Main

The additive manufacturing workflow, originally conceived more than 40 years ago, has largely remained unchanged. Although 3D printing plays a key role in medical devices, microfluidics, bioengineered tissues, as well as driving innovation in the automotive and aerospace sectors5,6,7, the printing process always begins with users defining the desired part by means of computer-aided design (CAD) software, which is then translated to the printer and fabricated either layer by layer or using volumetric methods1,2,[8](https://www.nature.com/articles/s41586-025-09436-7#ref-CR8 “Bernal, P. N. et al. The road ahead in materials and technologies for volumetric 3D printing. Nat. Rev. Mater. https://doi.org/10.1038/s41578-025-00785-3

(2025).“). Further research on embedded sensors and feedback loops aims to improve automation and is making substantial strides towards performing in-line quality control of printed objects9,10,11. Yet, 3D printers remain primarily tools that passively execute a command while being agnostic to the composition and nature of the environment in which the printing process takes place.

Enabling printers to detect and respond to contextual cues can open new avenues for many applications, including soft robotics, hierarchical composites and the bioprinting of living cells and human tissues. In fact, the functionality of such systems is intimately linked to both their architecture and relative positioning of their individual components (that is, particles, fibres, living cells)12,13, whose precise patterning within printed objects cannot be fully controlled. Recent advances in computer vision and artificial intelligence have the potential to greatly enhance this approach. Concurrently, the advent of volumetric printing, including tomographic volumetric additive manufacturing, has enabled the extremely fast production of large parts with virtually unconstrained design freedom, using visible light fields to polymerize photo-responsive resins in a layerless fashion. Owing to its contactless nature, volumetric printing excels at overprinting—that being, non-invasively printing onto or across existing objects, even when produced with other techniques14. This includes building multimaterial structures15 and safely encapsulating fragile living cells and organoids16. Such attributes position volumetric printing as an ideal demonstration platform for new fabrication pipelines that would be challenging to implement using traditional technologies.

In this study, we report a new technique to equip 3D printers with the ability to map the composition (chemical and architectural) of the printable material and to take autonomous, informed decisions on which geometries to print. This new workflow allows the printer to produce, within seconds, volumetric, generative designs that adapt and conform to embedded features within the resins, enabling exploration of diverse applications in data-driven additive manufacturing. It also supports complex materials, including living cells with biologically friendly hydrogels, and automates overprinting to create complex multicomponent structures, including models comprising several tissue types (that is, vascularized tissues; bone and cartilage in osteochondral models), and mechanical joints with interlocked movable parts, as examples among a broad array of possible printable designs.

3D printing guided by volumetric imaging

We termed this workflow GRACE, short for Generative, Adaptive, Context-Aware 3D Printing. We first demonstrated GRACE by integrating volumetric printing with light-sheet imaging. The light sheet rapidly maps the printing volume in 3D to extract positional, morphometric and spectral (that is, fluorescence) information from the contents of the vial. This serves as input for multiparametric modelling algorithms that process the data to generate precisely targeted geometries for printing, effectively enabling adaptive fabrication. Although GRACE printing is relevant for various fabrication processes, we first showcased its key capabilities in the context of bioprinting, using cell-laden hydrogels as resins. In living organs, key tissue components, including structures composed of several cells, develop to adapt to the needs of the resident cells. For instance, blood vessels grow into intricate networks to reach each individual cell, ensuring access to nutrients. At present, printing technologies cannot fully recapitulate this process, as cells are randomly or homogeneously distributed within a printed hydrogel. With GRACE, we demonstrated on-the-fly generation of 3D models to create positive and negative features, including targeted vessel-like channel networks that can precisely reach cells, cell clusters and organoids of interest, resulting in improved functionality of the bioprinted cells. We further demonstrated the production of interconnected geometries and the precise encapsulation of various features embedded within the resin. Moreover, we enabled the automated alignment of new prints onto pre-existing ones, permitting the generation of multicomponent constructs. GRACE also enabled the mapping of opaque features, countering shadowing artefacts and improving (over)printing quality.

The experimental device for GRACE comprises two main components: a custom-made tomographic volumetric printer and a light-sheet microscopy path (Fig. 1). We chose polar light-sheet microscopy as the primary scanning modality for its capacity for rapid, large-scale imaging and its ease of incorporation within our set-up, requiring no modification to the printing path. Light-sheet generation could be achieved through two methods: (1) by encoding a light-sheet pattern (that is, a single column of activated pixels) onto the digital micromirror device (DMD) or (2) using an external laser source with dedicated optics. Although the former required no extra hardware, for most of our experiments, we opted for an external light-sheet configuration, maximizing power and signal-to-noise ratio during scanning.

Fig. 1: Experimental GRACE printing. Schematic of the experimental device, with the light sheet (green), imaging and printing (violet) optical paths indicated. The printing path consists of a 405-nm continuous-wave laser source (CW405), collimating lens (L1), fibre coupling lens (L2), square core multimode-fibre optic (FO), collimating lens pair (L3 and L4), DMD, a 1:1 magnification 4f relay and Fourier filter (L5, Iris and L6), and a folding mirror (M1). The light-sheet path comprises three collimated laser sources at 450 nm, 532 nm and 650 nm (CW450, CW532 and CW650, respectively), combined by means of dichroic beam-combining optics (DM1 and DM2), 2.5× beam-reduction optics (L7 and L8), a 30° Powell lens (PL) and cylindrical lens (CL) for focusing the light sheet. Along the imaging path, a monochromatic CMOS sensor with imaging lens (IL) captures fluorescent signals by means of a fluorescence imaging filter (F1). All three optical paths intersect on the central axis of a resin-filled print vial, coupled to a rotational stage. To minimize refractive errors during both printing and imaging, the vial itself is immersed within a liquid-filled (refractive-index-matching fluid, depending on the resin to be printed, but typically water when working with low-concentration hydrogels) quartz cuvette (CUV) whose optical faces are precisely aligned orthogonal to each path.

To enable the 3D mapping of features embedded within the printing volume, we developed a protocol that synchronized the motion hardware with the imaging and printing systems (Extended Data Fig. 1a, Supplementary Video 1 and Supplementary Methods 1–4). On initial homing of the vial, the light sheet acquires images that serve as the input of the computer-vision routines. The resultant fluorescent emission of the optical section is captured by the camera at many angular increments to obtain a dense set of polar cross-sections through the axis of the vial. The process can be repeated for different fluorescence channels. The extracted data are then assigned corresponding per-pixel Cartesian coordinates, producing a 3D point cloud in print-volume coordinate space. Depending on the scanned features and intended model, the raw data can either be sent directly to parametric modelling software or undergo further processing in the form of cluster detection. For this, we used density-based spatial clustering of applications with noise (DBSCAN), chosen for its efficiency with large datasets, ability to identify arbitrarily shaped clusters, automatic outlier exclusion and lack of requirement for pre-defining the cluster count17. Clustering detection was applied when numerical indexing of individual features was required or when centroid coordinates were preferred, as is the case with symmetrical features such as particles, organoids or microspheres. We tested our imaging and feature-detection accuracy with suspensions of standardized polyethylene microspheres (diameter 150 or 500 µm), acquiring images using both a DMD-generated and an external light sheet. The results were benchmarked against conventional microscopy, showing no difference in size detection accuracy (Supplementary Methods 5 and Supplementary Fig. 1). This workflow provided robust feature detection in complex, heterogeneous samples, with the resultant coordinate data passed to the parametric modelling software to automatically generate geometries targeted around those scanned features (Extended Data Fig. 1b and Supplementary Figs. 2–4).

Context-driven parametric structures using GRACE

As a testing platform for printing context-driven architectures, we produced fluorescently stained alginate microspheres of various radii (0.15–0.90 mm, a size range compatible with that of organoids18) (Supplementary Methods 6). These were mixed into a gelatin methacryloyl resin (GelMA, 10% w/v), with lithium phenyl-2,4,6-trimethylbenzoylphosphinate (LAP) as a photoinitiator. Using only spatial information with clustering detection to determine the centroids of the alginate particles, we first successfully demonstrated context-aware prints with vascular-like channel networks (diameter 450 ± 20 μm) targeted around individual spheres, with a 300-μm offset (Fig. 2a, Supplementary Methods 7 and 8, Supplementary Figs. 5–7 and Supplementary Video 2). Further proof of perfusability of complex vascular-like structures with different hydrogel resins was also shown (Supplementary Methods 9 and Supplementary Fig. 8). We also fabricated interconnections between several particles (Fig. 2b) and showed the encapsulation of individual features (Fig. 2c). Furthermore, discriminating different particle subpopulations is possible through processing spectral and morphometric data. Polydispersed alginate particles were detected and their individual radii (0.15–0.90 mm) measured by means of cluster detection. Using this information, GRACE generated conditional geometries defined by a size threshold: spheres with r < 0.5 mm received a single grazing channel, whereas those with r ≥ 0.5 mm were enclosed by a more complex spherically wrapped channel network (Fig. 2d). Moreover, we applied a similar conditional approach to discriminate between different fluorophores. Here the model generated solitary grazing channels around Cy3.5-stained microparticles, whereas spherically wrapped networks were generated around Cy5-stained particles (Fig. 2e). Scanning, feature isolation, model generation and printing were completed within approximately 4 min (per spectral channel), underscoring the compatibility with the rapid, high-throughput capabilities of volumetric printing. Detailed processing times are reported in Extended Data Table 1.

Fig. 2: GRACE allows printing adaptive and feature-driven prints with complex geometries. A showcase of GRACE printing by generating targeted features around randomly distributed, fluorescently stained alginate spheres detected within the resin. a, A spherically wrapped channel network architecture is generated around the alginate spheres, showing the process from acquiring the raw data to generating the target model (shown rendered) and, finally, printing and imaging the resultant construct. The light-sheet sections at different depths of the printed gel show in greyscale the stained GelMA and the cross-sections of the convoluted channels printed within the gel, whereas the unstained particles are circled in red (or in green) to improve visualization and to distinguish them from cross-sections of the channels. Scale bars, 1 mm. b, Interconnection of randomly distributed alginate spheres with printed struts, showing a render of the target geometry (blue) and resultant light-sheet reconstruction in 3D after printing. c, Encapsulation, with corresponding light-sheet 3D reconstruction post-printing indicating presence of spheres. Scale bar, 1 mm. d,e, Parametric discrimination of generated features based on size (d) or spectral emission (e). Here different populations receive either a single grazing channel or a more complex spherical channel network. Scale bars, 2 mm. f, Automated alignment of a cartilage model to femoral head in a two-part sequential print, with computation for the alignment (post-scanning) taking <5 s to perform on a common personal computer. Our transformed cartilage model was printed directly onto the femoral head, resulting in a multicomponent sequentially printed construct with correct relative positioning of its components. Scale bar, 2 mm.

Finally, we demonstrated multistep fabrication through automatic alignment (Fig. 2f), eliminating the need for manual repositioning of sequential prints—a process1,19 that is error-prone and prohibitively slow. GRACE automatically detects and aligns with previously printed structures, enabling the precise positioning of subsequent prints relative to existing ones, streamlining the fabrication workflow to build complex, multilayered or hierarchical structures with high repeatability. We successfully demonstrated this by first printing a model of a human femur, washing it and resuspending it in GelMA at a random orientation. Through an iterative closest point algorithm, the optimal rigid transformation that aligns a pre-existing reference model (correctly positioned cartilage layer) to the scanned point cloud of the femur was identified (Fig. 2f). Our experiments demonstrated how GRACE adapts to various embedded objects (Extended Data Fig. 2) and generates precise, functional geometries with minimal user input after the initial parametric model definition. This level of automation and adaptability would be impractical, if not impossible, to achieve through manual positioning and 3D modelling.

Image-guided printing across opaque features

Shadowing artefacts from light-absorbing features is a common challenge in light-based printing. Occlusions within the projection path result in poor reconstruction quality, compromising dimensional accuracy and printing homogeneity when performing overprinting, especially for volumetric printing1,20. Addressing this challenge, we applied our light sheet as a profilometer, using the reflected signal to map the surfaces of occluding structures (Fig. 3a). To mitigate printing artefacts, we used object-space model optimization (OSMO) of tomographic reconstructions to iteratively optimize our tomograms for the presence of these occlusions21,22. This was demonstrated in two ways. First, we used an opaque, polymeric occlusion, consisting of ten 1-mm-diameter vertical pillars to provide a reproducible occlusive feature within our build volume. The quantity and spatial distribution of pillars was based on a Monte Carlo optimization by minimizing the Bhattacharyya coefficient (for contrast) and maximizing the Jaccard similarity index—both derived from the OSMO-corrected reconstructions (Fig. 3b and Supplementary Fig. 9). The resulting occlusion phantom (Fig. 3c) was embedded into our resin and scanned, to parametrically generate the representative occlusive geometry based on the pillar centroids and angles (Supplementary Methods 10 and 11 and Supplementary Fig. 10). On optimizing and printing the target geometry (eight-toothed cog model; Fig. 3b), the shadow-corrected structure demonstrated superior print quality compared with uncorrected output, with the latter showing regions of both under-crosslinking and over-crosslinking within the same print (Fig. 3d). This indicated that there would exist no possible dosage condition for which an accurate output could be attained without shadow correction. Meanwhile, the corrected output showed a more homogenous crosslinking behaviour, with finer details retained across the entire construct. This is corroborated by the Monte Carlo simulation, indicating improvements in both contrast—with the Bhattacharyya coefficient decreasing from 0.39 ± 0.01 to 0.15 ± 0.01—and in the similarity index, which increased from 0.70 ± 0.01 to 0.945 ± 0.007.

Fig. 3: Light-sheet mapping of occluding structures and shadow correction. a, Flow chart demonstrating the process of scanning, mapping and correcting for the presence of occlusions. b, OSMO-based reconstruction of a cog-like target model without and with shadow correction when influenced by ten randomly distributed pillar-like occlusions. Jaccard index and Bhattacharya coefficient demonstrate the relative improvements attained by the correction (mean ± s.d., t-test, n = 12, ***P < 0.001, DF = 22). c, Rendering showing the SLA-printed pillar occlusions and cog-like target geometry to be volumetrically printed around them. d, 3D light-sheet reconstruction of corrected and uncorrected prints, as well as a single optical section of the cross-section of each. Scale bars, 2 mm. e, Render showing the ball-in-cage model, with spherical target geometry inside. f, Light-sheet reconstruction of the resultant prints following the destructive removal of the occluding cage following printing. g, r.m.s. error of the printed part. A lower value indicates less deviation from the target geometry (mean ± s.d., t-test, n = 3, P = 0.0071, DF = 4, **P < 0.01). h, Sphericity of each printed sample. A higher value indicates that the sample is more spherical (mean ± s.d., t-test, n = 3, P = 0.0137, DF = 4, *P < 0.05).

Source Data

To tackle more complex, continuous occlusions, we extended this approach to print a ball-in-cage model (Fig. 3e). In this more challenging scenario, the cage creates multidirectional occlusions around the central sphere. Unlike our previous approach, in which we parametrically generated the representative occluding surfaces using the scanned point cloud data, here we instead combined our auto-alignment workflow to precisely match the a priori reference model to the scanned point cloud, thus aligning, rather than generating, the occluding geometry (Supplementary Fig. 11). The aligned mesh was used as the occlusion input for the OSMO-based optimization.

Shadow-corrected models demonstrated statistically significant improvements, compared with uncorrected prints (Fig. 3f and Supplementary Methods 12), with a reduction in surface root mean square (r.m.s.) error from 0.50 ± 0.09 μm to 0.18 ± 0.05 μm and an increase in sphericity from 0.830 ± 0.060 μm to 0.965 ± 0.006 μm (Fig. 3g,h). Finally, we demonstrated enhanced printing of vessel-containing structures surrounded by a shadowing cage (Supplementary Methods 13 and Supplementary Fig. 12). These experiments highlight the suitability of GRACE not only for creating context-aware geometries but also for mapping occlusive features within the printing volume and mitigating their influence.

Bioprinting with GRACE

Next we explored the potential of GRACE for biofabrication, a field aiming to produce engineered tissues for regenerative medicine, and as in vitro models for personalized medicine. Specifically, we aimed to fabricate adaptive geometries around living cells and organoids, showcasing: (1) the production and functionality of customized vessel-like channels; (2) the generation of automatically aligned multitissue prints; and (3) the compatibility of GRACE with further fabrication techniques. As a first demonstrator—inspired by how in vivo blood vessels grow to reach cells and provide nutrients—we assessed GRACE to automatically generate adaptive vessel-like networks optimized around dense cellular structures. Notable efforts in bioprinting focused on producing convoluted channel networks to nurture tissues23,24,25. However, none of these approaches can adaptively print channels reaching every cellular structure or organoid within the construct.

Here we combined GRACE with Embedded Extrusion‐Volumetric Printing (EmVP)3,26 to generate adaptive vascular-like architectures around toruses (extruded into GelMA) densely laden with insulin-secreting pancreatic cells (iβ-cells, 5.0 × 107 ml−1; Fig. 4a and Supplementary Methods 14).

Fig. 4: Bioprinting of functional living tissues with cell-location-driven features enabled by GRACE. a, Diagram of experiment process depicting the scanning of the volume, feature-driven model generation around the torus and printing of the scaffold. b,c, 3D segmentations of the resultant structures printed around the extruded toruses, both targeted and random, also showing light-sheet micrographs of the scaffold cross-sections at two planes. Scale bars, 2 mm. d, Graph showing the amount of insulin stored and released by iβ-cells by means of the bioluminescent reporter NanoLuc (Nluc) for the GRACE print, as well as for the random and bulk controls (mean ± s.d., ANOVA, n = 6, **P < 0.01, ***P < 0.001, F = 55.74). e, A light-sheet section of the construct immediately following printing with the automatically aligned femur-cartilage model showing distinct osteal and chondral regions. Scale bar, 2 mm. f, The histological sections following 4 weeks of maturation highlight the presence of a mineralized compartment (von Kossa positive, brown/black staining) and of glycosaminoglycans-rich cartilaginous tissue (SafO, red staining) (n = 3). Scale bars, 500 μm (left panels); 200 μm (right panels). g, 3D model of FLight construct generated with GRACE after scanning. Printing was performed around a subset of each spheroid population to avoid overcrowding. h, Resultant light-sheet 3D reconstruction after printing and washing. i, Filamentous construct is visible surrounding the spheroid. Scale bar, 100 μm. j, Discrimination of spheroids is evident in the schlieren image (captured immediately after printing) with fluorescence overlay, with the spheroids precisely positioned through the central axis of each star or circle FLight structure. Scale bar, 2 mm.

Source Data

Networks around the toruses were generated with channel diameters of 1 mm at the inlets, tapering to 0.4 mm at the scaffold midplane, maintaining a fixed surface area of approximately 180 ± 10 mm2 (Supplementary Fig. 13) and a 300-μm offset from the torus (Fig. 4b). To evaluate the efficacy of this approach, we compared these adaptive structures to two controls: randomly generated channels of the same surface area (Fig. 4c) and a bulk structure without channels as negative control. Following 24 h of dynamic culture, we observed a substantial increase in proinsulin secreted from GRACE-printed structures compared with both the random non-targeted and bulk controls (3.2 ± 0.3, 2.0 ± 0.4 and 1.4 ± 0.1 × 105 relative light units, respectively) (Fig. 4d). This result suggests superior mass transport within the adaptive geometries, probably because of the improved proximity of the cells to the surface of the channels. This therefore results in shorter diffusion distances when compared with randomly distributing channels throughout the whole construct.

Next, through automated alignment, a two-component cell-laden bone and cartilage model was produced. We prepared two GelMA-based resins containing articular cartilage-derived progenitor cells (ACPCs, 1.0 × 107 ml−1) and bone-marrow-derived mesenchymal stem cells (MSCs, 5.0 × 106 ml−1; Supplementary Methods 15). Femur models were first printed within the MSC-laden GelMA, washed, then placed within the printing vat filled with the ACPC bioresin. The MSC component was scanned to determine the orientation and position of the femur. The cartilage phase was then automatically positioned and printed to form a layer around the femoral head, building a native-like osteochondral architecture (Fig. 4e). Cells remained functional and persisted in their intended compartment (chondral or osteal) over 4 weeks, with ACPCs and MSCs synthesizing cartilage and mineralized bone matrix components, respectively, as confirmed by means of histology (Fig. 4f, Supplementary Fig. 14 and Supplementary Methods 16).

Finally, we demonstrated the versatility of GRACE by using an alternative printing modality: filamented light (FLight) biofabrication, a vat polymerization technique that generates structures composed of multicentimetre-long aligned microfilaments16. Here two differently stained populations of MSC spheroids (Supplementary Methods 17) were identified within the vat and the parametric model encased each group in a unique shape (stars or circles for green-stained or red-stained spheroids; Fig. 4g–j), resulting in the formation of elongated filamentous constructs spanning the width of the print volume (Fig. 4h–j). This showcases the ability of GRACE to adapt to different printing modalities while maintaining its core capability of generating context-aware, population-specific geometries.

Discussion and outlook

This study introduced GRACE, an innovative workflow that makes use of the unique attributes of volumetric printing to generate context-aware geometries. By integrating light-sheet microscopy, computer vision algorithms and parametric modelling, GRACE can detect and respond to features across several scales, from organoids to macroscopic structures, and rapidly fabricates complex geometries that dynamically adapt to arbitrarily distributed features within the print volume. This level of automated, context-driven fabrication would be prohibitively time-consuming and impractical with manual design. GRACE operates with minimal user intervention, requiring only experiment-specific adjustment of the parametric models (Supplementary Fig. 6), thus greatly streamlining the fabrication workflow while simultaneously expanding the complexity and functionality of achievable structures. The versatility extends beyond its current implementation, as the concept could be integrated into other printing approaches. Examples include xolography, which inherently uses light-sheet optics27, multiphoton printing, acoustic-based printing28,29 or extrusion printing in suspension baths30. As the parametric models are agnostic to the imaging method generating the input data, alternative scanning modalities could be explored. For instance, optical tomography, or holographic approaches31,32, for non-fluorescent imaging, could provide a broader safe photoexposure window, although we did not see any relevant phototoxic effects within our experiments (Supplementary Methods 18 and 19 and Supplementary Figs. 15 and 16). In terms of future directions, the ease of performing overprinting offered by GRACE could aid several innovative applications, for example, in soft robotics, for introducing more refined polymeric skins onto previously printed movable parts or accurately controlling the geometry of hydrogel-based osmotic actuators overprinted across skeletal-like scaffolds22.

The GRACE workflow has direct implications for biofabrication, with the creation of biomimetic scaffolds that can adapt to the spatial distribution of cells or organoids, and for the fabrication of tissue constructs with highly controlled architecture, relevant to regulate cell function and tissue maturation. These systems could already serve as models for biomedical and pharmaceutical research. Although light-sheet imaging at present enables scanning of multicentimetre-sized volumes33, future work would be necessary to produce constructs having (human) full-tissue-scale sizes. For instance, combining light-sheet imaging and GRACE with movable vats for volumetric printing could expand the imaging and fabrication range, allowing a form of mosaicking the sample volume34. Moreover, with larger imaging and printing ranges, new techniques to mitigate scattering will become increasingly needed35,36,37. In parallel, much progress has been made in designing self-assembly materials that can be used in bioprinting, which allow to tune the cellular microenvironment at the (sub)cell-level scale38. Converging these classes of materials with GRACE could further permit to better approximate the hierarchical composition of living tissues at the macro-to-micro scale (through printed architecture), down to the micro-to-nano scale (provided by the structured materials). GRACE also opens new avenues for adaptively modifying an object at any time point post-printing. Modifications could include spatial-selective grafting of biomolecules39 and modulating stiffness gradients40 or viscoelasticity41. Altogether, these advancements underscore the versatility of GRACE and its potential in both general additive manufacturing and bioprinting contexts, representing a shift in how printing is performed.

Methods

Tomographic volumetric printer

The custom-built volumetric printer used in this study (Fig. 1a) used a 405-nm laser source shaped into a flat-top intensity profile through coupling of the beam into a square core fibre (WF 70 × 70 μm, CeramOptec). This shaped beam was collimated and used to illuminate a DMD (Hi-Speed V-7000, ViALUX), for which the resultant projection was Fourier filtered and imaged with 1:1 magnification using a 4f relay onto the centre of the printing volume in a telecentric projection regime. During the printing process, vials were immersed within a refractive index compensating bath comprising a square water-filled quartz cuvette (OP36, eCuvettes). This was required to minimize refractive errors arising from the curvature of the vial surface, for both printing and feature scanning. To accommodate different print requirements, we used borosilicate glass vials (Readily3D) with internal diameters ranging from 8.8 to 15.3 mm, selected according to the specific model parameters and target print volumes. In terms of compatible resins, tomographic volumetric printing is compatible with processing light-crosslinkable materials, whereas other classes of materials can be added as inclusions in the resin vat, for instance, by means of embedded extrusion printing3, before polymerizing the resin.

Light-sheet imaging module

The light-sheet imaging path (Fig. 1) used three 40–50-mW diode lasers operating at 450 nm (RLDD450-40-5, Roithner), 532 nm (RLDD532-50-3, Roithner) and 650 nm (RLDH650M-40-5, Roithner) to cover a broad range of commonly used fluorophores. The beams were combined with dichroic mirrors, shaped into a flat-top fan profile with a 30° Powell lens (43-473, Edmund Optics) and then focused to the vial centre using an f = 50 mm cylindrical lens. Two scanning regimes were used: a half-sweep mode covering Θtotal = π (500 polar sections) and a Θtotal = 2π sweep mode (1,000 polar sections). The latter was preferentially used for samples exhibiting occlusive or highly scattering samples, as this provided bidirectional illumination of the features. Image acquisition for feature scanning/registration was performed using a monochromatic camera (Alvium 1800 U-240m, Allied Vision) in conjunction with an f = 50 mm C-Mount lens (MVL50M23, Thorlabs). With this hardware, we achieved a resolution of 14.47 µm along the image plane, suitable for capturing cellular aggregates and organoids. Higher-resolution hardware would be needed to extend the detection to the single-cell regime, hence the current system is limited to detecting larger particles. Exposure and gain parameters were manually set before each scanning session. A rotary filter mount containing band-pass filters for several commonly used fluorophores (GFP, Cy3.5, Cy5) was integrated post-objective to facilitate spectral selection of the emission signal while rejecting the backscattered laser line. Notably, although light-sheet imaging has been historically an expensive technology, new reports describe how to build open-source, affordable light-sheet systems. In our case, our simplified set-up required only the use of a laser source, a beam reducer, a Powell lens for beam shaping and a cylindrical lens for focusing the light sheet within the volume—all available as off-the-shelf components.

Image processing and feature registration for GRACE

A MATLAB script was prepared to perform several functionalities key to the GRACE workflow. This process (shown in Extended Data Fig. 1a, and detailed extensively in Supplementary Methods 4) was responsible for the following: (1) the initialization and synchronization of hardware (imaging, light sheet and printing) and software parameters; (2) the acquisition and processing of polar light-sheet image stacks; (3) the isolation and registration of features of interest from the background; (4) the conversion of image data to usable 3D coordinate data; (5) the processing of these data (for example, using clustering detection) to extract construct-specific information necessary for generating the desired construct; and (6) outputting these data for use with the parametric modelling software.

Data-driven parametric models

We used off-the-shelf software Rhino3D (Robert McNeel & Associates) in conjunction with its integrated Grasshopper (GH) visual programming environment. GH definitions were developed to create parametric models for each geometry type. These definitions performed three key tasks: (1) importing and synchronizing with data files containing coordinates, radii and other relevant data as described above; (2) using these data to generate the desired parametric geometry; and (3) baking and exporting the final model as an STL file suitable for 3D printing. Our work focused on three broad types of biologically inspired geometry: perfusable vessel-like channels surrounding scanned features with an inlet and an outlet, positive interconnected geometries and targeted single-layer encapsulation. In several of these cases, although not strictly necessary, we also used two freely available add-ons for GH. These were the Dendro plug-in[42](https://www.nature.com/articles/s41586-025-09436-7#ref-CR42 “Yein, R. Dendro. Github https://github.com/ryein/dendro

(2022).“), which facilitated convenient surface generation around point structures and the ShortestWalk plug-in[43](https://www.nature.com/articles/s41586-025-09436-7#ref-CR43 “Piacentino, G. Shortest Walk GH. Food4Rhino https://www.food4rhino.com/en/app/shortest-walk-gh

(2025).“), which makes use of the A* algorithm41 to calculate the shortest walk in a network of paths. We note that similar functionalities can be provided using other plug-ins (either inbuilt or third party) or by writing a custom script within GH. See Supplementary Methods 8 for further details on the creation of the parametric models and definitions.

Synthesis of GelMA

All chemicals were obtained from MilliporeSigma and used without further purification or modification, unless stated otherwise. GelMA was synthesized as previously reported[44](https://www.nature.com/articles/s41586-025-09436-7#ref-CR4