Adaptive Multi-Fidelity Physics-Informed Models for Accelerated Design Space Exploration

Abstract

High fidelity simulations are necessary to provide a correct modeling of complex physical systems, however their high cost of computation greatly restricts the ability to explore a design space efficiently especially in the under-data scenario. Current strategies usually have a hard time obtaining a balance between accuracy, computational speed and physical consistency causing poor exploration and slow convergence. To overcome these issues, this paper presents an adaptive multi-fidelity physics-informed model that combines low- and high-fidelity data using a single learning architecture. The framework uses physics-informed constraints to impose governing equations, and a smart adaptive sampling policy, which dynamically chooses informative samples according to uncertainty and exploration policies. The combination allows the efficient transfer of knowledge through levels of fidelity and maintains physical consistency. Experimental evidence indicates that the proposed method is more accurate in its predictions, much less costly in computational terms and converges faster than traditional single-fidelity methods and non-adaptive methods. The framework is proved effective on representative benchmark problems, and on real-life engineering problems, indicating its potential application in aerospace design optimization, fluid dynamics, and other computational science fields where scalable and efficient design space exploration is needed.