Integrating machine learning for the optimization of polyacrylamide/alginate hydrogel

Abstract

Hydrogels are highly promising due to their soft texture and excellent biocompatibility. However, the designation and optimization of hydrogels involve numerous experimental parameters, posing challenges in achieving rapid optimization through conventional experimental methods. In this study, we leverage machine learning algorithms to optimize a dual-network hydrogel based on a blend of acrylamide (AM) and alginate, targeting applications in flexible electronics. By treating the concentrations of components as experimental parameters and utilizing five material properties as evaluation criteria, we conduct a comprehensive property assessment of the material using a linear weighting method. Subsequently, we design a series of experimental plans using the Bayesian optimization algorithm and validate them experimentally. Through iterative refinement, we optimize the experimental parameters, resulting in a hydrogel with superior overall properties, including heightened strain sensitivity and flexibility. Leveraging the available experimental data, we employ a classification algorithm to separate the cutoff data. The feature importance identified by the classification model highlights the pronounced impact of AM, ammonium persulfate, and N,N-methylene on the classification outcomes. Additionally, we develop a regression model and demonstrate its utility in predicting and analyzing the relationship between experimental parameters and hydrogel properties through experimental validation.

Keywords: Bayesian optimization; alginate/polyacrylamide hydrogel; flexible electronics; machine learning; stretchability.

Graphical abstract

Figure 1.

(A) Design principle of interpenetrating double network hydrogel based on polyacrylamide and alginate. (B, C) SEM images of hydrogels at different magnification.

Figure 2.

Workflow chart of Bayesian optimization process for optimizing dual network hydrogels.

Figure 3.

The Y value change curve during the optimization process (A). A comparison of stress-strain (B) and GF curves (C) before and after optimization.

Figure 4.

Hydrogel as skin sensor for human motion detection. (A) Finger bending. (B) Wrist flex. (C) Knee movement. (D) Elbow movement. (E) Vocal cords. (F) Throat swallow.

Figure 5.

(A) F1 scores of different classification algorithms. (B) Learning curve of RF algorithm. (C) Learning curve of XGB algorithm. (D) Radar plot depicting the importance of feature in the classification process performed by an RF model. (E) The distribution of data points within the three-feature space encompassing MBA, AM and APS.

Figure 6.

A Comparison of R² and MAE of different regression algorithms when training different models on resistivity (A), elongation (B) and fracture energy (C). A comparison between the predicted and measured resistivity with varied concentration of NaCl (D), AM (E) and SA (F).