Original title: Machine learned designs of functional colloidal foldamers
Ryan van Mastrigt, Zorana Zeravcic · August 2026
Scientists often try to make self-assembling materials fold into a single fixed shape, treating any other shape as a mistake. In this study, researchers used reinforcement learning to design the folding routes instead of just the final shapes. They folded chains of DNA-coated droplets into rigid two-dimensional shapes. The algorithm revealed two kinds of paths. In downhill paths, bonds only form. In detour paths, bonds break and re-form to reach the goal, which is the only route available for certain shapes. Simplifying these paths by their interactions creates real-world laboratory protocols. Some protocols make one shape, while others create multiple shapes that can cycle between each other or assemble into larger superstructuresLarger structures formed from several smaller shapes assembling together., allowing materials to reconfigure and self-repair.
Imagine you are building a house out of blocks. You usually follow a plan from bottom to top, adding one block at a time. Now suppose that to get a particular roof shape, you first have to remove some blocks you just put in, then place them differently. That would seem like a waste, but it is the only way to finish. In this paper, the droplets act like blocks, and the folding process is the building sequence. The researchers found that breaking and remaking connections is not a mistake — it is a design choice that enables shapes and behaviors no straight path can achieve.
Today's programmable materials are built to one fixed shape — anything else is a defect. This work shows that by designing the folding process itself, droplets can be made to switch between multiple shapes or assemble into larger structures. That opens the door to colloidal materials that rearrange, repair, or adapt on demand, useful for everything from medical devices to smart surfaces.
The abstract does not report experimental validation of the coarse-grained protocols. The work is based on model droplet chains, so real-world implementation may face additional challenges. The approach is demonstrated for a specific system and may not generalize to more complex materials.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Machine learned designs of functional colloidal foldamers — Ryan van Mastrigt, Zorana Zeravcic
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.