ShapeShift

: A Physics-Informed Transformation Framework for computationally efficient Finite Element Analysis (FEA) in Material Extrusion (MEX)

Abstract

Material Extrusion (MEX) produces polymer components with mechanical behaviour
governed by process-induced anisotropy, layer interfaces, and infill architecture. As a result,
mechanical properties cannot be predicted using homogeneous material assumptions. Existing
evaluation methods fall into four categories: commercial slicer-simulation packages with
proprietary modelling assumption cannot be inspected or modified by users, Classical
Laminated Theory (CLT) based on homogenised orthotropic layers cannot capture filament-
scale mechanism, filament-level geometry Finite Element Analysis (FEA) requiring explicit
toolpath reconstruction and fine-mesh, and implicit or reduced-order formulations that avoid
filament discretisation but do not resolve filament-scale stress transfer and interlayer damage
mechanisms. The limitations motivate the development of a method that preserves the
mechanical response captured by filament-level FEA while removing the need for explicit
filament discretisation in routine mechanical evaluation.
This thesis introduces ShapeShift, a physic-informed surrogate modelling framework designed
to reproduce the elastic mechanical response of filament-level FEA models while operating on
solid surrogate geometries. Filament-level simulation data are constructed using an Automated
Filament-Level Simulation Module, which reconstructs filament-level geometry within
Abaqus environment. The module employs a simplified rectangular filament cross-section and
perfect bonded interlayer interface to enable automated dataset generation with controlled
computational cost. Validation against experiments on PLA specimen yield elastic stiffness
prediction errors of 5.7-11.5% in tension and 0.8-18.9% in torsion.
These simulation-derived dataset are then used to train a ShapeShift framework that maps MEX
process parameter and part geometry to surrogate model representations. A hybrid learning
architecture combining graph-based learning and regression modelling enforces mechanical
consistency by matching surrogate elastic response to filament-level FEA predictions under
equivalent boundary conditions. For keyed-shaft geometries subjected to torsional loading, the
resulting surrogate models reproduce filament-level torsional stiffness within -2.9% to 8.2%
error while reducing finite-element counts by factor of 5-7 times reduction.
vi
Physics-informed surrogate modelling enables preservation of filament-level elastic
mechanical behaviour while substantially reducing model complexity and computational cost,
providing a validated and extensible approach for efficient mechanical evaluation of MEX-
manufactured parts.

Date of Award	17 Mar 2026
Original language	English
Awarding Institution	University of Bristol
Supervisor	Aydin Nassehi (Supervisor), Fengyuan Liu (Supervisor) & Suchandrima Das (Supervisor)