Experiment-driven framework for predicting center-of-gravity shifts under liquid sloshing

Liquid sloshing in partially filled tanks induces time-varying loads and center-of-gravity (CG) shifts that can degrade stability and control. We present an experiment-driven framework that integrates laboratory measurements, computational fluid dynamics (CFD), and machine learning (ML) to accurately quantify and predict CG dynamics under sloshing excitation. The framework (i) reconstructs CG trajectories from synchronized load-cell and pressure measurements, (ii) validates a volume-of-fluid (VOF) OpenFOAM model against experiments using CG-centric metrics—peak-to-peak (P2P) amplitude, root-mean-square (RMS), and dominant frequency—with confidence intervals, and (iii) trains data-driven surrogate models that generalize within this jointly validated domain. To ensure transparent benchmarking, CFD validation is performed at water depths of 2, 4, and 6 cm (i.e., D/L ≈ {0.033, 0.067, 0.100}). Surrogate modeling (ML) is supported by a broader dataset covering depths from 3.0 to 7.2 cm, enabling interpolation across intermediate fill levels and testing generalization. Direct experiment–CFD overlays confirm agreement within 5–10% in RMS and peak-to-peak metrics. The ML surrogates, particularly the LSTM sequence model (two stacked layers, horizon H = 1000), achieve mean absolute errors around 1–2.5 mm across unseen fill levels, while simpler models (linear regression (LR), random forest (RF), and gradient boosting (GB)) remain competitive only in low-variability regimes. Results demonstrate that CG trajectory prediction and frequency content can be captured with high fidelity across fill levels, with ML surrogates providing substantial speedups for design-time trade studies. The surrogate' s limits relative to CFD are clarified, and representative overlays (experiment vs. CFD vs. surrogate) provide direct visual and quantitative comparison, enhancing clarity, transparency, and reproducibility.