Layered Compression of Micro-PIV Video That Keeps the Physics of Polymer Coiling

Engineers who push polymer solutions through rock to recover oil depend on long chains staying stretched out, because stretched chains make the water thick and springy enough to push oil evenly. Salty brine undoes this. Dissolved salt ions crowd around the charged groups on each chain and neutralize the repulsion that kept the chain extended. Chemists call this charge shielding. The chain then collapses into a small coil, and the fluid loses both its thickness and its springiness. The clearest evidence of what this does to flow comes from filming tracer particles moving through transparent pore-network chips, a technique called micro particle image velocimetry (micro-PIV), which recovers a velocity field from pairs of video frames. These videos are large, sparse, and highly redundant, which makes them a natural target for compression. The difficulty is that ordinary video codecs, and even learned flow estimators, remove exactly the features that reveal a collapsed polymer.

We propose a layered codec for micro-PIV of polymer flooding. It sends a small base layer that carries a smooth, mass-conserving velocity field. A second, sparse layer carries the "apparent divergence": places where in-plane flow appears to be created or destroyed because fluid is moving out of, or into, the imaged plane. A third layer carries the raw particle pixels for anyone who wants to check the result by eye. The central bet is that, for salt-coiled and sulfide-damaged solutions, the apparent-divergence layer is signal rather than noise. It should grow as charge shielding strengthens, and it should shrink for polymers designed to resist salt. A codec that keeps this layer should preserve the channeling and fingering statistics that decide whether a polymer is working, and it should need far fewer bits than codecs that optimize pixel fidelity.

Core Claim

Scalable video coding (a bitstream split into a required base layer and optional enhancement layers that refine it) has been adapted to serve machines first and humans second. Scalable Video Coding for Humans and Machines and its journal version Learned scalable video coding for humans and machines build the base layer for an object detector and use conditional coding in the enhancement layer to rebuild pixels. An Emerging Coding Paradigm VCM sends sparse skeleton points as the base layer and generates video from them. We reuse this machine-first layering. What neither work settles is what the "machine" layer should contain when the downstream task is physical measurement rather than recognition.

Our answer draws on a tension the file exposes. PIVNO reaches low error partly by penalizing any in-plane divergence of the estimated velocity, which assumes that 2D incompressible flow describes the imaged plane. Pore-Scale Flow Fields of the Viscosity-Lost HPAM Solution Caused by Sulfide Ion motivates the opposite view. It shows that elasticity loss in partially hydrolyzed polyacrylamide (HPAM, the standard flooding polymer) produces a roughly five-fold jump in the ratio of mainstream to margin velocity, eccentric radial throat peaks of 106.16 µm/s, and bimodal axial velocity profiles. These are channeling behaviors that exchange fluid with the pore depth the camera cannot see.

We therefore split the velocity field with a Helmholtz decomposition, which writes any 2D field as the sum of a swirling, mass-conserving part and a source/sink part. We then code the two parts in separate layers:

$$ \mathbf{u} = \nabla^{\perp}\psi + \nabla\phi, \qquad \nabla^2\phi = \nabla\cdot\mathbf{u}. $$

Here $\mathbf{u}$ is the measured in-plane velocity, $\psi$ is a scalar stream function coded coarsely in the base layer, and $\phi$ is a potential whose Laplacian equals the apparent divergence. That divergence is coded as a sparse map in the first enhancement layer. The testable claim is that the energy in $\phi$ is lower for a salt-free polymer, higher for a salt-coiled one, and highest for a sulfide-damaged one. The measurement must first subtract the $\phi$ energy seen for a Newtonian fluid of matched viscosity in the same chip, which captures artifacts from the chip geometry and from depth-averaging.

Research Program

The physics directions do more than supply an application. They supply the controlled variable. Salt concentration sets how strongly charges are shielded, which moves HPAM along a continuum from extended chains to coiled ones. The mechanism is the collapse into near-Newtonian coils described in Impact of polymer addition, electrolyte, clay and antioxidant on rheological properties. That same paper also shows formate salts preserving chain hydration, which gives a salt condition that should counteract coiling.

Against this continuum we set polymers built to exploit salt rather than suffer from it. One is the hydrophobically associating polymers of Study on Properties of Hydrophobically Associating Polymer in High Salinity Reservoirs, which hold or gain viscosity up to 80,500 mg/L. The other is an aromatic-associating copolymer in the spirit of the DT tackifier. This yields a grid of fluids whose predicted ordering of the $\phi$ energy is known before any compression is done. The ordering is a physics check on the representation, independent of any compression claim.

We would build a micro-PIV dataset using a sandstone-patterned chip and the fluorescent-tracer protocol of Yu et al. It would cover HPAM across a salinity sweep (0 to 80,500 mg/L), HPAM with 0 and 5 mg/L sulfide, formate-stabilized HPAM, an associating polymer, and a Newtonian glycerol–water control matched to each fluid's shear viscosity. Viscometry for each fluid anchors the rheology.

The codec has three parts. The base layer is a learned transform coder for $\psi$ at quarter resolution. The first enhancement layer entropy-codes a thresholded divergence map, conditioned on the base layer. The second enhancement layer conditionally codes particle images, using the decoded flow to warp the previous frame as its context, in the manner of the temporal context mining used by Hadizadeh and Bajić. The flow estimator that feeds the encoder is PIVNO retrained without its divergence penalty, so that apparent divergence is not suppressed before coding begins. Progressive truncation inside each layer borrows nested quantization from PLONQ.

Baselines

The naive reference point is to store raw frames losslessly and run PIVNO or classic cross-correlation on the decoded frames. This defines the target values of every diagnostic quantity, and the bit cost that any codec must undercut.

The strongest conventional competitor is single-layer VVC or HEVC applied to the raw particle video at matched bitrates, with the same retrained PIVNO run on the decoded output. It is strongest because these codecs are mature and hardware-supported, and the human–machine codec papers use them as the anchor. For the closest learned method, Hadizadeh and Bajić report a −16.8% average BD-Rate-mAP (the bitrate saving at equal detection accuracy) against VTM on SFU-HW-Objects-v1. The file contains no reported result on Lacuna for any codec evaluated on PIV or micro-PIV data, so the number we must beat here has to be measured.

The strongest learned layered competitor is the Hadizadeh–Bajić conditional two-layer design retrained on our data. Its base layer would be optimized for velocity endpoint error from PIVNO rather than detection mAP. This isolates whether our physics-motivated split beats a generic machine-optimized base layer at the same parameter count and training budget.

The key ablation replaces our split with a single divergence-free base layer. This means coding only $\psi$ from PIVNO trained with its original divergence penalty, at equal total bits. If this ablation matches our codec on the diagnostic quantities, the central claim fails.

All methods would see identical chips, fluids, and splits, receive the same GPU budget and hyperparameter search size, and have total bits counted across all layers they transmit. Beating a baseline means a lower BD-rate at equal error on the three diagnostics defined below, at no loss in endpoint error, across the salinity sweep.

Why not just…?

Why not just run PIV at the microscope and store the vector fields plus an HEVC copy of the video? This two-stream "simulcast" is the real threat, because it is simple and already separates machine data from human data. Our design must beat its total bit count. It must also preserve the diagnostics better than a simulcast whose vectors come from a divergence-penalized estimator. We would run simulcast as a baseline with vectors quantized to the same precision as our layers. If it matches us, the conclusion is that layering adds nothing, and only the Helmholtz split matters.

Why not just compress frames with a standard codec, or downsample them? MCFormer reports that sparse seeding already makes generic optical flow estimate zero motion where particles are missing. It also reports that downsampling merges tracer dots into gray blobs. The file gives no compressed-PIV numbers, so our rate sweep of VVC/HEVC is what settles the question.

Why not just keep PIVNO's divergence-free loss, since it improves accuracy on standard benchmarks? On synthetic 2D flows it does help: PIVNO reports an average endpoint error of 0.8 on Cylinder flow versus 1.4 for PIV-RAFT. On Backstep, however, it scores 1.9 versus 1.6 for PIV-RAFT, so the benefit is not universal even there. Those flows are truly 2D. The divergence-free ablation, together with the Newtonian control, tests directly whether the penalty erases the polymer-specific signal.

Proposed Experiments and Decisive Checks

Before running the physics experiments, we would verify that retraining PIVNO without its divergence penalty does not ruin accuracy. The check uses the PIVNO synthetic benchmarks and the MCFormer benchmark, where MCFormer reports an overall NEPE (normalized endpoint error) of 0.295.

On the new micro-PIV data, the decisive metrics are three diagnostics computed from decoded flow and compared with their values from lossless data: the mainstream-to-margin velocity ratio, the bimodality of throat axial velocity, and the Newtonian-subtracted $\phi$ energy. Each is plotted against total bits.

The first decisive check is purely physical. On lossless data, the Newtonian-subtracted $\phi$ energy should rise monotonically with salinity and with sulfide for HPAM, and should stay flat or fall for the associating polymer and the formate-stabilized fluid. If it does not, apparent divergence is measurement noise, and the proposal reduces to ordinary flow compression.

The second decisive check is the compression one. At equal bits, our codec should keep all three diagnostics within the lossless-data spread across the salinity sweep, at bitrates where VVC and the divergence-free ablation have already lost them.

A third check asks whether a classifier reading only the base layer and the first enhancement layer can recover the salinity class. The comparison is against the same classifier applied to decoded VVC frames at equal bits, which tells us whether the compressed stream remains a usable readout of charge-shielding state.