Data Efficient AI for Reduced Order Models: Meta-Learning the Optimal Training Dataset

Machine learning (ML) models often lack physical intuition and rely on large datasets, without a systematic and scalable framework for creating more informative training data. This thesis tackles this issue by proposing a closed-loop Artificial Intelligence (AI) approach that integrates learning with optimization on top of existing AI modules, aimed at building reduced-order models (ROMs) through targeted data selection. This pipeline focuses on performance optimization of ML algorithms while reducing the time needed for data collection, thereby improving data efficiency. This enhancement directly translates into reduced experimental costs, accelerated development timelines, and minimized wear on physical test systems. To enhance system interpretability and trust, the framework integrates explainability studies that leverage a large language model. In a practical application with real-world vehicle data, this enabled the identification of eight critical clusters of maneuvers, consequently reducing the required testing duration from 4 hours to merely 20 minutes. Evaluated across challenging real-world and synthetic domains, including vehicle and electric motor dynamics, the closed-loop optimizer consistently outperformed unguided random sampling baselines. With guided iterations, the optimizer converged to a smaller training dataset that produced a more accurate model, outperforming the average of thousands of unguided, random attempts. This highly efficient search for informative data thereby enables the creation of high-fidelity surrogate models using over 90% less training data. These results establish this integrated approach as a robust and scalable solution for developing data-efficient, trustworthy ROMs for complex engineering systems and can be extended to non-ROM based applications.