Hydrogels are water-rich polymer materials whose mechanical properties can be tuned through composition, crosslinking, processing and environmental conditions. In additive manufacturing, small changes in polymer concentration, crosslinker content, ionic environment or processing can lead to nonlinear changes in viscoelastic behavior, as expressed by storage and loss moduli. The design of such polymers often depends on iterative experiments, which can be time-consuming, expensive and difficult to generalize across material systems. Machine learning can learn relationships among structure, process and measured properties, while conditional generative models address the reverse design problem: identifying formulations that may satisfy a specified performance condition. Agentic Hydrogel brings these activities into one traceable sequence rather than introducing a new algorithm for any single task.

The system is organized around four specialized software components. A Planner Agent identifies the dataset, task type, regression or classification targets, and the conditioning and output variables needed for conditional generation. An Analysis Agent performs preprocessing, visualization and prediction. A Generation Agent trains generative models, produces probable material compositions and process parameters, and evaluates the synthetic data. An Orchestrator Agent coordinates the complete workflow, records states, manages communication among the agents and produces human-readable reports. Here, an “agent” is a software component that receives workflow information, applies task-specific rules, invokes assigned tools, updates the workflow state and passes structured outputs onward. Conventional algorithms perform data loading, schema discovery, model training, validation, evaluation and workflow management; a language model has no authority over these scientific decisions.

Human oversight is built into the command-line interface. Before execution, users must specify the dataset, task type, target variables for regression or classification, conditioning and output variables for generation, and any optional model settings. The framework then presents a summary listing the selected dataset, cleaned feature set, targets or conditions, analysis mode and planned computational steps. Users can approve, modify or terminate the workflow before processing begins. This confirmation is intended to prevent silent errors in consequential assignments, because treating a variable as an outcome or a target condition represents a scientific assumption that may not be inferable from data statistics alone. An optional large language model, available through a local deployment or cloud service, can summarize procedures and results, but it does not select data, define targets or control the analysis.

For regression and classification, Agentic Hydrogel uses adaptive multilayer perceptrons with rectified linear unit activation functions and optional dropout. Models are optimized with Adam, and validation-based early stopping is used during training. The base network in the reported experiments had two hidden layers with 64 neurons each, and the user-defined performance threshold was 0.60. If a model did not reach that threshold, network capacity could be increased and training repeated up to a user-set maximum, after which the model with the highest evaluation score was retained. This is a capacity-adjustment mechanism tied to a chosen threshold, not an exhaustive search for an optimal architecture. The authors selected multilayer perceptrons because they provide a unified nonlinear model for regression, classification and multi-output prediction within the same PyTorch workflow; they do not claim that this approach outperforms linear or tree-based models.

The polyacrylamide dataset contained 2,336 samples and nine variables and was used for multi-output regression of storage and loss moduli from material composition and additive-manufacturing parameters. In shuffled five-fold cross-validation, the multilayer perceptron achieved an R² of 0.782 ± 0.026, a mean absolute error of 247.079 ± 7.783 pascals, a root mean square error of 424.342 ± 16.857 pascals and a mean absolute percentage error of 19.25% ± 0.19%. The base network did not require capacity expansion. Because storage and loss moduli differed substantially in numerical scale, the authors caution against interpreting the metrics as a state-of-the-art benchmark or as evidence that the model learned the physical mechanisms governing rheology. The result instead demonstrates that the framework can configure, execute and document a pipeline for predicting continuous properties. The analysis was also repeated with five-fold cross-validation independent of the standard execution path, with preprocessing, early stopping and capacity decisions confined to the training folds to reduce data leakage.

Classification was evaluated on two granular hydrogel subsets, each containing 3,969 samples and ten variables. One subset used oscillation strain and the other oscillation stress as the differing experimental input, while both were used to predict a binary resuspension state. The two classes were approximately balanced. In stratified five-fold cross-validation, the oscillation-strain subset achieved an accuracy of 0.940 ± 0.017 and a macro-F1 score of 0.939 ± 0.017. The oscillation-stress subset achieved an accuracy of 0.937 ± 0.015 and a macro-F1 score of 0.937 ± 0.015. Neither base network required expansion. The closely matched results are consistent with the fact that the datasets share most formulations and rheological variables and differ in only one column. The authors present these scores as evidence that the workflow can perform discrete classification, not as state-of-the-art results or proof that the model captured the underlying intricacies of hydrogel formulation.

Conditional generation reverses the conventional direction of prediction: rather than estimating properties from a known recipe, the system searches for statistically probable composition and process variables associated with specified rheological conditions. Agentic Hydrogel implements a conditional variational autoencoder and a conditional denoising diffusion probabilistic model. The autoencoder receives paired output and conditioning vectors, maps them into a continuous latent distribution through mean and log-variance representations, samples latent values and decodes them with the conditions to reconstruct a material formulation. Its objective combines mean squared reconstruction error with a weighted Kullback–Leibler divergence term. The diffusion model uses a multilayer-perceptron denoising network that receives noisy output variables, rheological conditions and a sinusoidal embedding of the diffusion timestep. Its implementation adapts Gaussian diffusion and continuous-variable tabular diffusion ideas, but encoded discrete variables are handled numerically and restored during postprocessing rather than through a separate multinomial diffusion process.

Generated candidates were evaluated from distributional, structural and conditional perspectives. A Kolmogorov–Smirnov statistic was calculated separately for each output variable to compare real and synthetic marginal distributions, with lower values indicating closer agreement. The Frobenius norm of the difference between real and synthetic Pearson correlation matrices assessed preservation of multivariable relationships. Conditional fidelity was measured by finding the nearest experimental sample in scaled conditioning space for each generated sample and comparing the associated output variables with root mean square error and mean absolute error. A final range analysis flagged values outside the experimentally observed ranges. For the polyacrylamide data, the conditional autoencoder generated samples with a mean KS statistic of 0.406, a correlation Frobenius norm of 0.693, a conditional nearest-neighbor RMSE of 1.147 and a conditional nearest-neighbor MAE of 0.667 in standardized units. For the granular oscillation-strain subset, the diffusion model produced corresponding values of 0.291, 0.373, 1.041 and 0.651; for oscillation stress, they were 0.290, 0.390, 1.013 and 0.624. These measurements indicate partial agreement, but they do not establish a ranking of the two models because the hydrogel families, sample sizes, variables and architectures differ.

Postprocessing makes generated values more interpretable through optional clipping, non-negativity constraints, rounding of encoded categories and user-defined parameter bounds. These operations keep a candidate within the numerical domain represented by the source datasets, but they do not establish mechanical or chemical feasibility. Every execution produces model checkpoints, generated datasets, evaluation metrics, figures, workflow logs and Markdown reports, allowing the analysis to be inspected and documented. The study analyzed existing datasets and did not perform closed-loop fabrication or rheological testing of the generated candidates. It also did not provide systematic benchmarking of alternative predictive and generative models, uncertainty estimates, reinforcement learning or physics-informed constraints. Agentic Hydrogel therefore offers an auditable, human-bounded way to organize hydrogel machine-learning work, while leaving physical and chemical validation as a separate step before any computational candidate could be regarded as a working material.

Subject of Research: Human-supervised computational workflow for hydrogel analysis and conditional candidate generation

Article Title: Agentic hydrogel: a human-in-the-loop computational workflow for hydrogel analysis and conditional candidate generation

Article References: Mohammad, S., & Hossain, M. S. (2026). Agentic hydrogel: a human-in-the-loop computational workflow for hydrogel analysis and conditional candidate generation. Frontiers in Soft Matter, 6, 1936657. https://doi.org/10.3389/frsfm.2026.1936657

Image Credits: AI Generated

DOI: 10.3389/frsfm.2026.1936657

Keywords: hydrogels, agentic AI, machine learning, generative AI, rheology, conditional generation, materials informatics, additive manufacturing, polyacrylamide, granular hydrogels, Agentic, hydrogel