Monocular 3D bubble reconstruction: DNS-informed ray tracing and deep learning

Abstract

Bubble size, depth, and spatial distribution govern transport and uniformity in multiphase flows. A single shadowgraphic view cannot separate bubble diameter from focal-plane displacement, as image signatures depend on both, while multicamera systems add optical access, calibration, and cost. We develop a single-camera physics-guided framework combining direct numerical simulation (DNS), validated Monte Carlo ray tracing, preprocessing, and focus-conditioned multitask deep learning to recover three-dimensional bubble position and physical-equivalent diameter from two-dimensional shadowgraphs. DNS-resolved bubble label images rendered at multiple focal-plane settings. The ray tracer reproduces shadowgraphic signatures with errors below $3\%$ (intensity), $1\%$ (apparent size), $15\%$ (blur), $6\%$ (numerical aperture), and $4\%$ (edge-spread-function shape), running $\sim 30\times$ faster than the CPU implementation. The dataset has $900$ frames and $87,728$ labeled instances, partitioned by DNS timestamp to prevent leakage and expanded to over $10^5$ crops by broad-depth controlled-density rendering. A multitask network predicts signed depth and diameter from crops and focal-plane coordinates, achieving $\sim 97.6\%$ normalized depth accuracy on held-out data; a closed-loop DNS-to-image-to-reconstruction test of $180$ bubbles gives diameter error below $9.71\%$ ($\sim 90.29\%$ accuracy). On experimental images at four radial planes ($142$--$971$ bubbles each), PDF overlap is $68\%$--$92\%$ and log-PDF correlation is $0.72$--$0.87$. At the focal reference plane ($11,679$ bubbles), overlap is $\sim 86\%$, correlation is $0.93$, and full-distribution, $D_{99.8}$, and composite errors are below $7.7\%$, increasing with focal-plane distance and defining a recoverability limit. Without stereoscopic hardware, the framework offers a scalable diagnostic for dense bubbly flows.

arXiv comment51 pages, 11 figures