Memory Material: Neural Networks Replace Molecules express
Viscoelastic materials — polymers that, like dough, retain a memory of every touch: their response to impact depends on their deformation history. Maxwell and Boltzmann provided the mathematical description long ago. But for computer simulations, engineers had to model each molecule at every step — an impossible task even for supercomputers.
Now a neural network trained on thousands of virtual experiments comes to the rescue. It captures patterns without iterating over billions of particles, much like a skilled baker who, just by looking at dough, understands its entire history. The network's internal parameters accumulate information about the past, much like entropy gathers data on molecular disorder.
Tests show the model accurately reproduces the behavior of a carbon-based polymer, polyurea, including under cyclic bending, impact, and even considering heating. A remarkable fact: a layer of this polyurea just a few millimeters thick can shield a concrete wall from a blast wave. The neural net works thousands of times faster than classical models and can be easily integrated into any engineering software as a universal template — analogous to the standard model for materials physics.
Original abstract
as written by the authors arXiv:2603.27430
We present a neural-operator-accelerated concurrent multiscale framework that couples atomistic simulations with continuum finite-element analysis for history-dependent materials, thereby making atomistic-continuum multiscale simulations of viscoelastic materials tractable. The approach replaces direct molecular dynamics (MD) evaluation of the constitutive response with a Recurrent Neural Operator (RNO) surrogate trained on atomistic simulations. The surrogate learns the strain-history-to-stress operator from molecular dynamics simulations and provides a discretization-independent approximation of the atomistic constitutive mapping, enabling efficient evaluation of stresses and latent internal variables at each quadrature point. The framework is implemented within an explicit finite-element solver, where the constitutive update reduces to inexpensive operator evaluations rather than repeated MD solves. Memory effects are represented through learned internal states, and transfer learning across temperature enables the surrogate to capture thermally dependent viscoelastic behavior. The method is assessed using polyurea through cyclic loading, Taylor impact, and plate impact simulations and compared with an experimentally calibrated viscoelastic polyurea model and a Johnson-Cook model. The neural-operator surrogate reproduces correct viscoelastic response while enabling atomistically informed dynamic simulations at scales that are not tractable with direct MD-FEM coupling.
Polyurea is so viscoelastic that a coating just a few millimeters thick can protect a concrete wall from a blast wave.
This technology evokes sci-fi ‘smart’ materials that adapt to every impact, like Iron Man’s suit.