Chemical Generalization Boundaries in Fire-Extinction Prediction from a Public Surfactant Formulation Database
Year: 2026
Type: preprint
Abstract: Machine-learning models of firefighting surfactants can appear persuasive when repeated measurements or closely related formulations occur on both sides of a random split. We reconstructed the public Hinnant surfactant workbook at cell-level provenance and evaluated fire extinction with explicit failure labels and a hierarchy of increasingly stringent chemical splits. The reconstructed data contain 917 heptane and 595 gasoline fire tests; heptane extinction comprised 733 successes, 179 documented failures, and 5 unlabeled observations, whereas gasoline comprised 554 successes and 41 failures. Structure-based analyses retained 114 formulations built from 50 single-structure and 22 polydisperse represented materials. For the fixed Extra Trees/combined representation, pooled heptane success ROC–AUC was 0.812 for row-random, 0.813 for replicate-disjoint, 0.805 for formulation-disjoint, and 0.783 for component-set-disjoint evaluation. At the prespecified 0.55 Tanimoto stress threshold, 114 formulations formed seven families of sizes 87, 11, 10, 3, 1, 1, and 1; only five families had labeled heptane observations. A fixed FireSuppressionFP/Extra Trees leave-one-family-out stress test gave pooled ROC–AUC 0.557. The three unique historical family partitions gave macro, size-weighted, and pooled ROC–AUC values of 0.694, 0.572, and 0.557, respectively; model-seed repeats were not counted as independent partitions. Extinction-time prediction remained limited: formulation-disjoint MAE was 38.9 s with pooled R2 0.306, while the family stress test yielded MAE 45.8 s and pooled R2 0.076; the partition-macro family R2 was -0.105. Thus, extinction success generalized across unseen formulations and component sets within represented chemistry, but coarse-family extrapolation was unstable and poorly constrained. Results are retrospective associations for 19 cm liquid-pool fires, not prospective validation or evidence of wildland-fire performance.
Source: ChemRxiv
Author: Aris Tsai
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214703609