The Material That Learns

Evolution usually learns by changing genomes one generation at a time. This work asks whether ordinary material can do something closer to learning in real time: whether it can change its own stiffness map from repeated shaping targets and become better next time without rebuilding the whole thing by hand.

In Nature Physics (April 2026), Yao Du, Ryan van Mastrigt, Jonas Veenstra, and Corentin Coulais report a 2026 metamaterial that can learn shape changes by updating local stiffnesses, i.e., effectively the “weights” between unit cells, through examples. The authors describe three capabilities that are unusual together: it can forget old targets and learn new ones in sequence, it can learn non-reciprocal shape pairs, and it can learn multistable shapes that support separate stable configurations.

The details are worth reading for the weirdly software-like behavior: the material does not host a central controller but uses the mechanics of the network itself as memory. The paper’s supplementary material frames this as a contrastive scheme, where the structure is shown positive and negative examples and adjusts internal degrees of freedom. The result is not magic flexibility but a specific kind of physically constrained adaptation: a trained shape map that can later be re-called by a suitable input deformation.

Why should this matter outside the materials lab? Maybe narrowly. In the accompanying 2025 preprint, the same team frames it as “mechanical learning” and says this is a concrete step toward materials that move from fixed blueprints toward conditional competence. That has practical and philosophical relevance here: if infrastructure can carry both structure and learned response, then some “memory” no longer lives only in code or cloud logs. It can live in distributed stiffness, with all the maintenance, drift, and retraining costs that implies.

Sources: arXiv:2501.11958; Nature Physics (paywalled abstract); Nature Physics News & Views: Motorized paperclip learns functional reflexes.