Multi-fidelity transfer learning for complex bund overtopping prediction with varying input dimensions

Abstract

This study introduces a novel multi-fidelity transfer learning framework designed to predict the complex bund overtopping fraction in catastrophic tank failure scenarios. The framework addresses the challenge of varying input dimensions between low-fidelity and high-fidelity datasets by effectively integrating these disparate sources of data. In this case, low-fidelity data, generated from empirical formulas based on simple bund configurations, is first used for initial model pre-training. The model is then fine-tuned using a smaller, high-fidelity dataset obtained through computational fluid dynamics simulations, which account for more complex bund configurations, including additional breakwater parameters. This approach enhances the model's predictive accuracy and generalization capability, particularly in scenarios with limited high-fidelity data. Case studies demonstrate that the transfer learning model outperforms traditional models trained solely on high-fidelity data, offering significant reductions in computational cost while maintaining robust predictive performance. The proposed framework not only advances the understanding and prediction of bund effectiveness but also provides a versatile tool applicable to a wide range of engineering problems involving multi-fidelity data.