Metaheuristic-Driven Ensemble Learning for Robust Fracture Energy Prediction in FDM-Fabricated PLA Components

TL;DR

This paper introduces a machine learning approach to predict the fracture energy of 3D-printed PLA parts, improving accuracy by combining multiple models with optimization techniques.

Contribution

A novel metaheuristic-driven ensemble learning framework is proposed for enhanced fracture energy prediction in FDM-printed PLA components.

Findings

Impact toughness of FDM-printed PLA was evaluated using Taguchi DoE methodology.

Hybrid models with nature-inspired algorithms improved prediction accuracy significantly.

The ensemble method achieved a MAPE of 5.0847%, a 37.3% improvement over individual models.

Abstract

Additive manufacturing (AM) has reshaped production methodologies by enabling the fabrication of complex geometries for high-performance applications. As a leading AM technique, Fused Deposition Modeling (FDM) is widely used for its versatility. However, the structural reliability of FDM-printed parts is fundamentally dictated by their mechanical performance, where impact toughness functions as a critical benchmark across demanding industrial environments. Polylactic acid (PLA) has distinguished itself as a premier biodegradable polymer, favored for its superior stiffness and processability. Nevertheless, the inherent brittleness and anisotropic behavior of FDM-printed PLA pose significant challenges, necessitating investigation of their fracture mechanics. This study firstly evaluates the impact toughness of FDM-processed PLA Izod specimens using impact tests, structured within a Taguchi design of experiments (DoE) methodology. An L27 orthogonal array was employed to investigate the influence of manufacturing parameters on impact behavior and fracture energy. Then, to achieve high-fidelity predictions from experimental data, the parametric effects were systematically investigated through an advanced machine learning framework. In the first stage, optimal prediction models were identified by evaluating five mathematical formulations hybridized with five nature-inspired optimization algorithms (GWO, SMA, GSA, FPA, and KH) across nine dataset combinations. In the second stage, these best-performing models were integrated into a metaheuristic ensemble using the GWO to perform a weighted aggregation. This hybrid ensemble methodology significantly enhanced predictive accuracy, achieving a Mean Absolute Percentage Error (MAPE) of 5.0847%, which represents a 37.3% relative improvement over the best individual base model.

Figures