مادة ذات ذاكرة: الشبكات العصبية بدلاً من الجزيئات سريع

الملخص الأصلي

كما كتبه المؤلفون، بالإنجليزية 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.