Introduction Compartment fires are governed by strongly coupled interactions among heat release, buoyancy-driven flow, turbulent mixing, heat transfer, and ventilation. Among these factors, ventilation plays a particularly important role because the size and configuration of openings control the supply of oxygen to the fire and the transport of hot gases from the enclosure. Changes in doorway or window geometry can therefore alter burning behavior, gas temperatures, flow patterns, and the transition between well-ventilated and under-ventilated fire regimes. These interactions are inherently transient and nonlinear, making computational modeling an important tool for systematically investigating compartment-fire behavior. Computational fluid dynamics (CFD) has become widely used for resolving the spatial and temporal development of fire-driven flows. The Fire Dynamics Simulator (FDS), developed by the National Institute of Standards and Technology, solves a form of the Navier–Stokes equations appropriate for low-speed, thermally driven flows, with particular emphasis on smoke and heat transport from fires [1]. Its large-eddy-simulation formulation enables detailed investigation of quantities such as gas temperature, velocity, species transport, and heat transfer and has made FDS an established computational platform for fire research and engineering applications [1,2]. Ventilation effects in compartment fires have received considerable attention because opening geometry can fundamentally alter the available combustion regimes. Lafdal et al. [3] performed more than 300 FDS simulations of naturally ventilated compartment fires with different enclosure scales, opening heights and widths, and heat-release rates. Their analysis identified well-ventilated, transitional, and under-ventilated regimes and demonstrated the strong relationship between opening characteristics, ventilation, and the resulting thermal response. More recently, Beshir et al. [4] combined full-scale experiments and numerical modeling to examine the effects of ventilation position, opening area, and opening aspect ratio on compartment-fire dynamics. Their results further demonstrated that changes in opening configuration influence internal fire behavior, heat fluxes, and the conditions associated with flashover. Collectively, these studies show that ventilation cannot always be represented by a simple monotonic relationship between opening size and a single thermal response variable. Although high-fidelity CFD provides detailed information about compartment-fire dynamics, repeated simulations can become computationally expensive when many combinations of operating and geometric parameters must be investigated. This limitation becomes particularly important when the high-fidelity model is embedded within an optimization procedure, because the simulator may need to be evaluated repeatedly as the optimizer explores the design space. Surrogate modeling provides an attractive alternative by learning an approximate relationship between simulation inputs and outputs from a finite set of high-fidelity calculations and subsequently evaluating that relationship at substantially lower computational cost. Machine-learning methods have consequently attracted increasing interest as surrogate models for computationally demanding fire simulations. Nguyen et al. [5], for example, developed machine-learning surrogate models for calibration of fire-source properties in FDS simulations of façade-fire tests. Their study considered both multiple linear regression and artificial neural networks and demonstrated that surrogate models could be coupled with optimization while selected solutions were subsequently verified using FDS. This work illustrates the potential of data-driven models to reduce the number of direct high-fidelity evaluations required in fire-engineering applications. Neural networks are particularly attractive for transient surrogate modeling because they can represent nonlinear mappings between multiple operating variables, time, and system responses. The use of a surrogate within an optimization algorithm, however, introduces an additional challenge. Good predictive performance over a training or validation dataset does not necessarily imply that the surrogate will reproduce the local behavior of the high-fidelity model in the region explored by the optimizer. Small local errors may change the predicted direction of improvement, and an optimizer may exploit inaccuracies in the surrogate rather than identify genuine improvement in the underlying high-fidelity system. Consequently, surrogate-assisted optimization requires a mechanism for controlling the region in which the surrogate is trusted and for verifying candidate solutions against the original high-fidelity model. Trust-region model-management methods provide a systematic framework for addressing this problem. Eason and Biegler [6] developed a trust-region filter appr