Data-driven machine learning meta-analysis of process–property relationships in polymer additive manufacturing: A case study on FFF-printed PEEK

Abstract

Polymers are widely used in 3D printing technologies such as fused filament fabrication (FFF), where process parameters significantly influence the structure and performance of parts. Establishing process–property relationships is challenging due to the high dimensionality of the process space, limited and sparse data, and heterogeneity in the literature. We present a literature-driven meta-analysis focused on process–property measurements of FFF parts made from polyether ether ketone (PEEK). A dataset of 188 data points from 12 studies is compiled to examine the effects of nine printing parameters on ultimate tensile strength (UTS) of printed coupons. We train neural-network models and apply global sensitivity analysis (Sobol indices). To improve robustness and extract higher-order patterns, we adopt an ensemble strategy, training over 1,000 networks and selecting the top five based on performance metrics. This explainable AI (XAI) workflow quantifies both first-order and total-effect contributions; results show that while UTS is sensitive to parameters such as infill percentage and printing speed, sensitivity is significantly amplified when accounting for coupled interactions among parameters, elevating the influence of variables like nozzle temperature. The study provides a transferable, XAI framework that quantifies uncertainty and reveals higher-order interactions from literature-derived datasets, enabling data-driven process optimization in additive manufacturing.