Abstract
<title>Abstract</title> <p>Chemically aerated dense concrete produced by aluminum powder addition exhibits a mechanical penalty that scalar porosity models cannot explain. In six mixes (0-5% aluminum by cement mass, w/c = 0.35), apparent porosity increased only from 3.24% to 6.18% while 28-day compressive strength collapsed by 56% (52.9 to 23.3 MPa); splitting tensile strength fell abruptly at 4% Al; and stoichiometric gas retention efficiency remained below 0.68% throughout. Three macroscopic anomalies emerged: a disproportionate 39% strength collapse at 1% Al, a non-monotonic tensile-to-compressive strength ratio peaking at 3% Al, and a porosity decrease at 4% Al coincident with a 27.4% tensile strength loss. Conventional strength-porosity models fit the data well (R² = 0.953-0.990) but cannot explain these anomalies, being scalar in porosity alone. U-Net segmentation of 60 polished cross-sections (12,718 characterized pores) resolves all three: the 1% Al collapse reflects a 7.3× pore-density increase combined with the lowest circularity in the series, multiplying Inglis stress-concentration sites; the f_t/f_c peak coincides with a local circularity maximum one dosage step earlier; and the 4% Al anomaly is confirmed as bubble coalescence by five convergent indicators. Results show that pore multiplication, not enlargement, governs porosity development, that gas retention rather than generation limits foaming efficiency, and that geometry-resolved descriptors not scalar porosity are required to explain and guide the design of aluminum-aerated dense concrete.</p>