Structural Damage Prediction in a High-Velocity Urban Dam-Break Flood: Field-Scale Assessment of Predictive Skill

Dam safety and flood risk management programs are dependent on damage predictions that are difficult to validate and subject to considerable uncertainty. The 1963 Baldwin Hills dam-break flood caused high-velocity flows exceeding 5 m/s and structural failure of 41 wood-framed residences built in the mid-1940s, 16 of which were completely washed out. The flood is revisited here to examine the predictive skill and variability of established structural damage models when coupled with a hydraulic flood model that predicts parcel-scale depths and velocities. Two-way coupling is introduced so that predictions of structural failure affect localized flood predictions, which in turn affects damage predictions, in contrast to one-way coupling where structural failure has no impact on flood predictions. Two damage states defined by structural failure (Level 2) and washout (Level 3) are considered, along with 10 different structural damage models. One damage model considers flood depth alone, while the remaining nine consider a combination of depth and velocity defined by a constant discharge, energy, or force. Two-way coupling is shown to yield predictions with $\sim$30% higher skill and 10% higher false alarms than one-way coupled models. Hence, there is a tradeoff between skill and false alarms that favors two-way coupling. Predictive skill is also shown to be sensitive to the structural damage classification and the damage model. Depth-based damage predictions yield low predictive skill as expected; however, across the nine velocity-based damage models, skill varies from 50 to 78% for structural failure and from 79 to 95% for washout. These results reveal a similar level of predictive uncertainty from the hydraulic model implementation, damage model, and damage classification. Results also point to flow force as a good predictor of both moderate (Level 2) and severe (Level 3) levels of damage. Through calibration, force thresholds of 0.75 and 9.5 m³/s² are found to maximize model skill at 85 and 95% for Level 2 and 3 damage, respectively, and these values compare well with previously published thresholds for similar building types. Finally, the results reveal a model bias toward a relatively wide damage zone compared with the observations; therefore, the high skill predictions are accompanied by a high rate of false alarms.