Reconstructing the Fluid–Structure Interaction of an Earthquake-Triggered LP-Gas Deflagration and Progressive Collapse in a Two-Storey Shopping Mall in Kumamoto, Japan, on 28 July 2026

Abstract

On 28 July 2026, approximately 80 minutes after the Mj 7.1 Kumamoto earthquake (maximum JMA seismic intensity 7), a large gas explosion destroyed part of a two-storey steel-frame shopping mall in Kashima Town, Kumamoto, Japan. Seven fatalities had been reported as of 30 July 2026. Public reporting indicates that the earthquake damaged the supply piping from an outdoor bulk LP-gas vessel and that leaked propane accumulated in a second-floor retail zone before igniting. This paper reconstructs the event using a three-stage fluid–structure interaction (FSI) chain built entirely from open-source tools: (i) a 3-D premixed-deflagration CFD model of an approximately 330 kg stratified propane cloud in a 160 m section of the mall (constructed with OpenFOAM XiFoam using Weller b–Xi flame-wrinkling closure and k–omega SST) ; (ii) a quasi-two-way venting FSI scheme in which 255 breakable envelope and floor panels are monitored during the run and opened through a deterministic mesh-rebuild restart when their failure pressure is exceeded, with capacities derived from Japanese design references and degraded by an earthquake damage factor; and (iii) an event-driven progressive-collapse and debris/dust dynamics model that distinguishes a column-first storey-crush (pancake) core from a hinge-type slab-unzipping periphery. The effective turbulent burning velocity is calibrated using two observables: the contiguous facade-stripping extent and the near-frame-one envelope failure, which are therefore calibration targets rather than validation metrics. With that single calibration, the reconstruction independently reproduces the absence of a sustained external fireball, the radial debris scatter on the tank-side facade, the localized roof breach above the ignition zone, the mixed crush/draped collapse morphology, and a dust-front speed of 8.8 m/s against a measured 13 m/s (band 8–17 m/s). The debris model is further tested, without any adjustment, against the ballistics recorded in the footage: it reproduces the observed launch-speed/apex pair, and the back-calculated impulse duration of 22 ms agrees with the 25 ms assumed a priori. However, it does not reproduce the longest observed throw distances, a deficiency reported here rather than removed by tuning. A uniform 0.5 m grid-refinement study shows that the facade-stripping extent and the sustained compartment overpressure are essentially unchanged, while the near-kernel peak, the external blast spike, and the local failure counts are not convergent and are reported with that caveat. An ignition-location sensitivity run, in which the kernel is moved to the back-of-house corridor position suggested by post-event drone imagery, leaves the headline conclusions — south-side damage concentration, a roof breach above the epicentre, and few-kilopascal interior loads acting on seismically weakened members — unchanged. A numerical pitfall is that the subsetMesh utility of OpenFOAM-v2606 assigns the constraint type “empty” to newly exposed patches, which silently turns off momentum solution on 3-D meshes. All findings are hypotheses for the official investigation and imply no attribution of responsibility. On the grounds of damage magnitude, impulse, and debris velocity, the reconstruction supports a fast deflagration rather than a cloud detonation, while a transient, localized transition cannot be excluded.