Physics-Informed Generative Design for Nonlinear Topology Optimization of Aerospace Bracket Assemblies Compliant with ASTM Standards

Abstract

Structural mounting brackets in aero-engine and satellite assemblies carry combined static, maneuver, and vibratory loads, and their in-service failures are dominated by fatigue crack initiation at geometric stress concentrations, local yielding under peak maneuver loads, and buckling of slender load paths. Weight is a first-order design driver, but structural topology optimization and especially its recent neural and generative variants tend to produce“organic” material layouts that either violate powder-bed-fusion manufacturing rules or develop stress hot-spots once the structure is evaluated under its true nonlinear (large-deformation and elastoplastic) response rather than the linear-elastic model used during optimization. Such designs are attractive on paper yet non-manufacturable or unsafe in service. This paper presents aphysics-informedgenerativedesign frameworkthatreparameterizesthematerialdensityfield of a Ti–6Al–4V bracket as an implicit neural field and trains it against a composite objective that (i) en forces the residual of the geometrically and materially nonlinear elastic equilibrium, (ii) penalizes viola tions of a minimum feature length scale and a self-supporting overhang angle, and (iii) aligns achievable feature sizes and mechanical property assumptions with ASTM Committee F42 specifications for laser powderbedfusion. The resulting bracket is validated against independent nonlinear finite element analysis in ANSYS/ABAQUS. Relative to a compliance-only generative baseline, the physics-informed design achieves a comparable mass reduction while keeping the peak von Mises stress below the material yield strength with a positive margin and eliminating unsupported overhangs and sub-tolerance members, so that the geometry is directly manufacturable without post-hoc repair. The framework offers a route to generative designs that are both lightweight and certifiable.