International Surgery and Case Reports

  • High-Fidelity-Guided Surrogate Optimization of Compartment Fire Dynamics Using Trust-Region Model Management

    Case Report Pdf View 👁️

    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 approach for optimization problems containing computationally expensive black-box components represented by reduced models. Subsequent work extended the strategy and demonstrated its application to hybrid glass-box/black-box optimization problems [7]. More recently, Biegler [8] reviewed the trust-region filter strategy for optimization with surrogate models and emphasized the combination of computationally inexpensive surrogate optimization with intermittent evaluation of the high-fidelity truth model. Related recent developments have continued to investigate more efficient reduced-model construction and sampling strategies for computationally expensive black-box optimization [9]. These approaches provide an important conceptual basis for combining inexpensive surrogate calculations with selective high-fidelity evaluations rather than relying on the surrogate throughout the optimization process.

    Despite advances in both computational fire modeling and surrogate-based optimization, comparatively limited attention has been directed toward integrating transient FDS simulations, neural-network surrogate modeling, adaptive high-fidelity evaluation, and trust-region model management specifically for compartment-fire ventilation optimization. Previous fire studies have demonstrated the importance of ventilation and the usefulness of FDS for resolving compartment-fire behavior [3,4], while machine-learning studies have demonstrated the feasibility of replacing repeated FDS evaluations with computationally efficient surrogate models for calibration and related tasks [5]. Trust-region methods, meanwhile, provide a rigorous foundation for managing discrepancies between surrogate and high-fidelity models [6–9]. Bringing these elements together provides an opportunity to examine not only whether a surrogate can approximate fire simulations, but also whether its local predictions are sufficiently reliable to support optimization decisions.

    In this study, we develop an integrated computational framework combining high-fidelity FDS simulations, neural-network surrogate modeling, and trust-region model management for transient compartment-fire analysis and ventilation optimization. High-fidelity simulations are first used to investigate the effects of fire intensity and doorway width on temperatures measured at multiple elevations within a ventilated compartment. The complete transient temperature histories are then used to construct and evaluate a neural-network surrogate with heat-release rate per unit area, doorway width, and simulation time as inputs. Doorway width is subsequently treated as the design variable in a surrogate-assisted optimization framework, while additional FDS simulations are used to evaluate proposed candidates, quantify local surrogate discrepancies, and guide adaptive surrogate refinement. A grid-sensitivity assessment using three progressively refined computational meshes is also performed to characterize the numerical-resolution dependence of the predicted temperatures.

    The principal contribution of this work is the integration of transient fire simulation, data-driven surrogate modeling, and high-fidelity-guided trust-region model management within a unified computational workflow. Rather than assuming that a surrogate with acceptable overall predictive performance remains reliable during optimization, the framework explicitly evaluates local agreement between surrogate and FDS predictions before accepting optimization-driven conclusions. The resulting analysis therefore provides both an assessment of the potential of neural-network surrogates for reducing the computational burden of fire-dynamics optimization and a demonstration of the importance of high-fidelity model management when localized surrogate errors influence the predicted direction of improvement.

    DISCUSSION

    The results demonstrate the importance of combining surrogate modeling with high-fidelity verification when optimizing transient compartment-fire dynamics. Although the neural-network surrogate provided a computationally efficient representation of the FDS simulations, discrepancies remained in both the magnitude and local trend of the peak-temperature objective. In particular, surrogate-predicted improvements at some doorway widths were not confirmed by the corresponding FDS simulations. These findings show that surrogate accuracy measured over a broader dataset does not necessarily guarantee sufficient local accuracy in the region explored by the optimization algorithm. The trust-region model-management framework therefore provided an important mechanism for identifying disagreement between the surrogate and high-fidelity models before accepting candidate solutions.

    The high-fidelity simulations also revealed a non-monotonic relationship between doorway width and peak compartment temperature. At an HRRPUA of 400 kW/m⊃2;, the peak TEMP_LOW decreased as the doorway width increased from 0.4 to 0.8 m, with the lowest value among the original ventilation cases occurring at 0.8 m. Further increases in doorway width did not produce an additional reduction in the peak temperature. This behavior indicates that the thermal response cannot be represented adequately by assuming a simple monotonic relationship between ventilation and peak temperature. The result also highlights the value of evaluating the ventilation parameter over multiple high-fidelity simulations rather than inferring an optimum from a limited local trend.

    The adaptive refinement results further illustrate the distinction between improving surrogate accuracy and achieving optimization-relevant fidelity. Incorporating the additional high-fidelity trajectories improved the surrogate representation in portions of the locally sampled region, particularly the predicted trend between doorway widths of 0.900 and 0.967 m. However, substantial errors remained in the predicted peak-temperature magnitude, and the refined surrogate still failed to reproduce the high-fidelity trend between 0.800 and 0.900 m. Thus, additional training data improved aspects of the local surrogate response but did not uniformly resolve the model discrepancy. These results support the continued use of high-fidelity verification when surrogate predictions are used to guide optimization decisions.

    Within the high-fidelity-supported interval of 0.800–1.000 m, the model-management analysis identified a doorway width of 0.800 m as the best evaluated solution, with a peak TEMP_LOW of 1090 °C. This value represents a reduction of approximately 1.8% relative to the 1110 °C peak obtained at the initial 1.000 m doorway width. Importantly, this result should not be interpreted as a global optimum. Rather, it represents the best solution supported by the available high-fidelity evaluations and the local model-management analysis. This distinction is important because the surrogate model exhibited significant discrepancies in the optimization region, making extrapolation beyond the high-fidelity-supported interval inappropriate.

    More broadly, the results demonstrate the value of model management for fire-dynamics optimization when high-fidelity simulations are computationally expensive and surrogate predictions contain localized errors. The surrogate can rapidly explore the design space, while selected FDS evaluations provide high-fidelity information needed to assess candidate solutions and identify regions where the surrogate requires refinement. The present study therefore illustrates a practical coupling of transient fire simulation, neural-network surrogate modeling, and trust-region model management. At the same time, the observed surrogate–FDS discrepancies emphasize that high-fidelity evaluations remain essential when optimization decisions depend on relatively small differences in the predicted objective.

    Several limitations should be considered when interpreting these results. The optimization was conducted over a limited range of doorway widths and used peak TEMP_LOW over the 60-s simulation period as the primary objective. The surrogate model also exhibited residual errors in both peak magnitude and timing despite adaptive refinement. The grid-sensitivity assessment further showed that the absolute temperature predictions remained dependent on spatial resolution over the investigated mesh range, with complete grid convergence not achieved. Accordingly, the quantitative temperature values should be interpreted with numerical-resolution uncertainty, while comparisons among the parametric and optimization cases remain based on a consistent baseline mesh.In addition, the present implementation adopts trust-region model-management principles in a simplified objective-based form rather than implementing the complete trust-region filter algorithm. Consequently, the results should be interpreted as demonstrating the feasibility and value of high-fidelity-guided surrogate optimization for the present compartment-fire problem rather than establishing a global optimum or a universally applicable fire-control strategy.

    CONCLUSIONS

    This study developed a computational framework that integrates high-fidelity FDS simulations, neural-network surrogate modeling, and trust-region model management for the analysis and optimization of transient compartment-fire dynamics. Parametric FDS simulations characterized the effects of fire intensity and doorway width on compartment temperatures, and the resulting transient data were used to train and evaluate the surrogate model. The surrogate was then incorporated into an iterative model-management procedure in which candidate solutions were assessed against additional high-fidelity FDS simulations. This approach enabled discrepancies between surrogate predictions and the high-fidelity response to be identified explicitly during the optimization process.

    For the ventilation conditions investigated at an HRRPUA of 400 kW/m⊃2;, the high-fidelity simulations showed a non-monotonic dependence of peak TEMP_LOW on doorway width. Within the high-fidelity-supported interval of 0.800–1.000 m, the model-management analysis identified 0.800 m as the best evaluated doorway width, corresponding to a peak TEMP_LOW of 1090 °C. This represents an approximately 1.8% reduction relative to the 1110 °C peak at the initial 1.000 m doorway width. The result is interpreted as the best model-managed solution within the evaluated interval and not as a global optimum.

    The results demonstrate that surrogate accuracy alone is insufficient for reliable optimization when localized discrepancies occur between the surrogate and high-fidelity model. Additional FDS evaluations and adaptive surrogate refinement improved the representation of portions of the local response but did not eliminate errors in peak-temperature magnitude and trend. High-fidelity verification therefore remained essential for evaluating optimization candidates. The grid-sensitivity assessment also demonstrated that the absolute temperature predictions remained dependent on mesh resolution, emphasizing the need to distinguish comparative trends obtained on a consistent mesh from grid-converged quantitative temperature predictions. The framework developed in this study provides a foundation for future work involving broader ventilation and fire conditions, improved adaptive surrogate models, and more complete trust-region filter formulations for computational fire-safety optimization.

    ACKNOWLEDGEMENT

    The author thanks Dr. Biegler for teaching him the trust-region strategy.  The author thanks Dr. Carlos Ramirez for encouraging him to write single author papers.

    References

    1. McGrattan, K. B., McDermott, R., Weinschenk, C., Overholt, K., Hostikka, S., and Floyd, J. (2013). Fire Dynamics Simulator Technical Reference Guide, Volume 1: Mathematical Model, Sixth Edition. NIST Special Publication 1018, National Institute of Standards and Technology. https://doi.org/10.6028/NIST.SP.1018e6.
    2. McGrattan, K. B., McDermott, R., Weinschenk, C., and Forney, G. P. (2013). Fire Dynamics Simulator User’s Guide, Sixth Edition. NIST Special Publication 1019, National Institute of Standards and Technology. https://doi.org/10.6028/NIST.SP.1019.
    3. Lafdal, B., Djebbar, R., Boulet, P., Mehaddi, R., Koutaiba, E. M., Beji, T., and Torero, J. L. (2022). Numerical study of the combustion regimes in naturally-vented compartment fires. Fire Safety Journal, 131, 103604. https://doi.org/10.1016/j.firesaf.2022.103604.
    4.  Beshir, M., Wang, Y., Cicione, A., Krajcovic, M., Hadden, R., and Rush, D. (2025). Qualitative analysis of ventilation position and dimension effects on compartment fire dynamics: An experimental and numerical approach. Fire Technology, 61, 3615–3647. https://doi.org/10.1007/s10694-025-01747-5.
    5.  Nguyen, H. T., Abu-Zidan, Y., Zhang, G., and Nguyen, K. T. Q. (2022). Machine learning-based surrogate model for calibrating fire source properties in FDS models of façade fire tests. Fire Safety Journal, 130, 103591. https://doi.org/10.1016/j.firesaf.2022.103591.
    6.  Eason, J. P., and Biegler, L. T. (2016). A trust region filter method for glass box/black box optimization. AIChE Journal, 62(9), 3124–3136. https://doi.org/10.1002/aic.15325.
    7. Eason, J. P., and Biegler, L. T. (2018). Advanced trust region optimization strategies for glass box/black box models. AIChE Journal, 64, 3934–3943. https://doi.org/10.1002/aic.16364.
    8. Biegler, L. T. (2024). The trust region filter strategy: Survey of a rigorous approach for optimization with surrogate models. Digital Chemical Engineering, 13, 100197. https://doi.org/10.1016/j.dche.2024.100197.
    9. Liang, R., Han, Y., Hu, H., Chen, B., Yuan, Z., and Biegler, L. T. (2024). Efficient trust region filter modeling strategies for computationally expensive black-box optimization. Computers & Chemical Engineering, 189, 108816. https://doi.org/10.1016/j.compchemeng.2024.108816.