Edgar Torres

Edgar Torres is a Researcher at the Institute for Artificial Intelligence, University of Stuttgart, specializing in data-efficient machine learning methods for scientific simulations. His work focuses on solving partial differential equations through physics-informed neural networks, neural operators, and adaptive learning techniques including transfer learning and meta-learning, with applications in computational fluid dynamics and solid mechanics.

Research Focus: His primary research addresses computational bottlenecks in scientific simulations by developing adaptive ML models that minimize data requirements. Key interests include Physics-informed Neural Networks (PINNs), Neural Operators, Transfer Learning, Meta-learning, Few-shot Learning, Computational Fluid Dynamics (CFD), Solid Mechanics, Surrogate Models, and Neural Fields. His approach integrates sparse data observations with prior system knowledge to enhance model generalization across simulation tasks.

Publications Insight: His 2025 survey publication analyzes adaptive techniques in physics-informed neural networks, revealing how transfer learning and meta-learning frameworks enable rapid adaptation to new PDE systems with minimal data. This work identifies critical pathways for improving computational efficiency in scientific machine learning, particularly for design optimization and real-time assessment scenarios where traditional simulations are prohibitively expensive.

Academic Engagement: Torres serves as Teacher Assistant for the Deep Learning for the Sciences seminar (WS-2024/25) and actively supervises Bachelor's and Master's theses. Prospective students must submit CVs and Transcripts of Records; he encourages proposals aligned with his research areas when formal openings are unavailable.

Research Context: As part of the Machine Learning for Simulation Science group, he contributes to the Institute's mission of bridging AI and scientific computing. His current PhD project develops adaptive surrogate models that overcome limitations of reduced-order modeling through meta-learning frameworks, targeting applications requiring rapid multi-configuration evaluations.