Machine learning-driven strategies for managing precision and efficiency in multiphysics simulations

Abstract

The accuracy and efficiency of numerical modeling in scientific simulations have long been significant challenges, making these processes costly and uncertain. Researchers often prioritize maximum precision in parameter settings to minimize errors, which increases computation and execution times. This cautious approach persists due to the difficulty in ensuring absolute accuracy and validating results, leading to prolonged simulation periods in pursuit of reliable outcomes. To address these challenges, this thesis presents three interconnected contributions that enhance the accuracy, efficiency, and robustness of scientific simulations. The first contribution introduces a novel machine learning-based approach for predicting results accuracy and execution times in multiphysics simulations. Traditional methods, such as numerical modelling and analytical calculations, are often computationally intensive and time-consuming. By employing regression algorithms, this research provides a black-box solution that predicts potential errors and computational demands, streamlining the exploration of simulation options. This method maintains the integrity of physical equations and ensures simulation convergence, reducing the trial-and-error efforts typically associated with conventional methodologies. The efficacy of this approach is demonstrated through applications in heat transfer and fluid-structure interaction problems. Building on this, the second contribution develops an advanced framework for optimizing design parameters in scientific simulations. During our research, we identified the need for efficient parameter optimization to enhance design processes without increasing computing times. The framework employs a machine learning-based surrogate model integrated with the Non-dominated Sorting Genetic Algorithm II (NSGA-II), enabling rapid and efficient navigation of complex design spaces. This approach balances objectives such as minimizing costs and maximizing directional motion. The effectiveness of the framework is illustrated through the optimization of a mooring system for a wave energy converter, significantly reducing the need for time-intensive simulations and facilitating informed decision-making in engineering design. Finally, recognizing the necessity of verifying simulation results to detect anomalies caused by improper parameter selection, the third contribution focuses on anomaly detection.

Unsupervised anomaly detection methods are employed to identify deviations that indicate potential convergence issues. This approach enhances the reliability and accuracy of simulations by detecting anomalies that signal convergence failures. Validation through heat transfer simulations demonstrates the algorithm’s capability to refine parameter selection and improve the robustness of simulation processes. Together, these contributions form a cohesive framework that significantly enhances scientific simulations. By integrating machine learning techniques for predictive accuracy, optimization, and anomaly detection, this thesis provides a comprehensive solution that improves the efficiency, reliability, and fidelity of simulation outcomes in complex engineering applications.

Date of Award	Jul 2025
Original language	English
Awarding Institution	Queen's University Belfast
Sponsors	Northern Ireland Department for the Economy
Supervisor	Hans Vandierendonck (Supervisor) & Adrian Murphy (Supervisor)