Retinal degenerative diseases like retinitis pigmentosa and age-related macular degeneration cause severe vision impairment by damaging photoreceptor cells. However, inner retinal neurons (ganglion and bipolar cells) often remain functional. Electronic retinal prostheses offer a promising solution by electrically stimulating these surviving neurons using photoresponsive devices, generating visual perceptions (phosphenes). While subretinal prostheses offer stable mechanical fixation, they present greater surgical difficulty and risks. Epiretinal prostheses, placed on the RGC side, are less invasive but face challenges due to inconsistencies between the implant and the often uneven retinal surface in severely degenerated retinas. This mismatch increases the electrode-cell distance, leading to decreased spatial resolution and axonal stimulation causing irregular visual perceptions. To overcome these limitations, ultrathin and flexible optoelectronics have been explored, but even these devices struggle with the non-uniformity of the retinal surface. Three-dimensional (3D) microelectrodes offer improved selectivity and spatial resolution by reducing electrode-cell distance, but prior 3D electrodes made of rigid materials cause mechanical mismatch and damage to the soft retinal tissue. This research introduces a soft artificial retina to address these limitations.

Existing retinal prostheses, while showing promise, suffer from limitations in spatial resolution and biocompatibility. Subretinal implants, though offering stable fixation, are surgically challenging and can damage residual photoreceptors. Epiretinal implants are less invasive but struggle with the uneven retinal surface in advanced degeneration, leading to increased electrode-cell distance and consequently, reduced spatial acuity and inaccurate stimulation. Previous attempts to improve the situation using flexible materials still faced issues with conforming to the irregular retinal surface. The use of 3D microelectrodes, while offering advantages in selectivity and spatial resolution, has been hampered by the use of rigid materials causing tissue damage. The literature highlights the need for a biocompatible, flexible, and high-resolution device capable of overcoming these challenges.

This study developed a soft artificial retina (10 µm thick) combining flexible, ultrathin photosensitive transistor arrays with 3D liquid metal (LM) stimulation electrodes. Eutectic gallium-indium (EGaIn) alloy was 3D-printed to create the soft, biocompatible stimulation electrodes with high resolution. Platinum (Pt) nanoclusters were coated onto the tips of the LM electrodes to enhance charge injection efficiency. The device fabrication involved several steps: 1. **Phototransistor array fabrication:** A high-resolution array of silicon phototransistors was fabricated on a flexible polyimide substrate. This involved patterning silicon channels, depositing source/drain/gate electrodes, and depositing a dielectric layer. Parylene C was used for biocompatible encapsulation, with openings created for the 3D LM electrodes. 2. **3D LM electrode printing:** A high-resolution direct printing system was used to precisely deposit EGaIn pillars onto the transistor drain electrodes. The pillar height and diameter were controlled by adjusting the printing parameters (nozzle diameter and stage speed). 3. **Pt nanocluster coating:** Pt nanoclusters were electroplated onto the tips of the EGaIn pillars to enhance charge injection. Parylene C was used to encapsulate the sidewalls of the pillars, leaving only the tips exposed. 4. **Biocompatibility testing:** *In vitro* tests using human retinal pigment epithelium cells confirmed the biocompatibility of the device, showing negligible cytotoxicity. *In vivo* biocompatibility was evaluated by implanting the device into rd1 mice for 5 weeks, observing the retinal status (fundus imaging, immunohistochemistry) and confirming the absence of inflammation. 5. **Electrophysiological characterization:** The optoelectronic properties of the phototransistors (current-voltage characteristics, response time, light responsivity) were measured. Electrochemical impedance spectroscopy and cyclic voltammetry were used to characterize the LM electrodes, showing impedance significantly lower than uncoated electrodes. 6. **Ex vivo experiments:** Isolated retinas from wild-type (WT) and rd1 mice were used to assess the device's ability to elicit retinal responses. Electrical stimulation was applied, and recorded responses were analyzed to compare the evoked potentials and firing rates of RGCs between WT and rd1 retinas, with and without light illumination. 7. **In vivo experiments:** The device was implanted into rd1 mice. Full-field or localized laser illumination was applied, and the resulting neural responses (spike trains, firing rates) were recorded and mapped to assess vision restoration. Unsupervised machine learning (hierarchical and K-means clustering) was used to classify and analyze the recorded neural spikes, separating somatic RGC responses from axonal responses.

The fabricated soft artificial retina showed excellent biocompatibility both *in vitro* and *in vivo*. The 3D-printed EGaIn electrodes with Pt nanocluster coating exhibited significantly reduced impedance compared to uncoated electrodes, resulting in more efficient charge injection. *Ex vivo* experiments demonstrated that the device could elicit RGC spikes in both WT and rd1 mouse retinas with comparable magnitudes, even in the absence of functional photoreceptors. Notably, the rd1 retinas showed earlier and more pronounced firing activity. The firing rate increased proportionally with light intensity. Using 3D microelectrodes compared to flat electrodes significantly increased firing activity in both WT and rd1 retinas, highlighting the advantage of the 3D structure in proximity to RGCs. Unsupervised machine learning effectively classified RGC spikes, enabling the identification of somatic RGC responses, indicating selective stimulation of RGC somas. Critically, *in vivo* experiments with rd1 mice showed that light illumination induced spiking activity in RGCs within the illuminated area, providing strong evidence for vision restoration. Spatially mapped firing rates mirrored the light illumination pattern, further supporting the selective stimulation achieved with the device.

The findings demonstrate the successful development and implementation of a novel soft artificial retina capable of restoring vision in a mouse model of retinal degeneration. The use of biocompatible liquid metal electrodes allows for conformal contact with the uneven retinal surface, significantly reducing electrode-cell distance and improving stimulation efficiency. The integration of photosensitive transistors provides a light-responsive mechanism for stimulation, effectively mimicking the function of photoreceptors. The successful application of unsupervised machine learning provides a robust tool for analyzing complex retinal signals and separating RGC somatic activity from axonal activity, potentially leading to improved clarity of visual perception. The *in vivo* results showing localized light-induced RGC activation strongly suggest the potential for vision restoration in patients with retinal degeneration. The ability to achieve selective stimulation of RGCs minimizes axonal activation, leading to more natural visual perceptions.

This study successfully demonstrated a novel soft artificial retina using 3D liquid metal microelectrodes for vision restoration in a retinal degeneration mouse model. The device's biocompatibility, high charge injection efficiency, and ability to elicit localized light-induced RGC responses hold significant promise for future development of retinal prostheses. Future research should focus on increasing the pixel density and device size for larger animal models and ultimately, human clinical trials. Further exploration of nanoscale materials to enhance stimulation efficacy at smaller electrode sizes is warranted.

The current *in vivo* study used a small device (36 pixels) suitable only for mouse retinas. Scaling up the device for larger animal models and ultimately humans presents a challenge. While the study demonstrated biocompatibility over a 5-week period, longer-term studies are necessary to fully assess the long-term biocompatibility and efficacy of the device. The study focuses on a specific mouse model of retinal degeneration and the results may not be directly generalizable to all types of retinal degeneration.