Summary:

This paper introduces WULFFSS, a newly developed statistical model designed to simulate the monthly probability and size of large forest fires (size ≥100 ha) across the western US, on a 12-km spatial resolution. The modeling framework consists of 3 statistical models that are called stepwise, forced by climate, vegetation, topographic, and anthropogenic variables. Two of the three models each consist of three components representing different groups of predictor variables, as well as interactions in between them. The modeling framework features its stochastic nature and interpretability, and computational efficiency. The framework is designed to be couple with forest ecosystem model (such as DYNAFFOREST) to allow bidirectional feedbacks between the vegetation and fire events. The application of the framework for 1985-2024 is able to capture the majority interannual variability of forest fires in the studied western US with some limitation. 

Comments:

The authors did a good job in building the statistical modeling framework to estimate fire probability and sizes in western US from multiple influencing groups of variables. Allowing the model to couple with an forest ecosystem model for feedbacks indeed promotes its future potential. However, there are questions about methodological choices and performance analysis that further explanation needs to be made to clarify the robustness of this paper.

1. The modeling framework consists of three stepwisely connected statistical models, how sensitive is the final simulation to the prior step variables?

2. For model P and A, each consists of the three component groups of predictor variables. The variables are categorized into spatial, annual cycle and temporal domains to be examined. This is quite different from the more conventional method. Why took this approach, instead of treating all predictors in the same single pool? Is there quantitative consideration that the three-domain method better than the standard multi-variable regression?

3. When presenting the landcover data in the model, the author uses 41 grid cells from 8 directions surrounding the target grid cell to calculate the forest connectivity. Is there specific reason how the chosen area (or, the number of the grid cells) is set?

4. When introducing the human factor in the model, the author uses population density and distance to road, as predictors. Could the author please explain why some other widely used predictors, such as GDP or fire agency availability, are not included in the domain, relating to suppression capability? 

5. The model yields quite good simulations for the other regions in the western US, yet it evidently underperforms in the CA/NV region on both annual frequency and forest area burned (Fig. 16d and Fig. 17d). Apart from the potential reasons listed, did author analyse as to where the discrepancy comes from, was it mainly caused by ignition probability, or together with fire size? What is the dominant driver for this bias?

6. The model significantly underestimates the extreme fire events in year 2020 and 2024 (in terms of fire size), does this imply that the model systematically lack the ability to capture fire events induced by extreme weather conditions accurately? Can author suggest ways to improve this?

7. By introducing self-regulation effect that reduces fuels from previous fires, will this “saturation” effect limit the model performance in future projection, considering that the climate change and ecosystem feedbacks may alter its self-regulation point?

General comments

This manuscript presents the development and evaluation of WULFFSS, a novel stochastic monthly gridded model for simulating large forest fires in the western United States. The model operates at a 12-km resolution and leverages interpretable statistical methods to estimate fire probability and size. Its key strengths include the integration of spatial, annual cycle and temporal anomaly components that can interact, high computational efficiency enabling large ensembles, and its design for coupling with the DYNAFFOREST ecosystem model to simulate vegetation-fire feedbacks. The model demonstrates considerable skill in capturing frequency and extent of western US forest fires, as validated through rigorous temporal and spatial cross-validation. The manuscript is thorough, well-structured, and describes the model, data, and validation procedures in significant detail. The work represents a valuable contribution to the field of fire modeling. However, the text is somewhat lengthy, and there is room for improvement. Therefore, I recommend that this manuscript be accepted with minor revisions.

Specific comments

1. The model operates at a 12-km resolution, and it is possible that within a grid cell, the actual burned area constitutes only a small fraction. A key concern is the issue of intra-grid-cell heterogeneity and the representativeness of the predictor variables. It would be valuable if the authors could discuss the potential implications of this scaling issue.
2. The authors have considered a comprehensive list of potential predictors for constructing the Sp, Cp, and Tp components. However, the rationale or guiding principles for assigning specific variables to each of these three domains (Spatial, Annual-cycle, Temporal anomaly) is not sufficiently elaborated. A clearer explanation of the criteria used to categorize predictors into S, C, or T would significantly enhance the methodological transparency and reproducibility.
3. The model commendably incorporates interaction terms between the S, C, and T composite predictors, but these are limited to pairwise (two-way) interactions. The potential three-way interaction (S×C×T) is not considered. Could the authors please justify this methodological choice?
4. I recommend including a schematic diagram illustrating the general framework of WULFFSS. This figure should visually depict the relationships between the three core statistical models (P, N, A), their required input data streams (landcover, topography, climate, etc.), the key data processing steps (e.g., resampling), and the bidirectional coupling relationship between WULFFSS and the DYNAFFOREST model.

Figure 3: Adding text labels (e.g., "Pacific Northwest" or “PNW”) onto the map for the four quadrant regions would improve its immediacy and clarity.

Figure 4: Providing quantitative goodness-of-fit metrics for the curve fits in each panel would be welcome. Furthermore, the relationship for some predictors (e.g., wetds_mean_1mo in panel b) appears weak or poorly captured by the fitted curve.

Line 37: “fire” to “fires”