Xuesong Cai

Gas dispersion forecasting is a critical safety challenge for various engineering applications. For effective risk mitigation, it is essential to have a model capable of making accurate early predictions of the evolving hazard field right at the onset of a release. Although high-fidelity CFD models can provide detailed predictions, they are computationally expensive. In contrast, surrogate models can deliver near-real-time predictions, paving the way for digital twin frameworks that enhance safety design and operational decision-making for critical infrastructure. While existing neural surrogates have shown promise in related tasks, they often do not address the urgent need for direct, release-onset prediction only from parameters such as wind speed, leak intensity, and release angle, and cannot forecast the full spatio-temporal evolution of dispersion without relying on observed concentration frames. To address this problem, this study proposes the Modulated Fourier Neural Operator with Analytical Prior (MFNO-AP), whose central design combines an analytical Gaussian-plume prior with Feature-wise Linear Modulation (FiLM) on top of a recurrent Fourier Neural Operator (FNO) backbone. The proposed model was trained and validated on a floating liquefied natural gas example in both the 2D and 3D-slice benchmarks. Across both benchmarks, MFNO-AP provides accurate long-horizon dispersion forecasting with significant acceleration over the conventional numerical simulation. Controlled ablations further show a transferable mechanism split in which the Prior mainly regularizes the cold-start rollout entry and FiLM mainly stabilizes condition-dependent long-horizon evolution. These findings position MFNO-AP as a practical gas-dispersion surrogate for engineering applications.

The fire performance of Glass Fiber Reinforced Polymer (GFRP) reinforced concrete columns represents a critical knowledge gap in contemporary structural fire engineering. This study presents a comprehensive experimental and numerical investigation that compares the fire behavior of GFRP-reinforced and steel-reinforced concrete columns at elevated temperatures. Six full-scale circular concrete columns with a diameter of 380 mm and height of 3660 mm with different reinforcement configurations were subjected to theISO-834 standard fire exposure while maintaining constant axial loads at 30% and 50%load ratios. Experimental results revealed that under high axial load ratios (50%), GFRP and steel reinforced columns exhibited comparable fire resistance durations (34 minutes vs. 35 minutes). However, under moderate loading conditions (30%), steel-reinforced columns demonstrated superior performance with a 73% longer fire resistance duration(142 minutes vs. 82 minutes). Sequential thermal-mechanical finite element models were developed in ABAQUS and validated against experimental data, enabling parametric studies examining the effects of axial compression ratios (10%-60%), various concrete cover thicknesses (20-40 mm), and boundary conditions. A critical axial compression ratio threshold of 30% was identified for both reinforcement types, beyond which fire resistance deteriorated significantly. GFRP-reinforced columns exhibited enhanced thermal compatibility with the concrete matrix, characterized by reduced differential thermal expansion and more gradual, ductile failure progression. The numerical analysis confirmed that under axial compression ratio (≤30% axial ratio) and adequate concrete cover (≥30mm), GFRP-reinforced columns can achieve fire resistance performance equivalent to conventional steel systems while providing superior deformation control and progressive failure characteristics. These findings establish design criteria for GFRP reinforcement in fire-critical applications and support performance-based fire engineering methodologies for advanced composite structural systems.