A hydrogel can look uniform from the outside while deforming very differently from one region to another. Once stiffer features are printed into a softer matrix, average stretch only tells part of the story. Local strain may begin to collect around edges, interfaces, or geometric constraints long before the overall specimen response looks unusual.
A study out of Pennsylvania State University followed that process in stereolithography-fabricated architected hydrogels. The researchers used a CellScale UniVert mechanical tester for hydrogel tensile testing while digital image correlation captured the changing surface strain field. Those experimental strain maps were then used to train a deep learning model to predict the next stage of deformation.
Why Hydrogel Tensile Testing Needs Full-Field Strain Mapping
Conventional hydrogel mechanical testing often produces a force-displacement or stress-strain curve. That is useful for comparing stiffness, strength, extensibility, and failure. It can still miss where deformation is happening inside a mechanically heterogeneous sample.
A soft matrix containing a stiffer printed framework may experience very different local stretches from one region to the next, even though the entire specimen is pulled by the same grips. A single global strain value averages those differences together.
This is where digital image correlation becomes useful. DIC tracks a visible surface pattern across a sequence of images and calculates displacement and strain over the specimen surface. Instead of one strain value, it produces a full-field strain map showing where the sample is stretching more, where it is constrained, and how those regions change during loading.
In this study, the question was not only how much the hydrogel stretched. The researchers wanted to see how strain localization appeared and moved through the material as hydrogel tensile testing progressed.
Printing Mechanical Heterogeneity into an Architected Hydrogel
The samples were thin, two-dimensional hydrogel films fabricated through grayscale stereolithography. The resin was based on N-isopropylacrylamide, or NIPAm, and the printing process varied ultraviolet exposure across the structure.
Longer exposure created more highly crosslinked and stiffer regions. Shorter exposure left the surrounding matrix softer. This allowed the researchers to print a re-entrant honeycomb framework inside a continuous hydrogel film rather than making the whole sample mechanically uniform.
The geometry was chosen to produce a more complicated deformation field. Re-entrant honeycomb patterns are associated with auxetic behaviour, while the stiffness mismatch between the framework and matrix redirects load around the printed structure.
Similar spatial control is being explored in stimuli-responsive hydrogels, soft robotics materials, and other printed soft systems. It also creates a practical testing problem. The response of soft membranes and thin films can depend heavily on local geometry, so a bulk modulus or average strain may not be enough.
How the UniVert Was Used for Hydrogel Tensile Testing
The researchers performed uniaxial tensile tests using the UniVert. Each hydrogel film had a 20 mm gauge length and was stretched at 1.2 mm/min. Sample thickness was measured before testing so the mechanical response could be interpreted using the specimen geometry.
The test was carried out in a water bath maintained at 35 °C. That temperature kept the NIPAm-based hydrogel in its deswollen, or shrunken, state. This is one reason hydrated and temperature-controlled testing matters for responsive hydrogels. A change in temperature or hydration can alter the material before the tensile response is measured.
For DIC, the researchers dispersed amine-functionalized graphene oxide particles through the resin. These particles formed a visible marker pattern that could be tracked as the material deformed. A 12 MP camera recorded the specimen during loading, and GOM Correlate software was used to calculate full-field displacement and strain maps.
How can digital image correlation be combined with tensile testing?
The mechanical tester controls the loading history, while the camera records how the specimen surface moves at each stage. Built-in DIC software compares those images with the undeformed reference frame and calculates local strain across the field of view. In practice, uniaxial tensile testing can provide both the overall mechanical response and a spatial picture of where deformation is concentrating.
For this experiment, hydrogel tensile testing produced a sequence of strain maps rather than a single image at the end. That time component became the input for the machine learning model.
Can Machine Learning Predict Strain Localization in Hydrogels?
The researchers collected 120 time-resolved strain maps from the DIC measurements. They removed the surrounding background and colour bars, converted the maps to normalized grayscale images, and resized them to 256 by 256 pixels.
The model was given three consecutive strain maps and asked to predict the next one. It combined convolutional image processing with convolutional long short-term memory layers, usually shortened to ConvLSTM. The convolutional layers extracted spatial features, while the recurrent part followed how those features changed from one frame to the next.
That distinction matters. The model was not predicting an entire hydrogel tensile testing experiment from the material formulation alone. It was making a next-frame prediction from three immediately preceding experimental strain fields.
The measured maps changed from a relatively diffuse strain distribution to more pronounced localization along pathways related to the printed honeycomb architecture. The predictions followed much of the same pattern. By the later stage, the horizontal bands visible in the experimental data also appeared in the predicted field, although the predictions were somewhat smoother.
The authors report a validation mean squared error of 5.49 × 10⁻³. More useful than that single number, though, is where the errors appeared. The largest discrepancies tended to sit around regions where strain changed abruptly over a short distance.
High Strain Gradients Were Harder to Predict
One question the paper addresses is whether prediction error simply rises as strain becomes larger. It appears that the answer is no, or at least not in a straightforward way.
Panel A of the figure below compares normalized strain with prediction error. The red mean-error curve remains fairly low across most of the strain range, showing only a weak overall dependence on strain magnitude.
Panel B tells a clearer story. As the spatial gradient of normalized strain increased, mean prediction error also increased. A high strain gradient means neighbouring regions are deforming very differently. Those sharp transitions often occur near interfaces, structural boundaries, and localized deformation pathways.
A model can reproduce broad deformation patterns while still losing accuracy at the locations where strain localization is most mechanically interesting. Smoother predictions are not automatically wrong, but they may soften local peaks and boundaries that matter for damage, failure, or interface design.
What This Adds to Hydrogel Mechanical Testing
The study shows how hydrogel tensile testing can move beyond one stress-strain curve. The UniVert provided controlled tensile loading in a hydrated, temperature-controlled environment. DIC converted surface motion into evolving strain fields. The neural network then tested whether those experimentally measured patterns could be forecast one step ahead.
There are limits. The dataset contained 120 strain maps, but the paper does not clearly state how many independently printed specimens contributed to them. The model used an 80:20 training-validation split, with no separate held-out test set reported. It also focused on one architected hydrogel design, so it is not yet clear how the same trained model would perform on a new geometry, formulation, or loading rate.
Still, the workflow is useful. It shows how measured strain fields can become training data for material modelling, while keeping the prediction tied to a real mechanical experiment. For researchers involved in soft materials testing, that connection between prediction and measurement may be the part worth watching.
Related research has used mechanical testing to examine hydrogel fibres, stimulus-driven shape changes in hydrogels, and the validation of AI-based elastography measurements.
Citation
Hongtao SunSteven TraczikVictoria Devine-Ducharme; Machine Learning-Enabled Prediction of Spatiotemporal Strain Evolution in Stereolithography-Fabricated Architected Hydrogels. ACS Appl. Eng. Mater. 22 May 2026; 4 (5): 2718–2725. https://doi.org/10.1021/acsaenm.6c00433
About the UniVert for Hydrogel Tensile Testing
The UniVert is used to mechanically test soft tissues, hydrogels, polymers, scaffolds, and other biomaterials in tension, compression, and related loading modes. In a hydrogel tensile testing workflow, specimens can be loaded in a fluid bath while force and displacement are recorded throughout the test.
Imaging can be incorporated when the research question depends on local deformation rather than only crosshead motion. That may involve DIC, region-of-interest tracking, or simply watching how a printed feature changes shape during loading. The tester provides the controlled mechanical input. The imaging and analysis show what the material did with it.
For architected hydrogels, that combination can be especially useful. Internal geometry may influence where strain first appears, how localization develops, and where failure eventually begins. A global curve still matters, but it does not have to be the only output from the experiment.
