Programmable physical models for understanding complex self-assemblies and designing functional materials

Predicting functions from material microscopic structures is fundamental to both understanding the complex behaviors of biological self-assemblies and designing functional materials with desired properties. However, the design space of current models is too vast to efficiently search for the desired functional features. Here we present an end-to-end differentiable physical model with novel features that can reproduce the functional behaviors of some biological self-assemblies, such as the dynamical growth of microtubules. Furthermore, our proposed model can directly optimize desired properties in high-dimensional parameter space (building block geometry, interaction strength, etc) using automatic differentiation. These results advance a substantial step towards understanding the complex biological behavior and inverse design of functional materials.