A Heterogeneous Multi-Output Stacked Learning Framework for Mechanical Property Prediction of FDM-Printed ASA: Experimental Validation

Abstract

Accurate prediction of the mechanical performance of polymer components fabricated by fused deposition modelling (FDM) remains challenging owing to the complex nonlinear relationships between process parameters and material properties, limiting reliable process planning and broader industrial adoption of polymer additive manufacturing. This study develops and experimentally validates a heterogeneous multi-output stacked ensemble learning framework for the simultaneous prediction of tensile strength, flexural strength, compressive strength, Rockwell hardness, and Charpy impact strength of acrylonitrile styrene acrylate (ASA), a high-performance engineering thermoplastic with excellent weatherability and ultraviolet resistance that remains comparatively underexplored in data-driven FDM research. A Definitive Screening Design (DSD) was employed to investigate eight critical process parameters: extrusion temperature (ET), bed temperature (BT), infill density (ID), layer height (LH), print speed (PS), raster angle (RA), build orientation (BO), and cooling fan speed (CFS). Multiple supervised learning algorithms were systematically benchmarked, and the highest-performing complementary models were integrated into a heterogeneous stacked ensemble for simultaneous multi-output prediction. The proposed framework achieved an overall R2 of 0.9943 with an overall RMSE of 0.9758, while the individual prediction models attained R2 values ranging from 0.9898 to 0.9967. Beyond improving predictive accuracy, the proposed AI-assisted framework provides a data-driven basis for mechanical-property prediction and establishes a surrogate modelling framework that may subsequently be coupled with dedicated optimization or decision-making methods.