolar cells^27,51,52. The DR is a current range from I_th to the maximum current (I_max), beyond which inter-channel interference occurs between adjacent electrodes. The representative E-field distributions at INL from planar and 80-μm-protruding MEA are presented with increasing stimulation currents from planar in Fig. 5a, b. Notably, the 80-μm-protruding electrodes lowered the I_th to 4.5 μA from 6 μA of the planar electrodes. Increasing current beyond the I_max began to induce inter-channel interference, at which the INL area activated by a channel (contoured by white dashed lines in Fig. 5b) overlapped with the area activated by adjacent channels. The onset of interferences defines the upper boundary of the stimulation DR, which was extended from 14.5 μA for the planar array to 18.9 μA for the 80-μm-protruding array. A similar analysis for varying protruding heights from 0 to 100 μm (see Supplementary Fig. 19) suggested that the higher protrusions resulted in lower I_th and wider DR, as plotted in Fig. 5c. Such enhancement of 3D electrodes can be explained by more focused current distribution at INL, as shown in Fig. 5d (I = I_th). The MEA with higher protrusion generated higher contrast in the E-field profiles between the targeted area and the untargeted area. However, the 100-μm-protruding array induced an uneven E-field distribution, presumably due to an unmitigated edge effect from an excessively close electrode–cell distance. Therefore, we concluded that an 80 μm protrusion is the optimum height for efficient subretinal stimulation, which was adopted for the proof-of-concept 25-channel 3D array presented in the previous sections.

The protruding MEA is also predicted to enhance the spatial resolution of retinal stimulation, which was quantified using the Michelson contrast (MC), measuring the ratio of (E_max − E_min) to (E_max + E_min) (more details in the Methods), as shown in Fig. 5e. Higher protrusion of the electrodes led to higher E-field contrast (Supplementary Fig. 20) and correspondingly higher MC values across the entire range of channel pitches from 250 to 600 μm.

When the electrode diameters were varied from 100 to 500 μm, the smaller electrodes were expected to exhibit enhanced stimulation performance in terms of I_th, DR, and MC, as shown in Supplementary Fig. 21. The smaller electrode presumably led to higher current density and mitigated edge effects, resulting in reduced I_th and a wider DR, while the more focused activation area contributed to higher MC. In general, all protruding structures offer advantages over planar counterparts by reducing the effective electrode–cell distances, thereby enhancing I_th, DR, and MC.

Benefits of μETF MEA for subretinal stimulation in ex vivo experiments

The effectiveness of μETF MEA for enhanced neural interfaces in subretinal stimulation were evaluated through ex vivo retina experiments, by comparing the protruding μETF MEA and planar MEA in terms of stimulation threshold and spatial resolution. Activation of RGCs was monitored by imaging the calcium transients in response to electrical subretinal stimulation, using a custom-built fluorescence microscopy setup (Fig. 6a). As shown in Fig. 6b, the MEA placed under the mouse retina patch included both planar and 80-μm-μETF protruding electrodes with a diameter of 100 μm (Supplementary Fig. 7). A genetically encoded calcium indicator, sRGECO, was introduced to the retina via adeno-associated viral (AAV) vectors and its expression in RGC layer was confirmed three weeks after the injection (Fig. 6c). A typical calcium transient, in response to biphasic current pulses, is represented by normalized changes in fluorescence, ΔF/F₀ = (F_peak – F₀)/F₀, as shown in Supplementary Fig. 22.

**Fig. 6: Ex vivo retinal experiments for evaluating the effectiveness of μETF MEA for subretinal stimulation.**

The stimulation threshold was evaluated by quantifying ΔF/F₀ of RGCs placed within the diameter of the stimulating electrode (Fig. 6d) while increasing the current injection. As shown in Fig. 6e, the resulting responses were fitted into sigmoidal functions, from which the threshold was defined as the stimulation current at half of the maximum ΔF/F₀. The protruding electrode significantly lowered the median stimulation threshold to 0.91 μA, compared to 1.55 μA of the planar electrode, as shown in Fig. 6f. This suggests that the closer proximity of protruding structure to bipolar cells and RGCs allowed for neural activation with a lower current.

The spatial extent of electrical activation was quantified by ΔF/F₀ of RGCs depending on their distances from the center of the stimulating electrode (Fig. 6d) at a fixed current of I = 20 μA, as shown in Fig. 6g. The response curves were fitted by Gaussian functions to determine the spatial extents, defined as the half-width at half maximum (HWHM) of ΔF/F₀. The protruding electrode produced more focused retinal activation by reducing the HWHM from 94.8 to 43.7 μm, which well agrees with the FEA estimation in Fig. 5d. Stimulation with I = 10 μA resulted in consistent outcome (99.7 to 39.2 μm), as shown in Supplementary Fig. 23. The HWHM of the protruding electrode comparable to its radius (50 μm) indicates that the μETF MEA activates the retina with high contrast and minimized inter-channel interference, potentially providing artificial vision with higher spatial resolution.

Discussion

In this study, we presented a novel μETF process for creating simple and versatile 3D microstructures on thin-film polymer-based neural interfaces using 3D-printed molds to locally deform planar MEAs into desired protruding or recessing shapes. This approach builds upon the widely used polymer thermoforming technology in the plastic industry, but adapts it for microelectronic 3D structuring to create neural interfaces that are intimate with the nervous system. Our proposed technology uses one-step pressing of conventionally prepared planar MEAs to enable virtually any 3D microstructure to be created on microelectrode arrays without requiring additional fabrication steps. This implies that complex structures with different shapes and heights can be created within a single array using the same one-step thermoforming process. This contrasts with previous 3D-shaping techniques for neural interfaces typically, which require complicated fabrication steps and are limited in their ability to create different structural profiles on the same array, primarily because of the nature of the MEMS process.

The 3D printing technology was adopted in this study for preparing master molds, as it enables the creation of complex geometries such as cones, pyramids, and domes with unrestricted sidewall profiles, including vertical, tapered, curved, or spherical shapes. Moreover, the complexity of 3D structures does not increase the fabrication complexity, unlike conventional photolithography-based 3D structures. As a result, different structures and profiles can be created within a single 3D mold, which can subsequently be transferred to a single MEA embedding various 3D structures. These characteristics may highlight the advantages of 3D-printed mold over photolithography-based 3D structuring, which provides high precision but it is relatively more complicated to produce versatile structures. The simplicity and versatility of our proposed method in forming 3D microstructures expand design possibilities, enabling optimized electrode–neuron interfaces that reflect the anatomical and neurophysiological characteristics of various nervous systems. Additionally, by combining gross macrothermoforming and local microthermoforming, it allowed for an even wider range of customized macrostructures and microstructures beyond the 3D structures presented in this study. We expect that the proposed technique will have potential applications in various fields, including wearable electronics, lab-on-a-chip, organoids, and 3D neuronal network^53,54,55,56. Despite the diversity of these applications, they share the benefits of versatile, scalable protruding/recessed 3D structures with integrated microelectronic interfaces for recording and stimulation for optimized performance and functionality. Besides, the μETF is compatible with various thermoplastic materials commonly used in implantable and wearable biomedical applications, including parylene-C¹⁴, cyclic olefin copolymer⁵⁷, perfluoroalkoxy alkane⁵⁸, thermoplastic polyurethane⁵⁹, polymethyl methacrylate⁶⁰, polycarbonate^22,61, and polyvinyl alcohol⁶².

The μETF produces 3D structures with inherent round edges and gently sloped sidewalls, which help minimize tissue damage during and after surgical implantation. Mechanical stress within MEA and surrounding tissue was assessed via FEA during forced sliding of an 80-μm-height 3D structure into viscoelastic tissue, as shown in Supplementary Fig. 24. The μETF LCP electrode with native round edges was estimated to reduce the maximum stress within the electrode and on tissue by more than 60%, compared to LCP MEA with sharp edges. The stress mitigation of the μETF LCP structure was more pronounced when a comparison was made against silicon-based 3D structures with the same shapes.

Although degenerated retinas are known to undergo structural changes over time^63,64,65, a normal human retinal model was employed in this study for the computational simulation of subretinal stimulation. This was primarily due to a lack of consistent structural information available for degenerated retinas in humans. While discrepancies may exist between computational simulations and clinical applications, it is anticipated that the benefits of protruding electrodes predicted in a normal retinal model can also be applied to degenerated retinas, primarily because of the reduced distance between the target cells and electrode surfaces. Also, it is noteworthy to mention about the threshold criteria for retinal activation in the FEA study. This study adopted an E-field threshold of 3000 V/m, which is assumed to induce cross-cellular membrane depolarization by 30 mV in a 10 μm long cell²⁷. Alternatively, the gradient of the E-field has also been proposed as an important factor contributing to the facilitated depolarization of neural cells^66,67. Exploring the impact of this criterion on retinal activation will be part of our future work to further refine the model.

The accuracy and minimum feature size of the μETF structures are associated with two factors: the precision of the 3D-printed mold, and the mold–MEA alignment. The diameter and height of a single pillar from the 3D-printed mold was measured to have ~3% deviations from the CAD design (Supplementary Fig. S25). Although this resolution is generally suitable for a wide range of neural interfacing applications, recent rapid advancements in 3D printing suggest the potential future availability of molds with even higher precisions. The mold was aligned with the planar MEAs by matching the laser-drilled holes on the arrays with corresponding pins on the metal jig, ensuring a minimum in-plane feature size of approximately 100 μm. However, this approach may be susceptible to mismatches caused by different thermal expansion ratios between the metal jig, epoxy mold, and polymer array. To enhance the accuracy, the 3D microstructures and aligned pins could be printed monolithically, and error-compensating laser drilling could be employed to further reduce the impact of thermal expansion mismatches.

The maximum protruding height dependents on various gemetric parameters of the protruding structure, inlcuding the height, base diameter, top diameter, sidewall profiles, as well as serpentin shapes. In the proposed design strategy for subretinal stimulation, the process was optimized for a target height of 80 μm and reliably achieves a protrusion height of up to 100 μm with a top diameter of 100 μm, corresponding to an aspect ratio of 1. Optimizing these parameters could enable even higher protrusion, potentially exceeding 200 μm. For instance, larger electrodes may achieve protrusion higher than 200 µm; however, this is a function of multiple variables and requires a comprehensive design optimization. Establishing the relationship between these geometric parameters and the resulting 3D structures is one of our ongoing research.

Finally, the thermoplastic LCP substrate employed in this study is known to exhibit an exceptionally low water absorption rate (<0.04%), potentially ensuring the long-term stability of the 3D MEA^32,68. Therefore, combining versatile 3D forming with durable LCP packaging is considered a suitable approach for various neural interface applications.

Methods

Microfabrication of LCP MEA

A planar LCP MEA was fabricated using a previously reported procedure based on conventional microfabrication technologies for thin-film polymer–metal–polymer structures^26,32,68. Briefly, Ti/Au seed layers (50 nm/50 nm) evaporated on a 25-μm-thick LCP film (Vecstar CTQ-25, Kuraray) underwent photolithography to define negative patterns (8 μm thickness, AZ P4620, Merck), followed by gold electroplating. After seed layer removal (gold etchant and buffered oxide etchant 10:1, Sigma-Aldrich), a 25-μm-thick LCP cover layer (Vecstar CTF-25, Kuraray) was thermally laminated at 285 °C with a pressure of 4 kgf/cm² for 30 min using a heating press (Model 381, Carver).

3D mold preparation for μETF

A master mold carrying microscopic 3D structures and alignment keys was created via precision 3D printing (microArch S130, Boston Micro Fabrication) and then replicated using a high-temperature epoxy to tolerate the thermoforming process at 200 °C. The replication was performed via double casting using polydimethylsiloxane (PDMS) as the negative mold. High-temperature-resistant epoxy (Duralco 4460, Cotronics) was poured over the PDMS negative mold, which was cured in two steps: first at 120 °C for 4 h, followed by post-curing at 230 °C for 16 h to reinforce its thermal and mechanical tolerance.

Alignment and pressing setup for μETF

The alignment and pressing setup consisted of a pair of top and bottom metal jigs; a rubber sheet (RBSM5-100, Misumi) and a rubber frame for uniform pressure; and a cylindrical rubber stamp and a rubber frame for pressing the electrode array against the mold. The epoxy mold was assembled onto the bottom metal jig, on which the planar MEA array was loaded. Precise alignment was achieved by matching the laser-drilled align holes on the electrode layer with the align pins on the mold. The μETF was performed at a pressure of 6.4 kgf/cm² applied at 200 °C for 30 min. Then, a UV laser cutter (Samurai UV marking system, DPSS Lasers Inc.) was used for site opening and outlining. Oxygen plasma cleaning (CIONE 4, Femto Science) was performed at 80 W and 80 sccm for 30 min to remove the laser burrs.

Electromechanical considerations of 3D MEAs with varying pattern designs

The 25-channel MEAs were prepared with varying gold thicknesses and line shapes. The 300-nm-thick gold patterns were prepared via evaporation of Ti/Au layers, photolithography with a negative photoresist (NR9-3000py, Futurrex), and wet etching. The MEAs were not encapsulated by cover layers to facilitate the observation of cracks in the gold patterns. The failures of gold patterns after μETF were categorized into two mechanisms. The “electrode crack” is defined as a failure due to distributed cracks within the gold electrode sites, whereas “line disconnection” refers to a disconnection of gold lines. Number of samples for analyzing the failure rate are: 20 for (I)-straight, 20 for (I)-wavy, 40 for (II)-straight, 50 for (II)-wavy, 50 for (III)-straight, and 50 for (III)-wavy.

IrOx electrodeposition

The laser-opened gold electrodes were electrodeposited with IrOx, following the previously reported protocol^42,43 using a potentiostat (CompactStat, Ivium Technologies). The electrodes were subjected to four cleaning cycles of voltage sweep in 1 M sulfuric acid from −0.4 to 1.4 V at 50 mV/s. Triangular potentials (0 to 0.55 V) were then iterated for 200 cycles at 50 mV/s, followed by 2000 cycles of rectangular pulses (0.55 V, 0.5 s, and 1 Hz).

Electrochemical characterization

The electrochemical characterizations were performed with a three-electrode system in the frequency range of 1 Hz to 100 kHz using the potentiostat and phosphate-buffered saline (1X, PH 7.4, Gibco). The CV curves were measured at a scan rate of 50 mV/s in the voltage range of −0.6 to 0.8 V, from which the CSC_C was calculated using the time integral of the cathodically enclosed area in Eq. (1):

$${{rm{CSC}}}_{{rm{C}}}=frac{1}{{vA}}rm{int }_{!!!{E}_{c}}^{{E}_{a}}left|iright|{dE},,[{rm{mC}}/{{rm{cm}}}^{2}]$$

(1)

where v is the scan rate, A is the geometric area of microelectrode, i is the current, and E_a and E_c are the anodic and cathodic potential limits, respectively¹¹. Number of samples for comparing EIS of the MEAs before and after μETF are: 56 for planar gold, 31 for protruding gold, 43 for planar IrOx, and 63 for protruding IrOx. A number of samples for comparing CSC_C of the MEAs before and after μETF are: 12 for planar gold, 16 for protruding gold, 10 for planar IrOx, and 15 for protruding IrOx. A single data point for the box plot represents a single electrode channel from multiple MEA devices.

FEA simulations

Mechanical stress induced in the gold traces was computed using the structural mechanics module in COMSOL Multiphysics 6.0. The in-plane serpentines were implemented by applying a constraint in gold layers to only deform in the axial direction, while no constraints were applied to the out-of-plane serpentines allowing both axial elongation and 3D buckling.

The I_th, DR, and spatial resolution in subretinal stimulation were numerically simulated using an AC/DC module in COMSOL Multiphysics 6.0 based on a human retinal model^49,50. The stimulation threshold was defined as the stimulating current that generates an electric field of 3000 V/m at the location of the bipolar cells in the INL, which is the target layer for subretinal stimulation^27,51,52. The spatial resolution estimates the extent of inter-channel interference as evaluated by the Michelson contrast (MC)¹¹. The MC measures the ratio of the minimum and maximum electric field strengths at the plane of the INL upon simultaneous current injection from neighboring channels using Eq. (2):

$${rm{MC}}=frac{{left|Eright|}_{max }-{left|Eright|}_{min }}{{left|Eright|}_{max }+{left|Eright|}_{min }}$$

(2)

Fluorescent calcium imaging system

A custom-built optical system incorporated both fluorescence and bright field microscopes. The bright field microscope utilized a Köhler illumination configuration, consisting of a near-infrared light emitting diode (LED) at λ = 780 nm (M780L3, Thorlabs), a collector lens, a condenser lens, and two iris diaphragms to ensure uniform illumination of the sample. The fluorescent microscope involved a 565 nm LED (SOLIS-565C, Thorlabs), a dichroic mirror in combination with excitation and emission filters (MDF-MCHA, Thorlabs) to separate the excitation light from emission light. Images were focused onto the CMOS camera using a 40X objective lens (LUMPLFLN40XW, Olympus) with a numerical aperture of 0.8, and a tube lens with a focal length of 200 mm.

AAV-meditated gene transfection and retinal preparation

All animal experimental procedures were approved by the Institutional Animal Care and Use Committee (IACUC; PNU-2023-3237) of Pusan National University. Male C57BL/6 mice (10-week-old, 20–25 g) were housed in a cage under 12 h light/dark cycles with controlled temperature (22 ± 2 °C) and humidity (55 ± 5%), and ad libitum access to food and water. The AAV vectors were used to transfect mouse retinal neurons with sRGECO (AAV8-Ef1a-sRGECO; 2 × 10¹³ GC/mL in HBSS, plasmid number of 137125; Addgene). The mice were anesthetized with avertin (250 mg/kg, Sigma-Aldrich) and placed on a heating pad at 37 °C. A 30G syringe was used to puncture a hole around the limbus. The AAV vectors carrying the genetically encoded calcium indicator were injected into the vitreous for 10 s using a 33G blunt needle (World Precision Instruments)^69,70. After 3 weeks of injection⁶⁹, the retina was prepared following the previously reported protocol by the authors⁷¹. The detached retina was transferred into the electrode-mounted chamber (RC-27L, Warner Instruments LLC) with the photoreceptor facing down, and immobilized by a nylon-mesh anchor (HSG, Ala Scientific Instruments). A total of four retinas were used in this study.

Ex vivo retinal experiments

During the procedures, the prepared chamber was continuously perfused by oxygenated Ames’ medium (A1420, Sigma-Aldrich) containing a synaptic blocker cocktail of CNQX (25 μM; Hello Bio) and D-AP5 (25 μM; Hello Bio) at 34 °C with a constant flow rate of 5 mL/min^70,72. The stimulating MEA was connected to a stimulus generator (STG4004, Multi-Channel Systems) with a reference platinum wire electrode (MW-4310, BASI Research Product). Electrical stimulation was delivered in a burst of seven cathodic-first biphasic pulses at a frequency of 60 Hz, duration of 60 μs, and inter-phase delay of 100 μs with varying amplitudes.

Data processing for ex vivo experiments

The stimulation threshold was determined by fitting ΔF/F₀ = (F_peak – F₀)/F₀ versus current amplitudes, using a modified sigmoid function, S(x) in Eq. (3):

$$Sleft(xright)=frac{a}{a-frac{a}{1+exp left(btimes cright)}}times left(frac{a}{1+exp left(-btimes left(x-cright)right)}-frac{a}{1+exp left(btimes cright)}right)$$

(3)

where x is the stimulation current, and a, b, and c are fitting parameters. The threshold is defined as the current corresponding to 0.5F_max. The spatial extents were determined by fitting ΔF/F₀ versus distance from the electrode center using a Gaussian function, (G(x)) in Eq. (4):

$$Gleft(xright)=atimes exp left(-btimes {x}^{2}right)$$

(4)

where x is the distance away from the electrode center, and a and b are fitting parameters. The spatial extent was defined as half of HWHM. The calcium intensities were normalized with respect to the steady-state values of fitted functions (i.e., S(∞) in Fig. 6e and G(∞) in Fig. 6g), which allows for a straightforward definition of the thresholds as 0.5 after normalization. All the fittings were performed through nonlinear least-squares methods in Matlab.