Experimental investigation and machine learning prediction of steel fiber-aggregate mixture packing density for shield tunnel segments

A thorough understanding of material-process interactions and machine learning modeling strategies is essential for cost-effective precast production of steel fiber-reinforced tunnel segment concrete. This study designed tunnel segment-based vibration-densification tests to investigate the coupled effects of fiber geometry, aggregate combinations, and vibration compaction on mixture packing density. Using the experimental dataset, four machine learning models were developed to compare the mixture packing density predictabilities. The model's applicability was rigorously verified with extensive test and validation sets, and the significance of input material factors was assessed through feature importance analysis. Results indicate that mixture packing density first increases and then decreases with increasing coarse-aggregate content w_ₐ; optimal coarse-aggregate content w_ₐ = 30%–60% and aggregate-size ratio d₂/d₁ = 3.59–6.41 provide a quantitative basis for densification-skeleton design. Increasing steel fiber content and aspect ratio consistently reduces packing density, with hooked-end and waved fibers exhibiting greater adverse effects than straight-milled fibers due to hooking interactions—providing direct guidance for fiber selection in practice. Loose and dense packing densities have a stable linear relationship (R² > 0.94), enabling quick estimation of vibration-compacted density from loose measurements. By explicitly incorporating aggregate and fiber characteristics as inputs, the backpropagation neural network model demonstrates high predictive accuracy than others, while feature importance analysis identifies the most influential factors as coarse-aggregate content and steel fiber dosage. This study provides both experimental insights and an intelligent predictive tool to support performance-driven and cost-effective mixture design for steel fiber–reinforced concrete tunnel segments. © 2026 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license. http://creativecommons.org/licenses/by-nc-nd/4.0/