Scaffolding

Scaffolds provide the structural support and microenvironment that let cells grow and organize into functional 3D tissue rather than an unstructured mass. For cultivated meat the design target is demanding: a scaffold (or scaffold-free mould) must steer cell alignment, density, and mechanical texture toward something that recapitulates muscle and fat, while remaining manufacturable and, ideally, edible. The design space — geometry, material composition, and print or moulding parameters — is large and the physical readouts are slow and expensive to obtain, which is precisely the regime where machine-learning surrogates earn their place: they learn to predict the outcome of a slow biophysical simulation or a costly fabrication run, so the design loop can be run thousands of times instead of dozens.

Designing mould and scaffold geometry with learned surrogates

One line of work, from the Andrews group, targets the geometry of tethered moulds that shape self-organizing tissue. Andrews et al. (2023, ref #35) trained an image-to-image deep-learning model (a pix2pix implementation) on 6,400 cases generated by the CONDOR biophysical simulator, learning to map a mould’s geometry to the resulting fields of cell density and alignment. On 100 unseen moulds the surrogate’s cell-alignment statistics correlated with the full simulation at r > 0.98 (and mean density at r ≈ 0.88) while running at least 10,000× faster — roughly 0.1 s on a laptop CPU versus a CONDOR run — which is what makes high-throughput rational mould design feasible. Andrews et al. (2025, ref #34) closes the loop: it couples a genetic algorithm (populations of 1,000, ~15,000 evaluations) with a fast ML surrogate of the same biophysical framework to search for mould tethering strategies that maximize cell alignment and uniform density, completing a full optimization in about 45 minutes on a commodity laptop — a search that would take over 100 core-years if each candidate were evaluated by direct simulation. Both papers name cultivated meat as a target application.

Predicting the mechanical and print quality of scaffolds

A second line predicts the properties of a scaffold from its design rather than its shaping field. Bermejillo Barrera et al. (2021, ref #19) trained 3D convolutional neural networks on “virtual tomographies” — lattice geometries sliced into stacks of 2D images — to predict Young’s modulus, shear modulus, and porosity, with finite-element simulation supplying the mechanical training labels. Working from an initial library of only 20 lattices expanded by data augmentation and tested on 8 entirely new geometries, the authors are candid that accuracy is not yet ideal, but the study validates CNNs as a surrogate for FEM when repeated mechanical simulation is computationally unaffordable. Rafieyan et al. (2024, ref #20) works at the bioprinting end: the authors assembled MLATE, an open dataset of 1,171 3D-(bio)printed scaffolds spanning 60 biomaterials and 49 cell lines, deliberately restricted to pre-print, controllable features (bioink composition, extrusion pressure, nozzle speed and temperature), and trained more than 40 ML/DL algorithms to predict printability, cell response, and overall scaffold quality. Tree-ensemble classifiers — XGBoost, gradient boosting, extra trees, random forest, and LightGBM — gave the best performance, turning the curated dataset into a “virtual lab” that estimates scaffold quality from a candidate recipe in seconds.

Three more recent studies broaden this property-prediction approach across distinct fabrication routes. Nair et al. (2021, ref #214) applied random-forest regression with feature-importance analysis to ice-templated collagen scaffolds, identifying collagen concentration and pH as the dominant determinants of pore size and connectivity and surfacing circular-dichroism peak intensity as an unexpectedly good structural predictor — a worked example of using interpretable ML to decide which fabrication variables are worth controlling. Golbabaei et al. (2024, ref #216) encoded polymer chemistry as SMILES strings alongside electrospinning process parameters and trained a voting classifier plus a neural-network regressor to predict electrospun-fiber properties, reaching a balanced accuracy of 0.95 for fiber-diameter classification and R² ≈ 0.72 for diameter regression, while candidly reporting that conductivity prediction was unreliable for lack of published data. Shin et al. (2025, ref #215) moved to the bioprinting nozzle: comparing several algorithms (a multilayer perceptron performed best) to predict cell-laden droplet size from print parameters generated on a high-throughput parallel droplet printer, finding dispensing time, nozzle size, and pressure the most influential controls and deploying the model behind a user-facing tool.

Nondestructive characterization of fabricated scaffolds

A complementary use of ML is reading back what was built. Serpe et al. (2025, ref #217) imaged decellularized spinach-leaf scaffolds — an edible, food-relevant scaffold candidate that the authors frame explicitly for cultivated meat — with grayscale ultrasound, then used a trainable random-forest pixel classifier to segment scaffold from background and seeded from unseeded regions, distinguishing fresh, decellularized, and cell-seeded leaves at p ≤ 0.001 and validating the readout against histology. The ML role is deliberately modest (an off-the-shelf pixel classifier rather than a bespoke model), but the paper points at nondestructive, in-bioreactor scaffold QC — confirming decellularization and cell coverage without destroying the construct — which any scaled cultivated-meat scaffold pipeline will need.

Tools and data

Biophysical simulation substrate. The slow simulations these surrogates accelerate belong to the same family as the open cell-mechanics and tissue-morphogenesis simulators catalogued in Software.md — CompuCell3D and PhysiCell (the Bioprocess Modeling & Scaling section) — which model the mechanics of cells interacting within a scaffold or mould.
Open scaffold data. The MLATE dataset released with ref #20 (1,171 (bio)printed scaffolds with printability, cell-response, and quality labels) is one of the most extensive openly available train-on artifacts for scaffold design and is bundled with its ML pipeline.

Open challenges for cell-ag

The surrogate-modeling results are encouraging but narrow. The mould-design work (refs #35, #34) is validated on phenomenological self-organization in tethered hydrogels rather than on edible cultivated-meat scaffolds, and both papers flag the move to true 3D geometries and to bioprinting as future work; the property-prediction work is similarly bounded — ref #19 by a small training library and imperfect accuracy, ref #20 by its deliberate exclusion of post-print features such as realized porosity, pore size, and fiber diameter; and the newer fabrication-route studies (refs #214, #215, #216) are each trained on a single lab’s collagen, electrospinning, or bioprinting setup. Across all of this work, the binding constraints for cell-ag are the same: training sets are small and idiosyncratic to each lab’s fabrication setup, there is no shared, standardized scaffold-design dataset analogous to the reference atlases other research areas enjoy, and almost none of the work yet ties a predicted geometry or recipe to edibility, food-safety, or scale-up cost. A surrogate trained on cultivated-meat-relevant cell types and food-grade scaffold materials, against standardized structural and textural endpoints, is the open opportunity here.