3. Benchmarks information
This section gives an overvies of the FireBench Benchmarking Process.
Collection of observational data FireBench collects and curates observational datasets related to fire across multiple scales, including:
Large wildfire events
Prescribed burns
Laboratory-scale experiments
Diversity of observations Observational datasets may describe different aspects of fire-related phenomena, such as:
Fire spread and progression
Weather and atmospheric conditions
Fuel properties
Building damage and impacts
Standardization of observational data All observational datasets are standardized and stored in a FireBench standard file format. This common format:
Simplifies benchmarking operations
Ensures consistency across datasets
Centralizes heterogeneous observations under a single structure
Standardization of model outputs Evaluated model outputs are converted to the same standard file format using a limited set of FireBench tools.
These tools ensure compatibility with the benchmarking framework
Interested users should contact the FireBench team for access and guidance
Benchmark execution Once both:
an observational standard file, and
a model output standard file
are available, FireBench provides benchmark scripts (Python files) that:
Run a predefined or custom set of benchmarks
Compare model outputs against observations
Scorecard generation Each benchmark run produces a scorecard that summarizes evaluation results:
Qualitative and quantitative performance indicators
Delivered as both JSON (machine-readable) and PDF (human-readable) formats
Metrics and evaluation methodology The definitions of:
Metrics
Key Performance Indicators (KPIs)
Normalization and aggregation functions
are described in detail in the documentation and are shared across all benchmarks for transparency and reproducibility.
Distribution of benchmarks Benchmark datasets and benchmark scripts are distributed via Zenodo. Running the benchmarks requires the FireBench Python library.
Certification and authenticity (optional) At multiple stages of the process, the FireBench team can deliver a certificate of authenticity using hardware-based cryptographic methods. These certificates can be used to:
Authenticate datasets
Certify benchmark executions
Validate published benchmarking results