A Multiphysics Neural Network Framework for Microdamage Evolution with A Shadow Plasticity Observer

Abstract

Microdamage evolution in ductile solids is influenced by plastic deformation, stress redistribution and material degradation. A computational framework is presented in which gradient damage mechanics is combined with a neural constitutive predictor. Plastic history is retained through a mechanically inactive auxiliary observer, termed shadow plasticity here. Its state is supplied to a multilayer perceptron without inserting the plastic tensor into the actual elastic deformation. A local damage increment is proposed by the network and converted into a correction of damage resistance. The original constitutive corrector and finite element balance equations are retained. The framework is examined using a single edge notched specimen under nonlinear loading. Complete reference and control histories are compared with an independently verified neural interval covering the first 205 s of a 1000 s loading program. Damage initiation, spatial localization and force response are assessed on explicitly matched supports. Early neural damage error is substantially reduced relative to the original damage only model. The completed controls demonstrate recovery of the reference damage region and structural response. The strongest complete agreement is obtained after the base damage resistances are aligned. Later neural propagation remains unverified. The findings provide a basis for physically structured prediction of microdamage and demonstrate the importance of separating constitutive interaction from material calibration.