Hybrid modeling framework for flow estimation in concrete three-dimensional printing using pump dynamics and machine learning

This work develops a control-oriented hybrid modeling framework for concrete three-dimensional (3D) printing that couples first-principles pump dynamics with machine-learning (ML) estimators of fresh slurry density ( ρ ) and viscosity ( μ ) to dynamically predict the slurry mass flow rate ( M ̇ ). A physics-guided feature set is screened using correlation, mutual information, and permutation importance and trained under grouped cross-validation across mixes (36 nominal mixes; 97 density and 195 rheology points). In the density predictor pipeline, a non-negative least-squares (NNLS)-stacked ensemble using Ridge, Extra Trees (ExT), and Random Forest (RF) outperforms single learners (coefficient of determination ( R 2 ) ≈ 0.73 ). For viscosity, high-capacity learners dominate; the best performance is achieved by a ridge-meta stack trained over the full base set ( R 2 ≈ 0.83 ). Unlike prior screw-extrusion flow-control work, validated on a single rig without a systematic sweep of mix compositions and reporting mean absolute error (MAE) of ≈ 6.7 % , this study targets progressive-cavity pumping and learns ρ and μ from a multi-mix dataset for dynamic, control-oriented flow prediction. Embedding the learned ρ and μ maps within the pump model yields simulated flow-rate curves that closely follow laboratory measurements across water-accelerator settings and motor speeds, with MAE on the order of 1–2% of mean flow. The predictor is interpretable and low-latency (milliseconds per evaluation) and integrates with Simulink for feedforward planning, soft sensing, and closed-loop use. The workflow provides a reproducible template, from data preparation and feature screening to model training and deployment, for hybrid modeling in large-format additive construction.