# Data-Driven Optimization and Modelling of the Gap Bridgeability Performance of Multi-Pin Friction Stir Welded EN AW 7020-T651 Joints

## TL;DR

This paper presents a data-driven approach to optimize and predict the quality of friction stir welded aluminum joints under weld line gaps.

## Contribution

A novel data-driven framework with a new metric (WAP) and machine learning models to assess gap bridgeability in FSW joints.

## Key findings

- A Random Forest model achieved 92.5% accuracy in classifying weld acceptability.
- Interaction terms welding speed × gap size and rotational speed × gap size were most influential for weld quality.
- A multi-pin tool improved weld soundness under gap conditions.

## Abstract

Friction stir welding (FSW) of high-strength aluminum alloys, including EN AW 7020-T651, encounters significant challenges under weld line gap conditions, leading to compromised joint integrity. This study develops a predictive, data-driven framework to assess and optimize the gap bridgeability performance of FSW joints with weld line gaps ranging from 0 to 4 mm in 2 mm thick plates. A structured experimental matrix was implemented, systematically varying rotational speed, welding speed, axial force, and tool shoulder diameter. To promote stable material flow and consistent weld quality under varying gap conditions, a multi-pin tool was employed throughout the welding trials. This configuration supported defect-free weld formation across a broad process window and contributed to improved weld soundness under gap conditions. Weld quality was evaluated using a comprehensive, multi-criteria approach that required (i) defect-free joints verified by visual and cross-sectional (metallographic) inspection, (ii) an ultimate tensile strength of at least 230 MPa, and (iii) a novel metric termed weak area percentage (WAP). Derived from micro-hardness mapping, WAP quantified the proportion of the heat-affected zone (HAZ) exhibiting hardness below 96 HV, providing a more robust and spatially sensitive measure of mechanical integrity than conventional average hardness values. Two machine learning models, Logistic Regression and Random Forest, were trained to classify weld acceptability. The Random Forest model demonstrated superior performance, achieving 92.5% classification accuracy and an F1-score of 0.90. Feature importance analysis identified the interaction terms “welding speed × gap size” and “rotational speed × gap size” as the most influential predictors of weld quality.