Transfer learning enhanced heat transfer neural operators (HTNO) for efficient thermal analysis in concrete structures

Abstract

Heat transfer analysis in concrete structures is crucial in structural fire engineering for determination of temperature distribution and fire resistance of members. Due to the nature of nonlinear parabolic partial differential equations coupled with nonhomogeneous boundary conditions, traditional numerical solvers are often computationally expensive and limited in their ability to adapt to numerous varying fire scenarios. This study introduced an end-to-end framework for development of Heat Transfer Neural Operators (HTNOs) leveraging on Fourier Neural Operator architecture, tailored for structural fire engineering applications. These HTNOs could efficiently map both initial and boundary conditions (Dirichlet and Neumann types) to temperature profiles for various time domains with high fidelity. The developed HTNOs exhibited mean L2 relative errors ranging from 0.111 % to 0.840 % across an extensive test dataset. Once trained, these operators could provide accurate solutions significantly faster than traditional solvers by at least three orders of magnitude. This study also highlighted the critical role of incorporating domain-specific knowledge in the development of HTNOs. Moreover, transfer learning techniques were employed to enhance the learning process and reduce data dependency, notably improving the accuracy, generalisability and convergence speed. This study demonstrated the potential of HTNOs in revolutionising thermal analysis in structural fire engineering and providing a robust, rapid and highly accurate method for fire resistance analysis.