AI-driven soft materials design for superadhesives

Main text

The design of advanced soft materials such as hydrogels and elastomers represents one of the most persistent challenges in materials science.^1,2 Their macroscopic mechanical and functional properties arise from a complex interplay of molecular sequence, dynamic noncovalent interactions, and multi-scale network architectures spanning molecular, mesoscopic, and macroscopic regimens. In contrast to crystalline inorganic solids, whose structure-property relationships can often be deduced from well-defined atomic arrangements and symmetry principles, soft materials defy such straightforward characterization. Their behavior is governed by entropic elasticity, hydration forces, and stochastic sequence variations, leading to nonlinear, environmentally responsive phenomena that elude description by classical theoretical frameworks. For decades, researchers have thus relied on labor-intensive cycles of empirical trial and error, where incremental adjustments of polymer composition or processing conditions gradually refine performance but rarely achieve transformative breakthroughs.^3,4

A recent study in Nature unveils a transformative approach to overcome these long-standing limitations.⁵ By extracting functional motifs from tens of thousands of natural adhesive protein sequences and embedding them into statistically controlled copolymer networks, the authors established a biologically informed design space for programmable soft materials (Figure 1A). They further employed machine learning to efficiently navigate this space, enabling iterative cycles of prediction and synthesis. This data-driven methodology produced hydrogels exhibiting underwater adhesion strengths surpassing 1 MPa, nearly an order of magnitude greater than hydrogel in literature. Beyond setting a new performance benchmark, this work illustrates how deep learning can act not only as a predictive tool but as an active co-designer in materials creation, redefining the logic by which soft matter is engineered, optimized, and deployed.

Figure 1.

Machine learning-guided discovery of superadhesive hydrogels

(A) Workflow integrating sequence mining, formulation, and machine learning optimization.

(B) Schematic workflow for deriving bioinspired formulations by mining adhesive-protein sequences, encoding features, and computing relative compositions.

(C) Schematic of data-driven superior adhesive hydrogel preparation.

(D) Images of data-driven hydrogels showing tough adhesion.

The central advance lies in translating biological sequence patterns into synthetic polymer formulations. Drawing from a curated dataset containing 24,707 adhesive proteins gathered from the National Center for Biotechnology Information (NCBI) protein database using the keyword “adhesive protein,” the authors extracted sequence-derived features to guide the design of bioinspired copolymer blocks (Figures 1B and 1C). Using an ideal copolymerization framework, they statistically emulated the sequence heterogeneity of natural proteins under tightly controlled synthesis conditions. Monte Carlo simulations confirmed close agreement between the synthetic and natural systems in terms of block length distributions and pairwise residue frequencies, thereby bridging sequence-level statistics with macroscale properties. This physicochemical fidelity provided a high-quality dataset that proved essential for subsequent machine learning prediction. Notably, when synthesis deviated from ideal conditions and introduced compositional drift, adhesion strength dropped by over 70%, underscoring the critical importance of precise sequence-level control.

Based on a dataset of 180 hydrogels, the researchers implemented Gaussian process and random forest models within a sequential model-based optimization framework. Through three iterative prediction-validation cycles, the hydrogel library was expanded from the initial 180 bioinspired hydrogels to a final dataset comprising 341 hydrogels, ultimately enabling the discovery of superadhesive hydrogels with megapascal-level underwater adhesion (Figure 1D). The study further delivered quantitative insight into monomer-level contributions. Shapley analysis revealed that aromatic and hydrophobic monomers enhance adhesion by displacing interfacial water, while cationic monomers contribute synergistically with aromatics via electrostatic interactions, though excessive cation content induces swelling and weakens contact. In contrast, nucleophilic, amide, and acidic monomers generally diminish adhesion strength. These structure-activity relationships not only corroborate the model’s predictive accuracy but also furnish interpretable design principles for guiding future soft material design.

The significance of this work extends beyond a single material system. It establishes an integrated and scalable framework that combines bioinspiration, statistical polymer chemistry, and machine learning-guided optimization. This methodology is generalizable to a wide spectrum of functional soft matter, including conductive hydrogels, ionic elastomers, and bioinspired lubricating systems. Furthermore, it opens pathways for incorporating physics-aware machine learning with advanced polymerization techniques to achieve protein-like sequence precision in synthetic polymers. On the practical front, the achievement of megapascal-scale underwater adhesion approaches performance thresholds relevant to biomedical adhesives, marine engineering, and extreme-environment sealing, underscoring the translational potential of this computationally driven design paradigm.

Nevertheless, the diversity of available monomers remains narrow, as they are still largely derived from adhesive protein motifs. This limitation hinders the exploration of a broader design space, underscoring the need to develop systematic and expanded libraries that are not solely reliant on specific biological templates. Hydrogel synthesis and testing throughput, still relatively low, restrict dataset scale and complicates the training of more sophisticated deep learning architectures. Environmental robustness also poses a concern, as adhesive performance varies significantly across saline, seawater, and deionized water conditions. Overcoming these limitations requires the creation of integrated high-throughput platforms capable of rapidly generating large and well-annotated structure-property datasets. These platforms could be based on automated microfluidic or array-based reactors integrated with parallel mechanical and adhesion testing. This effort must be complemented by the development of broader and more systematically curated monomer libraries that intentionally incorporate non-biological building blocks, including conductive, zwitterionic, and aromatic chemistries.

This study marks a paradigm shift in soft materials design from intuition-led iteration toward AI-accelerated discovery. The significance of this superadhesive hydrogel lies not merely in the material itself but in proving that the closed-loop, data-driven logic of digital design can be directly deployed in soft matter. In principle, this same framework is readily extended to other functional polymers (e.g., for conductivity or energy storage) by replacing protein-inspired features with property-driven descriptors such as conductivity, redox stability, or ion-transport capability and integrating them into automated synthesis-testing loops. Looking ahead, this approach foreshadows a research environment where algorithms and experimentalists co-design materials through real-time, iterative prediction-synthesis loops. As AI emerges as a foundational language for material science, this study delivers one of its first and most compelling translations, turning computational insight into tangible, breakthrough matter.

Funding and acknowledgments

This study was supported by the National Natural Science Foundation of China (22508229 and 22478046) and the Shandong Provincial Natural Science Foundation (ZR2025QC354). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Declaration of interests

The authors declare no competing interests.

Published Online: December 27, 2025