Adaptive wildfire spread prediction for complex terrain: modeling the effectiveness of sprinkler systems

Abstract

Background

Because the threat of wildfires to global ecosystems and society continues to rise, this study provides an experimental simulation framework that assesses the spread and reduction of wildfires to evaluate the effectiveness of adaptation methods in reducing their impact. The process entails selecting a vulnerable wildfire area and adaptation method, then generating the computational fluid dynamics (CFD) model. Monitoring data are then used to configure the model, set boundary conditions, and simulate the fire. The effectiveness of the adaptation method in minimizing damage in the area of interest is evaluated by comparing simulations with and without the chosen adaptation method. Our focus area was a natural recreational forest in Wonju, Gangwon-do, Korea, and our adaptation method was a water sprinkler system.

Results

Our framework provides aims to provide an experimental means of assessing the wildfire spread path and spread area based on exogenous variables of wind speed, wind direction, relative humidity, and more. The sprinkler adaptation had a reduction effect of 20% in the wildfire spread rate for the 10-h period, which refers to the time limit of the simulation after ignition. We revealed that at higher wind speeds, the fire primarily follows the wind direction; whereas at lower wind speeds, the fire is more influenced by the topography. Additionally, 60 min after ignition, the adaptation methods can suppress wildfire spread by > 70%. Notably, sprinklers reduce smoke concentrations by up to 50% (ppm) over the affected area.

Conclusions

This study demonstrates the potential effectiveness of a comprehensive CFD model in mitigating wildfire spread using sprinkler systems as an experimental analysis. Key results include a 20% reduction in wildfire within 10 h of ignition, significant influence of wind speed on spread patterns, and a reduction of smoke concentrations, improving air quality. These findings highlight the potential of CFD-based frameworks to enhance wildfire response strategies. However, it is important to note that this study’s limitations include the lack of experimental or measured fire behavior data, which should be considered when interpreting the effectiveness of the CFD model.

Resumen

Antecedentes

Dado que la amenaza de los incendios de vegetación a los ecosistemas globales y a la sociedad continúan incrementándose, este estudio provee de un marco de trabajo operacional de un simulador basado en la física, que permite determinar la propagación y reducción de incendios de vegetación para evaluar la efectividad de métodos de adaptación para reducir sus impactos.  El proceso implica seleccionar un área vulnerable a incendios y un método de adaptación, generando luego el modelo computacional basado en la dinámica de fluidos (CFD). Luego, los datos del monitoreo son usados para configurar el modelo, establecer las condiciones límites, y simular el fuego.  A posteriori, la efectividad del método de adaptación para minimizar los daños en el área de interés es evaluada mediante la comparación de simulaciones con y sin el método de adaptación elegido. Nuestra área focal fue un bosque recreativo en Wonju, Gangwon-do, en Corea, y nuestro método de adaptación fue un sistema de regadores de agua.

Resultados

Nuestro marco de trabajo provee de un medio efectivo para mitigar el camino de propagación del fuego, el área de propagación, el área de difusión, y el área de difusión basada en variables exógenas como velocidad del viento, humedad relativa, y otras.  Nuestra adaptación del regador tuvo un efecto de reducción del 20% en la propagación del fuego luego de 10 h posteriores a la ignición, y confirmamos que, a mayores velocidades del viento, el fuego sigue al viento: de otra manera trepa con la topografía. Basado en nuestros resultados, los métodos de adaptación pueden suprimir la propagación del fuego en un 70%, más que suficiente para que los bomberos puedan intervenir y salvar vidas. Notablemente, los regadores reducen la concentración de humo por hasta un 50% (ppm) sobre el área afectada.

Conclusiones

Llenamos un vacío en nuestro entendimiento colectivo sobre las capacidades ofrecidas por técnicas de adaptación y mitigación de incendios, y por lo tanto contribuyen a los esfuerzos de preparación necesarios para afrontar desastres naturales a nivel global a través de una eficiente selección y simulación de tecnologías adaptativas, e intentamos permanecer comprometidos en futuros esfuerzos.

Background

In recent years, the increasing threat of wildfires to society and global ecosystems has accelerated (Oliveira et al. 2021; Preisler et al. 2004; Yoo & Song 2023), and human-driven climate change continues to exacerbate the problem (McKenzie et al. 2004; Westerling & Bryant 2008). Indeed, the impact extends far beyond air pollution and forest damage, as topsoil loss, human casualties, biodiversity destruction, and ecosystem collapse are common repercussions (Gill et al. 2013; Kim et al. 2019; Vilar et al. 2021).

As far back as 1910, industrial civilizations have been unable to cope well with wildfires. At that time, the Devil’s Broom Wildland Fire in Idaho killed 78 people. More than 100 years later, the 2013 Yarnell Hill Fire in Arizona killed 19 people (NFPA 2017). In 2020, California endured severe wildfires, with the annual area burned increasing from roughly 54,000 hectares in 2010 to over 1,600,000 hectares (Cheung & Giardino 2023, CAL FIRE 2019).

Since 2019, annual wildfires in Gangwon-do have dramatically increased the demand for emergency medical services including the number of call-outs, hospitalizations, medical treatments, and emergency room visits, leading to higher public and private expenditures (Lee & Oh et al. 2022); hence, efforts are underway to mitigate the socioeconomic damage using improved preventative measures. As such, wildfire risk assessments have become a key component of planning for the consequences of our 4th Industrial Revolution (Chuvieco & Kasischke 2007; Finney 2005; Jain et al. 2020; Linn et al. 2002; Ning et al. 2024; Pais et al. 2021). Mitigation refers to the strategies and technologies used to reduce the severity and impact of wildfires. For example, Jazebi et al. (2019) surveyed wildfire mitigation technologies, including water, foam, and polymer gel sprays, as well as advanced ground robots with fire-extinguishing systems. Additionally, several new technologies have been proposed for blocking and/or redirecting wildfires. For example, Wang et al. (2021) developed an ecological green fire barrier solution that places natural-looking barricades up to 50 m in length to support firebreaks and ecological restoration.

In terms of advanced computational methods, Agranat and Perminov (2020) created a computational fluid dynamics (CFD) model that enables effect-based wind-speed change analyses. These types of innovations allow us to trace the intricate correlations between vegetation type, environmental factors, and fire spread. Advanced simulation studies have considered numerous variable and parameter combinations in recent years to better predict wildfire behaviors (Duff et al. 2016; Valero et al. 2021). Interestingly, Luo et al. (2019) found that occupant behaviors, such as window operations and HVAC usage, significantly impact indoor air flow patterns and pollutant concentrations during wildfire events, and the pioneering studies of Morvan and Dupuy (2004) provided novel numerical analytical methods to gauge the effects of wind effects on heat transfer between plant species, laying the groundwork for improved simulation techniques.

In terms of modern artificial intelligence (AI), Labati et al. (2013) proposed an adaptable neural network image processing system that monitors fire-prone environments to identify wildfire smoke, and Piao et al. (2021) utilized Google Earth and a random forest machine-learning algorithm to predict changes in land use land cover (LULC). Toan et al. (2019) developed a deep-learning model that uses hyperspectral satellite imagery to detect wildfires, and Lee and Lee et al. (2022a, 2022b) combined a forest growth model with the Weibull function and leveraged empirical mortality rate data to predict future wildfire fuel loads in an effort to assist local decision-makers and researchers.

Despite these new methods and the significant range of parameters they address, there is a gap in our understanding of the efficacy of current adaptation (mitigation) techniques (Dunn et al. 2020; Tedim et al. 2018). Notably, current computational modeling and AI prediction methods are limited in their ability to represent the effects of changing weather conditions, dynamic topographies, vegetation growth and decay, and most crucially, human-influenced climate change (Thompson & Calkin 2011). Therefore, we advocate for the expansion of CFD methods over other types due to their detailed ability to represent crucial fire dynamics, such as turbulence, combustion, and radiative heat transfer, which significantly enhance the development of fire management strategies.

Numerous studies have validated CFD models through comparison with lab experiments involving wildland fuels and field experiments. Mell et al. (2007) provided a physics-based approach to modeling grassland fires, which has been foundational in understanding fire dynamics in such environments. Moinuddin et al. (2020) focused on tree fires and the transition of fires from the forest floor to the canopy, emphasizing the need for numerical rigor in CFD simulations. Further, Raposo et al. (2018) analyzed junction fires at both laboratory and field scales, demonstrating the effectiveness of CFD models in predicting fire spread and behavior. McGrattan (2017) presented a sensitivity study for CFD simulations of grassland fires, comparing them with experimental results, which highlighted the crucial role of wind speed and moisture content in fire spread. Additionally, detailed physical modeling by researchers such as Mell et al. (2013) and Morvan et al. (2007) has been instrumental in refining CFD models through empirical validation. These studies collectively underscore the importance of validating CFD models with experimental and observational data to enhance their reliability and applicability in wildland fire scenarios.

To help fill the capability and knowledge gaps, this study elucidates the complex interplay between adaptive structures and the wind–topography nexus to provide an improved understanding of extant mitigation strategies used in fire-prone areas. Specifically, our goal is to incorporate the most comprehensive set of environmental variables to date and produce an adaptable CFD model that handles complex terrain–atmosphere interactions on the largest scale possible so that society may benefit from accurate and generalizable scenario-based wildfire spreading prediction. To meet this goal, we have established three distinctive objectives for our experimental operational model, as follows:

• 1. Select a vulnerable wildfire area and adaptation method; generate the CFD model.

• 2. Use monitoring data to prepare the CFD model, set boundary conditions, and simulate the spread and damage.

• 3. Analyze the effectiveness of the adaptation method in minimizing damage in the area of interest.

In terms of impact, we plan to contribute to the needed preparation and mitigation efforts in the face of global natural disasters through the efficient selection and simulation of adaptive technologies. Prior to delving into the methodology of achieving this goal, we identified four key performance variables that we use to assess the available adaptation methods:

• Spread path—Trajectory/directional pattern of wildfire expansion, considering the rate of spread.

• Spread area—Total area of land affected by the wildfire burn.

• Diffusion path—Trajectory/directional pattern of pollutants generated by the wildfire.

• Diffusion area—Area over which wildfire-generated pollutants spread at a concentration of 0.01 ppm or greater.

Incidentally, we have provided the following table of acronym and abbreviation definitions for use in this manuscript:

Acronyms and abbreviations
AI	Artificial intelligence
CAE	Computer-aided engineering
CBA	Cost–benefit analysis
CFD	Computational fluid dynamics
FVM	Finite volume method
IoT	Internet of things
RANS	Reynolds-averaged Navier–Stokes
RH	Relative humidity
RMSE	Root mean-square error
URANS	Unsteady Reynolds-averaged Navier–Stokes

Methods

Our methodology was structured according to the framework presented in Fig. 1. Based on the research goal of this study (i.e., to incorporate the most comprehensive set of environmental variables to date and produce an adaptable CFD model that handles complex terrain–atmosphere interactions on the largest scale possible so that society may benefit from accurate and generalizable scenario-based wildfire spreading prediction), this work was constructed around the three primary experimental objectives listed in the “Background” section.

Fig. 1

Schematic diagram of the study

Step 1, as detailed in the “Selection of vulnerable wildfire area and adaptation method—CFD model generation” section, involves selecting a vulnerable wildfire area, choosing an adaptation method, and generating a CFD model based on our configuration. Site selection entails ascertaining the frequency of visitors, the size of the forest area, and the historic wildfire frequency in Gangwon-do. 3D spatial data, building information, and tree types are included, and the reduction effects of adaptation technologies are analyzed (typically with internet-of-things (IoT)-based environmental data sensors). Step 2, as detailed in the “Preparation of the CFD model, addition of boundary conditions, and simulation validation” section, involves acquiring environmental data for the selected area and configuring the model (e.g., establishing boundary conditions) to simulate the scenario with and without the chosen adaptation technology. This step emphasizes the experimental nature of this study, as it relies heavily on accurately simulated environmental conditions. Step 3, as detailed in the “Model and adaptation method evaluation” section, entails evaluating the effectiveness of the adaptation technology in minimizing the area, altering the spread direction, and reducing pollutant transport in the area of interest. Performance variables include the diffusion path and diffusion time of the simulated fire, generated smoke, and changes in the spread area.

Selection of vulnerable wildfire area and adaptation method—CFD model generation

For our study, we selected a natural recreational forest area at Chiak Mountain in Gangwon-do. Gangwon-do is a local government located on the Korean Peninsula that covers 21.7% of the national forest area, with approximately 81% of the entire Gangwon-do area being forested (Fig. 3). The vulnerability of this area is obvious, as an uncontrolled wildfire would be devastating. Indeed, between 2014 and 2023, an average of 78.4 wildfires occurred annually in this region, highlighting the continuing risk (Korea Forest Service 2024). Moreover, the increasing number of recreational tourists and frequent visitors, as indicated in Fig. 2, creates a clear sense of urgency. The number of recreational forests in Korea has rapidly increased since 2006 (Fig. 2a), and the number of visitors to this particular area has more than doubled (Korea Forest Service 2021a).

Fig. 2

Site selection process: a increase in forest users and b causes of wildfires

In 2021, the Korea National Statistical Office released a report detailing wildfire causes for the previous decade (Fig. 2b), showing conclusively that unintended human behaviors were the largest culprit (Korea Forest Service 2021b). The Pinocchio Natural Forest, where the target site is located, covers an area of 86 ha, of which 33 hectares are dedicated to camping. The specific coordinates for the center of Pinocchio Natural Recreation Forest are (37.258807, 128.164110). There were three campgrounds in this area, and we established a 4 × 4 km subregion for our simulation, which entailed an ignition source at campground 1 (the largest camping area). In terms of adaptation methods, we opted for a sprinkler system that matches reality quite closely. A “sprinkler” refers to the pipe through which water is delivered, and a “sprinkler inlet” refers to a nozzle through which water is sprayed onto the affected area. These systems have been demonstrably effective in real-world configurations for decades (Labossière & McGee 2017), as nozzles installed on the outer walls of buildings can mitigate fire damage and spread (Green 2019). Moreover, high-speed imaging systems have been used to verify the effects (Green & Cooper 2019). Sprinklers are also installed on forest roads and around wooden buildings to prevent the loss of vital transportation corridors and cultural assets (Kim et al. 2012, Nam & Keum. 2013, Kong 2007) (Fig. 3).

Fig. 3

Site analysis: Chiak Mountain in Gangwon-do, South Korea

To create an experimental CFD model of the site, digital elevation model (DEM) data, building information, road configurations, and terrain slopes were obtained from the National Geospatial Information Platform (National Geospatial Information Institute Platform 2022). The chosen area included 1249 buildings with an average height of 4.8 m and 26 km of improved roads. The terrain elevation extended to 1227 m, with a steep northwest slope. Vegetation information and average tree sizes were obtained from the Forest Spatial Information Service (Korea Forest Service 2013). The vegetation consisted mostly of Larix kaempferi and comprised medium-sized trees with an average diameter of ~ 30 cm. Simulated sprinkler nozzles were installed at 60-cm intervals based on actual sprinklers installed in and around the campground area (Fig. 4). A simulated 550-m sprinkler line was installed along the forest road in the upstream direction of the mountain, starting 50 m away from the wildfire ignition point (Fig. 5a).

Fig. 4

Installation of sprinklers for wildfire reduction: a example installation at Mt. Gwanak and b spacing of sprinkler nozzles

Fig. 5

Validation of wildfire adaptation model: a environmental validation methods and sensor-based monitoring; b layout of sprinkler system and simulation scenario (Google Earth); c validation methods of flame model

Preparation of the CFD model, addition of boundary conditions, and simulation validation

We fabricated our experimental computational model using STAR-CCM + v.2210.1 (Siemens), a computer-aided engineering (CAE) multiphysics CFD solution used to simulate realistic environments (Kim & Kang 2022b). The model uses finite volume methods (FVMs) and has a good track record of simulating fluid flow changes in small indoor spaces. Notably, FVMs have been successfully used to simulate various large spaces for disaster preparedness and training, including reproducing realistic heat island phenomena, which are important to our study (Zou et al. 2018).

Two simulated IoT sensors were installed at campground 1 to monitor the spread of wildfire and smoke toward campground 3, nearly 500 m away. Real IoT weather sensors collected environmental information (e.g., temperature, humidity, wind direction, wind speed, and sunlight) every minute for 1 year from 1 January to 31 December 2022, which we used for our system configuration (Table 1).

Table 1 Sensor specifications

We calibrated and validated our experimental model by comparing simulated versus real-world airflow data on a specific spring day (15 April), as spring is when wildfires occur most frequently. In Korea, approximately 56% of wildfires occurred in the spring and 27% occurred in the winter over the past 10 years from 2014 to 2023 (Korean Statistical Information Service 2024). After loading the data, we analyzed the difference between simulated and real weather data, noting that the simulated and real installation locations corresponded (Fig. 5a). Weather station A was positioned at campground 1 (the ignition site), and weather station B was placed 500 m northwest at point 7 at campground 3. Data from 0800 to 2100 h on 15 April were assessed. Importantly, these data supplied many of the boundary conditions, which were distinct from the model’s training data. That is, the model did not directly learn from weather station A’s data. Likewise, weather station B’s data were used for independent validation to insure that the validation process was fully separated from model training. This method also prevented overfitting and ensured a fair demonstration of model generalizability. To verify environmental factors, we computed the R² value as a measure of the direct correlation between the simulation and observed data. The R² value indicates the proportion of variance for a dependent variable as explained by an independent variable in the regression model. A model is considered reliable with an R² value of 0.8 or higher (Acero & Herranz-Pascual 2015; Erell & Williamson 2006; Ghaffarianhoseini et al. 2015; Lee et al. 2016; Nasrollahi et al. 2017).

We also calculated the root mean-square error (RMSE) to assess precision (i.e., reliability). For regression analysis, we selected data points based on their proximity to the wildfire ignition site and the consistency of meteorological conditions on the simulated day. The presentation of cases A and B was distinguished by the absence or presence, respectively, of sprinkler installations, as shown in Fig. 5b. In addressing the placement of sprinklers, our selection was informed by practical considerations and the position of the real installation.

Our second verification method evaluated flame and smoke data as we implemented a dual-chamber simulation to reproduce the dynamic properties of a fire in a controlled environment, as guided by the foundational experiments of Steckler et al. (1982). We adapted their approach to measure the spread of flames and smoke through a vertical array of sensors (i.e., line probes) extending from the ground to 1.83 m in height (Fig. 5c). Chamber A served as the ignition point, and chamber B was used to observe the propagation of fire and smoke. A door between the two chambers allowed for the spread of flames, which we monitored using a line probe comprising 36 points from the ground to the maximum height. This arrangement facilitated the accurate capture of thermal changes across the plane section in the middle of the chamber spaces. These data provided a benchmark for experimental CFD validation using STAR-CCM + . The maximum flame temperature was set to 150 °C for visualization purposes, with actual temperatures being higher. The reliability of simulation is grounded in the validated Steckler experiment, ensuring the accuracy of our experimental approach in representing realistic fire dynamics. The simulation’s reliability is based on the validated Steckler experiment. By comparing the thermal measurements from the physical experiment with the CFD results, we verified the simulation accuracy and provided a basis for validating the model’s applicability to wildland fuels in a natural environment.

This experimental approach demonstrates the potential of CFD models in wildfire research. Although there are inherent limitations due to the lack of real-world fire behavior data, the results obtained from these methods are valuable. They offer significant insights into wildfire dynamics and the effectiveness of mitigation strategies, providing a foundation for future research and real-world application. Since this study has not validated the behavior of wildfires in large forested areas, future research should focus on validation to enhance the robustness and accuracy of the experimental models.

Model and adaptation method evaluation

Mesh setting and physics model selection

A computational model was created using Rhino v.7.0 to simulate the spread and evolution of the wildfire. Rhino v.7.0 was used primarily for terrain and shape modeling in our study. For simulating the wildfire dynamics, including phenomena such as crown fires, we employed STAR-CCM + , an experimental CFD-based model. This approach has been widely used and validated in numerous studies to accurately model complex fire behaviors. This model includes buildings, vegetation, terrain, sprinklers, roads, and other surfaces, all of which are known to affect flame and smoke propagation. The vegetation information included ignition temperature, which refer to the specific temperature at which each type of tree begins to ignite and are used to express fire spread. Additionally, topographic information was modeled to simulate air flow using recorded historic wind speed data. The geometry of the configuration was meshed using the FVM technique with different grid sizes for each component to adequately assess airflow, flame and smoke migration, and the impact of water droplets. The minimum thickness percentage of the prism layer was set to 0.01, and the layer reduction percentage was set to 10.0. The prism layer was set to ten as it naturally spreads the flow when the size of the mesh suddenly decreases. The grid sizes of buildings, rivers, and sprinklers are listed in Table 2, and mesh sizes were verified using a dependable mesh sensitivity test (Bakovic et al. 2017; Kim & Kang 2022b; Norton et al. 2010). Case A (no sprinklers) used 6,027,878 meshes, and case B (sprinklers) used 26,568,595.

Table 2 Mesh sizes (m) used in the CFD model

The mesh setup in this study was performed by creating three levels of grid density: fine, medium, and coarse, based on the traditional methods proposed by Bayon et al. (2016). The mesh densities were set to 6.03 million (fine), 2.46 million (medium), and 1.02 million (coarse) cells, respectively (Fig. 6). For the mesh sensitivity analysis, we evaluated the performance of each mesh density in simulating the combustion and spread of wildfires. Mesh sensitivity was accessed by comparing of the simulations with different mesh densities to determine the impact of grid size on the accuracy of the CFD model.

Fig. 6

CFD models with different mesh sizes

To quantify the uncertainty and ensure the reliability of our simulations, we used the Grid Convergence Index (GCI), a method widely recognized for assessing the accuracy of numerical solutions. The GCI provides a measure of the numerical error associated with the grid resolution. When considering a 95% confidence interval in predicting fire spread, the fine mesh with 6.03 million cells exhibited a GCI21 uncertainty of 1.56% (Table 3). This low uncertainty indicates that the fine mesh provides a high level of accuracy.

Table 3 Results of mesh sensitivity test

Further refinement of the fine mesh would yield minimal improvements in numerical results, as indicated by the low GCI uncertainty. Therefore, the fine mesh represents an optimal balance between computational cost and accuracy. By using the fine mesh for the grid setup in this study, we ensured that the simulations were both accurate and computationally efficient.

This detailed mesh sensitivity analysis highlights the robustness of our experimental approach, demonstrating the reliability of CFD model in representing the complex dynamics of wildfire combustion and spread.

Implicit unsteady analysis was performed to examine changes in the wildfire over time, and a passive scalar function was used to calculate smoke spread. To consider sprinkler water injection, a heat-flow analysis model that included relative humidity (RH) was used, and a multicomponent gas model was applied. A realizable k–ɛ two-layer model was used to model turbulence. Additionally, 19 models, including 3D, gradient, gravity, Reynolds-averaged Navier–Stokes (RANS), turbulent, segregated flow, and segregated fluid temperature types, were used. The unsteady RANS (URANS) model provided the equation governing the change in momentum over time. RANS models are based on the time averages of turbulent fluctuations (Blocken et al. 2012; Du et al. 2021; Neofytou et al. 2006). In the URANS version, flow variables f are divided into mean (time-averaged) $\overline{f}$ and fluctuating ${f}{^{\prime}}$ components. The URANS equation is defined by Eqs. (1) and (2) (Mei & Yuan 2021):

$$\frac{\partial \overline{{u }_{i}}}{\partial {x}_{i}}=0,$$

(1)

$$\frac{{\partial}\overline{{u}}_{i}}{\partial_{t}}+\frac{\partial(\overline{u_{j}u_{j}})}{\partial{x}_{j}}=-\frac{1}{\partial}\frac{\partial_{p}}{\partial_{xj}}+\frac{\partial}{\partial_{xj}}(\frac{\mu}{\rho}(\frac{{\partial}\overline{u_{l}}}{{\partial}_{xj}}+\frac{{\partial}\overline{u_{j}}}{\partial_{xi}})-\overline{{\rho}{x}{u}^{\prime}_{i}{u^{\prime}_{j})}}-g_{i}{\beta}(T-T_{ref}),$$

(2)

where $\overline{{u }_{i}},\overline{{u }_{j}}$ are the mean air velocities (m/s), ${u}_{i}{^{\prime}},{u}_{j}^{{\prime}}$ are the fluctuating velocities (m/s), $-\rho \overline{{u}_{i}{^{\prime}}{u}_{j}{^{\prime}}}$ is the turbulent Reynolds stress tensor (kg·m⁻¹·s⁻²) that represents the turbulent stresses caused by velocity fluctuations, and ρ is the density of the air (kg/m³). Dynamic viscosity μ has units of kg·m⁻¹·s⁻¹. Furthermore, − g_iβ(T − T_ref) is the buoyancy force per unit volume, with g being the acceleration of gravity (m/s²), and β is the thermal expansion coefficient (1/°C), where T is the mean temperature (°C). T_ref is the reference potential temperature (°C), and the buoyancy term is expressed as force per volume, which is equivalent to the unit of pressure (kg·m⁻¹·s⁻²). The term $t$ represents time (s). The realizable k–ɛ model for turbulent kinetic energy k is expressed by Eq. (3), which was designed to minimize unknowns in turbulent flow simulations (Ashgriz & Mostaghimi 2002; Chen et al. 2021; Eymard et al. 2000; Moukalled et al. 2016; Van Maele et al. 2003; Versteeg & Malalasekera 2007):

$$\frac{\partial (\rho k)}{\partial t}+\frac{\partial (\rho k{u}_{i})}{\partial {x}_{i}}=\frac{\partial }{{\partial x}_{j}}\left[(\mu +\frac{{\mu }_{t}}{{\sigma }_{k}})\frac{\partial k}{{\partial x}_{i}}\right]+{P}_{k}-\rho \varepsilon +{S}_{k}.$$

(3)

Turbulent kinetic energy k and turbulent dissipation rate ɛ were calculated using the steady transport equation (Passandideh–Fard et al. 2002, Saleh 2002, Shirani et al. 2005):

$$\frac{\partial (\rho \varepsilon )}{\partial t}+\frac{\partial (\rho \varepsilon {u}_{i})}{\partial {x}_{i}}=\frac{\partial }{{\partial x}_{i}}\left[(\mu +\frac{{\mu }_{t}}{{\sigma }_{\varepsilon }})\frac{\partial \varepsilon }{{\partial x}_{i}}\right]+{C}_{1}{S}_{\varepsilon }-{C}_{2}\frac{\partial \left(\rho \varepsilon {u}_{i}\right)}{k+\sqrt{\nu \varepsilon }} +{S}_{\varepsilon },$$

(4)

where the turbulent kinetic energy production term, P_k, was computed using the mean velocity gradients. The model constants in Eq. (3) are C₁ and C₂, and the strain rate, S, is given by Eq. (5). C₁ is defined by Eq. (6), and ratio η by Eq. (7):

$$S=\sqrt{2{S}_{ij}{S}_{ij}},$$

(5)

$${C}_{1}=\text{max}(0.43,\frac{\eta }{\eta +5}),$$

(6)

$$\eta =\frac{Sk}{\varepsilon }.$$

(7)

Here, v is the kinematic viscosity (m²/s), and σ_κ and σ_ɛ are the turbulent Prandtl numbers for k and ɛ, respectively. C₂ was set to 1.9.

The temperature in the STAR-CCM + model used degrees Celsius, and RH was evaluated as a combination of temperature and the specific composition of air, including the presence of water vapor and other relevant gasses. The saturation vapor pressure, water vapor pressure, and absolute humidity used to evaluate RH are expressed by Eqs. (8)–(11) (Castellvi et al. 1996):

$$RH= \frac{100*{P}_{v}}{{P}_{s}},$$

(8)

where P_S represents the saturation vapor pressure, P_a, and P_v represents the partial pressure of the vapor. Saturation vapor pressure P_S is expressed as follows (Buck 1981):

$${P}_{s}=611.21{e}^{\frac{(\left(18.678-\frac{T}{234.5}\right)T)}{(257.14+T)}},$$

(9)

Partial vapor pressure P_v is expressed as (Green & Southard 2019):

$${P}_{v}= \frac{(H{P}_{abs})}{(\left(\frac{{M}_{w}}{{M}_{a}}\right)+H)},$$

(10)

where P_abs denotes the absolute pressure, P_a, M_W indicates the molecular weight of water (g/mol), M_a denotes the molecular weight of air (g/mol), and H represents the absolute humidity (g/m³). H is expressed as follows (Green & Dyer 2009):

$$H=\frac{{X}_{w}}{{X}_{a}},$$

(11)

where X_w and X_a represent the mass fractions of water and air, respectively.

The evaporative cooling of fog operates based on the principles of the latent heat of vaporization and wet-bulb depression. In principle, wet-bulb depression refers to the difference between bulb and wet-bulb temperatures, wherein heat Q(J) varies the phase of a sample of mass (kg):

$$Q=m{L}_{(T,u)},$$

(12)

where L_(T,u) represents the latent heat of vaporization (J/kg) (Lampinen et al. 2001). The vaporization heat of water, which depends on humidity, can be accurately determined as follows:

$${L}_{(T,u)}={l}_{T}+{r}_{(T,u)},$$

(13)

where l_T represents the latent heat of vaporization due to temperature (J/kg), and r_(T,u) indicates the required auxiliary (sorption) heat (J/kg). In Eq. (13), r_(T,u) is the auxiliary heat affected by ambient humidity, representing the additional energy required during sorption, and the vaporization heat, l_T, depends on the temperature and is less influenced by humidity.

To model the vertical distribution of wind speed, a wind profile U was set at the inlet, as defined by Eq. (14) (Hong et al. 2012):

$$U={U}_{r}{(\frac{Z}{{Z}_{r}})}^{0.33},$$

(14)

where U represents the wind speed at a specific height (m/s), and Z and U_r refer to the wind speed (m/s) at height Z_r (m).

Four additional field functions were used for the generation and spread of fire and smoke. The field function included the heat release rate from the flame source, smoke generation function, flame propagation function, and smoke propagation function. We referred to previous research to configure the field function (Table 4). Most of the properties are tailored to Larix kaempferi, but where data was unavailable, properties from other conifer species were used as substitutes.

Table 4 Field function value setting in the CFD model

Although wildfires occur in complex ways, they are generally classified into surface, stem, crown, and ground types Kim & Kang 2022a; Morvan 2007; Quintiere 2016). The ignition and smoke temperature points in our experimental simulation were set to emulate crown fires, which rapidly spread through the canopies of trees and shrubs (Bond & Van Wilgen 2012; Brown & Smith 2000; Pyne 1984).

Initial and boundary condition selection

The temperature, humidity, mass fraction of air, and mass fraction of vapor were set. In the initial condition, the temperature was set to 20 °C, humidity at 60%, turbulence intensity at 0.01 m/s, and turbulence velocity scale at 1.0 m/s based on sensor results (Table 5). The boundary conditions were set for wildfire analysis. The temperature and humidity values represent typical spring conditions, but they do not significantly affect the results due to rapid temperature changes caused by wildfires. This study primarily focuses on experimental wildfire behavior in relation to wind direction and velocity, so the wind speed range was determined based on measured values.

Table 5 Boundary conditions of the CFD model

The boundary conditions of a nearby river were included to more accurately reproduce the real environment, and the respective temperature and heat-transfer coefficient were set using data from a previous study (Hannesdóttir et al. 2012). The mass production of dry air and vapor was calculated as a combination of temperature, humidity, and the mole fraction of air and water vapor. The main building’s boundary condition included an internal temperature of 25 °C and a heat-transfer coefficient of 5.0 W/m²/K. Thermal resistance, which depends on thermal conductivity, was set to 1.21 W/mK, assuming a concrete wall thickness of 0.3 m.

A “side” refers to the space at which wind begins to influence the flow of fire and air. The initial temperature and RH conditions were set to initial conditions of 20 °C and 60%, respectively, and the mass fraction was calculated. Wind speed was simulated at speeds ranging from 1 to 5 m/s, and the correlation between wind speed changes and wildfire reduction effects was examined. The sprinkler system was designed with nozzles with a radius of 5 mm. Across the 550-m sprinkler, 917 nozzles were installed at 60-cm intervals, each engineered to dispense water at a high pressure of approximately 2.2 L/min or 2 t/h. The system utilized a single water tank with a capacity of about 20 t. The results after 10 h were examined, and the wildfires were set to start 10 min after interpretation began so that a sufficient wind flow field would exist according to the topography and space characteristics of the target site.

Fuel properties and structure

The following summarizes the fuel properties and structures used in the experimental simulation based on the information provided above section:

• 1. Properties of the fuel: The primary fuel used in the simulation is Larix kaempferi, which has an average moisture content of 15%, a density of 0.65 g/cm³, a heat of combustion of 17,020 kJ/kg, and an ignition temperature of 460 °C. Moisture content is crucial as it affects the ignition and combustion properties of the fuel and the value of density is important for calculating the heat release rate and understanding how the fuel will burn.

• 2. Structure of the fuel: The fuel is represented as a combination of surface and crown layers. The surface layer includes ground vegetation and fallen leaves which act as the initial fuel layer that can easily ignite and help in spreading the fire. While the crown layer represents the tree canopies. This layer consists of the higher parts of trees and is crucial for modeling crown fires that can spread rapidly. The vegetation is modeled with an average tree height of 20 m and a diameter of 30 cm. These dimensions are used to simulate the physical structure of the forest and how the fire spreads through it.

• 3. Spatial distribution of fuel: The spatial distribution of the fuel is heterogeneous, taking into account variations in vegetation density, types of vegetation, and the presence of clearings and forest roads. This distribution is based on real-world data from the selected site in Chiak Mountain, Gangwon-do. Using real-world data including DEM data, building information, and road configurations ensures the simulation closely mimics actual conditions.

By including these fuel properties and structures, we aim to provide an understanding of the fuel properties, structure, and spatial distribution used in our experimental simulations.

Wildfire adaptation evaluation

The main wind direction at the target site was analyzed using environmental monitoring values, and the behavioral changes in fire and smoke were analyzed based on changes in wind speed from 1 to 5 m/s. To quantitatively analyze the wildfire reduction effect based on effective sprinklers, the following three tasks were executed:

• 1. We evaluated the effects of various wind speeds and directions on wildfire spread paths and times. A set of 16 control points was placed in a circular pattern 500 m from the ignition source at 22.5$^\circ$ intervals beginning in an eastward direction. The points at locations farther along the same radius and angle were created using JavaScript code.

• 2. We evaluated the spread path of smoke using CFD simulations. We simulated the times at which smoke concentrations reached or exceeded 0.01 ppm, employing criteria informed by significant air impacts, as seen in the 2022 Uljin wildfire in South Korea.

• 3. We evaluated the changes in the wildfire spread area according to wind speed based on sprinklers versus no sprinklers.

In the 2022 Uljin wildfire, the concentration of SO2 was 0.005 ppm and NO2 was 0.028 ppm, making an average smoke concentration value of 0.01 ppm representative of the conditions observed during the wildfire (Hankyoreh, 2022).

Results

Sensor results and simulation verification

Two weather stations were used for the experimental CFD simulation, and WRPLOT (Lakes Environmental Software) was used to examine wind direction and speed changes from 1 January to 31 December 2022 (Lakes Environmental Software, 2022). The results of environmental monitoring at the site are shown in Fig. 7. At weather station A, the starting point of the wildfire, 36.81% of the winds were below 0.5 m/s, and more than 50% were between 0.5 and 2.10 m/s. The average wind speed was 0.91 m/s, and the windward direction was from the east. The wind environment at weather station A had a dominant influence and was the basis for our boundary conditions. Generally, wildfires spread according to weather conditions, including wind direction, with increasing rapidity under dry conditions, high wind speeds, and steep slopes. Nevertheless, our study employed a singular boundary condition based on dominant trends in wind speed and direction from our monitoring data, which varied within the anticipated ranges. At weather station B, wind speeds of 0.5 m/s or less occurred 26.18% of the time, and windward directions of 112.5° and 135° were the norm. Weather station B had wind more than 50% of the time between 0.5 and 2.10 m/s (average 1.50 m/s). Hence, weather station B was used to verify our experimental CFD model (Fig. 8). The variability in wind direction was low, with a standard of approximately 0.23 for weather station A and 0.28 for weather station B.

Fig. 7

Results of environmental monitoring

Fig. 8

Results of environmental validation: a temperature validation, b relative humidity validation, c wind speed validation, d wind direction validation, e simulation results—spatial distribution of environmental factors (after 10 min)

At weather station A, the R² value was 0.852 for temperature, 0.906 for TH, 0.85 for wind speed, and 0.85 for wind direction between the experimental CFD simulation and the measured value. At weather station B, the R² value was 0.96 for temperature, 0.94 for RH, 0.85 for wind speed, and 0.82 for wind direction. Figure 8e presents a simulation scene that reflects changes in temperature, RH, wind speed, and wind direction over time.

Our second experimental verification method focused on flames and smoke. Results from previous studies based on the STAR-CCM + v.2210 User Guide (Siemens 2022) were used to obtain the i value for temperature and wind speed. The R² value for temperature was 0.93, and its RMSE was 11.8. The R² value of the wind speed was 0.91, and its RMSE was 0.38, which reflects reliability (Fig. 9). The differences in the RMSE values for temperature and wind speed can be attributed to the range and scale of the measurements. Temperature measurements at the probe range up to 140 °C, while wind speed measurements range up to 2.5 m/s. The broader range and higher values in temperature measurements naturally lead to a larger RMSE. The values were shown in time when the experimental CFD simulation reached stability. For flames, the measurements were taken at the point when the flames reached their maximum intensity.

Fig. 9

Comparison of CFD simulation results with empirical measurements: a temperature comparison between CFD and measurement; b correlation plot for temperature between CFD and measurement; c velocity comparison between CFD and measurement; d correlation plot for velocity between CFD and measurement

Our fire flame velocity validation raised important considerations about the relationship between simulated and observed wind speeds. Notably, although the experimental conditions primarily reflected wind speeds of around 1.5 m/s, the objective of the simulation was not to replicate specific conditions but to validate flame spread patterns in a controlled setting. Precise wind speeds in the field are generally above 2 m/s. Other discrepancies can be attributed to the inherent challenges of controlling variables in a complex experimental deployment. In any case, the experimental simulations achieved R² values over 0.9, indicating high levels of correlation and reliability.

Wildfire propagation and adaptation analysis

Here, the spreading phenomenon of wildfires is examined based on case A (no sprinklers) and case B (sprinklers). A summary of the changes in the wildfire area and spread path are shown in Fig. 10 and Table 6, respectively. The upper part of Table 6 illustrates the spread rates of the wildfire, and the lower part illustrates the wildfire propagation.

Fig. 10

Wildfire area changes over time with varying wind speeds: a case A without sprinklers and b case B with sprinklers

Table 6 Increase in wildfire area and propagation

Case A—wildfire propagation without sprinklers

For case A, our analysis showed that with higher wind speeds, the wildfire spread to larger areas (Fig. 11). One hour after ignition, the wildfire spread at a 1 m/s wind speed to an area of 0.23 ha; at 3 m/s, it spread to an area of 8.24 ha, which represents a 35-fold increase (Fig. 10). At 10 h after ignition, a wind speed of 1 m/s spread approximately 8 ha of wildfire, whereas 3 m/s winds spread it to 477 ha, which represents a 60-fold increase. At a wind speed of 5 m/s, the wildfire spread rapidly to 41 ha after 60 min, 966 ha after 360 min, and 1361 ha after 600 min (Fig. 11). The spread of smoke was higher than 0.06 ppm in all regions, and the temperature within the flame envelope of the wildfire increased to 2000 °C.

Fig. 11

Comparative visualization of wildfire spread, temperature, and smoke propagation at 5 m/s, with and without sprinkler installation

In case A, the wildfire spread to a radius of 500 m when the wind speed exceeded 2 m/s. After 168 min at wind speeds at or below 2 m/s, the spread to point 5, which demarks a steep slope in a mountainous area. We found that the wildfire moved from the ignition point, progressing westward in the lee direction. It took 533 min for the wildfire to spread to point 9, which was the first point reached. However, when the wind speed was 5 m/s, it took 45 min to reach point 9. The wildfire then spread to the north and northwest, and the topography allowed it to spread quickly to the upper mountainous areas (point 5). Approximately 200 min after ignition, the fire spread to most monitoring points. In summary, with faster wind speeds, the fire adheres more to wind direction, resulting in higher spread rates. However, with slower wind speeds, the fire first spreads to higher elevations, affecting the fire spread rates differently.

Wildfire propagation with sprinkler

To analyze the land area that can be covered by a single sprinkler, we assessed the range of the sprinkler’s spray. Previous studies have indicated that the range spans from 3 to 20 m (Martín–Benito 1992, Fischer 1988). Upon examining the mist spray range of our nozzles 10 min after beginning the experimental simulation, we observed a maximum spray range of 19.66 m, a minimum of 15.54 m, and an average of 18.33 m. These values were derived from experimental simulations using the defined boundary conditions, considering the terrain and conditions of the target site area. These data illustrate good coverage and correspondingly high potential for mitigating spread.

We found that in case B, the area of the wildfire spread with a wind speed of 1 m/s after 1 h was 0.07 ha, and that with a wind speed of 3 m/s was 0.36 ha (~ five times larger; Fig. 10). After 10 h at a wind speed of 1 m/s, the spread was 7.84 ha, and at 3 m/s, it was 411 ha. At 5 m/s, the wildfire spread to 0.8 ha after 60 min, 415.1 ha after 360 min, and 1010.6 ha after 10 h (Fig. 11). Wildfires did not spread into areas where sprinklers were used at first, but over time, they gradually penetrated those areas as well. The temperature at the center of the wildfire was 2000 °C, the same as when sprinklers were not used. However, sprinklers mitigated smoke concentrations, keeping them between 0.03 and 0.06 ppm.

Overall, case B showed similar spread patterns according to wind speed and topography. That is, the faster the wind speed, the more the wildfire follows the wind’s lee direction; the slower the speed, the more the spread ascends the topography. In the case of winds of 2 m/s, the wildfire spread to point 5 after 486 min and reached point 10 after 589 min. At 5 m/s, the wildfire reached point 9 first, as in the case without sprinklers. It then proceeded toward point 5. Between the ignition point and point 7, where the sprinkler was installed, the spread was significantly delayed. However, in the case without sprinklers, the time needed to reach point 7 was 76 min versus ~ 312 min in case B. At 1 m/s, the wildfire did not spread past 500 m in either case because the necessary heat was not provided until a sufficient number of trees were fully ignited.

Effects of wildfire adaptation techniques

The variations in wildfire mitigation effects with wind speed are summarized in Fig. 12. The faster the wind speed, the greater the reduction in the wildfire-affected area (the initial rate was between 70 and 100%). The effect of the wildfire reduction consistently decreased over time. In particular, this rate was remarkably high at wind speeds between 2 and 5 m/s and the range of the water spray was also larger.

Fig. 12

Changes in the wildfire area reduction rate according to wind speed

After 300 min at 5 m/s, the wildfire reduction rate was approximately 62%, and we found that a wind speed of 2 m/s had a reduction effect of ~ 35%. At 600 min, the rate fell below 20% in all winds between 1 and 5 m/s, and large wildfires of 1000 ha or more occurred, depending on wind speeds. At 1 m/s, the damage reduction rate was numerically low, and the area was less than 10 ha, which is also low.

Discussion

This research offers novel insights into the utility of computer modeling and analysis methods for the direct benefit of human society. Moreover, the verification process was quite rigorous in terms of its comprehensive use of environmental factors and its data reliability. Considering that the areas affected by wildfires in Korea in 2022 comprised 24,015 ha, the cost of damages totaled KRW 226.1 billion and an additional restoration cost of KRW 417 billion. Hence, our results imply that the use of prediction and mitigation methods could save lives, infrastructure, and public funds.

In 2010, the cost of maintaining a 100-m section of sprinklers was ~ KRW 1.02 million (Kim et al. 2010). Considering that it cost KRW 530 million to install the system covering the 500-m section of our focus area (Chungchung News 2020), it cost KRW 5.61 million to maintain in 2010. Applying the average inflation rate of 5.1%, the maintenance cost in 2022 was KRW 9.23 million, and by 2030, it is projected to be KRW 15.17 million (Statistics Korea 2022). By extension, this would require KRW 131.75 million in 2030. Compared with the expenses of battling and recovering from uncontrolled and unmitigated wildfires, this cost seems reasonable. Moreover, in terms of the amount of water used by sprinklers compared with that used by fire-control helicopters (between 2 and 8 t), preventative suppression is clearly more desirable (Safety News 2022).

Politically, our results provide valuable resources for decision-makers and policy planners both in Korea and globally. This study confirmed notable correlations between wind speed and the direction and topography of the wildfire spread based on simulated values. However, there was no environmental verification related to flame propagation. Operationally, our results emphasize the importance of early wildfire mitigation. In particular, we showed that inexpensive, widely accessible sprinkler systems can greatly improve the wildfire reduction rate within 60 min of ignition, which would be extremely valuable in terms of gaining firefighting advantages and evacuating people, preferably prior to canopy ignition. Despite the wildfire’s eventual catch-up to the control case in our simulation, sprinklers and other adaptation methods have the potential to create timely buffer zones that can delay the spread and intensity of fires. Hence, the strategic placement of adaptation technologies is vital and should be further examined for the benefit of policy and strategy. By integrating these findings into disaster management practices, the use of adaptation systems can be optimized to improve human survivability while preserving delicate ecologies. As such, future studies should integrate emergency methods and firefighting capabilities to take advantage of this temporal benefit, which can be predicted using our CFD/CAE method to realize future returns on investment. Additionally, further experimental and operational validation/testing of these observations and conclusions is necessary to ensure their accuracy and applicability.

This study was limited in that we could not provide a statistically sound analysis of the offset related to the total costs of applying both adaptation methods versus the resulting mitigated firefighting and loss expenses. A detailed cost–benefit analysis (CBA) was out of the scope of this study but presents a valuable opportunity for further study. Moreover, research should focus on regional CBAs first and then apply them to larger geographical areas and perhaps nations.

The findings of this study may lack full generalizability, as they were derived from simulations conducted at a single site. Furthermore, we applied a single ignition point, which is not as frequent in reality as one may assume. Future research should incorporate multiple simultaneous ignition locations to robustly assess the efficacy of adaptation methods when confronting runaway, convergent disaster scenarios. Moreover, an assessment of the type and number of systems needed to fully protect areas would be worthwhile, especially in terms of estimating the infrastructure needed and the computational/AI models that could be brought to bear. The scenario considered in this study was limited to a simulated fire occurring near an existing and fully operational sprinkler system. Notably, it is unlikely that our global society will be able to fully equip entire forests with fully functional adaptation systems in the near future. Hence, reasonable risk assessments should be performed as they are now more feasible than before.

Additionally, our current model focuses primarily on simulating crown fires and does not encompass all extreme fire behaviors such as fire whirls and firebrands, as noted in Liu Naian et al. (2021). Future work should aim to incorporate these aspects to provide a more comprehensive understanding of large wildfire dynamics and improve the simulations.

In conclusion, this experimental study demonstrates the potential benefits and limitations of using sprinkler systems for wildfire mitigation. While sprinklers can significantly reduce the spread of wildfires, their long-term effectiveness under high wind conditions is limited. Future research should focus on validating CFD models with real-world data and exploring a wider range of environmental conditions and mitigation strategies. This will help develop more robust and effective wildfire management practices, ultimately reducing the devastating impacts of wildfires on ecosystems and human communities.

Conclusions

Based on the growing threat of wildfires in Korea and globally, which are clearly exacerbated by human-driven climate change, this study sought to help fill the gap in our collective understanding of the capabilities offered by modern wildfire adaptation and mitigation techniques. As such, we selected CFD techniques as our baseline due to their demonstrated ability to represent realistic fire and weather dynamics, and we devised a sensor-based environmental factor monitoring system for use in meeting our larger goal of incorporating the most comprehensive set of environmental variables to date and producing an adaptable CFD/CAE model that can be used to assess wildfire spreads in complex areas where terrain–atmosphere interactions occur at large scales. That is, we built an operational model and framework that allows users to select a vulnerable wildfire area and adaptation method and generate a CFD model. By using our method, users can acquire suitable monitoring data to calibrate the model, set boundary conditions, and simulate the spread and damage of a wildfire based on ignition sources and weather influences. As a result, this person can analyze the effectiveness of their adaptation method in minimizing and delaying damage in the site area. However, at a minimum, some baseline validation of fire behavior predictions is required to ensure the reliability of such an approach.

We quantitatively verified our method’s realism and used CFD analysis to assess the effectiveness of our chosen adaptation method (i.e., sprinklers) as an innovative means of mitigating the wildfire spread path, spread area, diffusion path, and diffusion area based on exogenous variables of wind speed, wind direction, RH, and more. The application of the URANS technique and the realistic k–ε model were crucial to the success of our simulation and analyses. As such, the following key findings were obtained:

• The sprinkler adaptation demonstrated a reduction effect of 20% based on the wildfire spread 10 h after ignition.

• At wind speeds greater than 4 m/s, the wildfire spread strongly follows the wind direction; below 3 m/s, it follows the topography to higher ground.

• It only takes 10 h for a single ignition-source wildfire to spread to 1360 ha of forested area, which is 165 times larger than the area observed with a wind speed of 1 m/s. These sprinkler reduction results are applicable when considering vegetation consisting mostly of Larix kaempferi, with medium-sized trees.

• After 60 min from ignition, the programmed sprinkler system, reflecting real-world conditions, suppressed wildfire spread over 70%, significantly aiding responders in saving lives and controlling the flames.

• Sprinklers are also effective in reducing the concentration of smoke. Without sprinklers, a high concentration of more than 0.06 ppm was confirmed in all regions; however, with sprinklers, concentrations of 0.03–0.06 ppm were observed.

In terms of impact, we believe that this study contributes to preparation and mitigation efforts in the face of global natural disasters through the efficient selection and simulation of adaptive technologies. We intend to remain engaged in future studies in this endeavor.

While this study provides valuable insights into the interactions between atmosphere and terrain, the lack of validation under actual wildfire conditions limits the applicability of the findings to real-world scenarios. This study verified these factors without incorporating experimental or measured fire behavior data for direct comparison and validation. This represents a limitation, as previous studies have demonstrated the importance of validating CFD models against both environmental and fire behavior observations. Future research should aim to include empirical fire behavior data to enhance the robustness and accuracy of the simulation results, thereby addressing this critical weakness. Subsequent studies should integrate empirical data from actual wildfires to validate the combined fire-atmosphere-terrain model, ensuring more reliable and practical mitigation strategies.