Locating Parasitic Phases in Laser-Grown Films from Area-Averaged Labels

Pulsed laser deposition (PLD) grows thin crystalline films by firing laser pulses at a solid target. Each pulse vaporizes a small amount of material into a plasma that lands on a heated crystal. It is the standard way to make complex oxides with the right ratio of elements. The films often fail locally, though. Patches of an unwanted crystal structure (a "secondary phase") nucleate during growth or annealing, and molten droplets splash onto the surface. The measurements that tell a grower whether a film is good are slow and blurry in space. X-ray diffraction (XRD) gives a secondary-phase fraction averaged over a spot millimetres wide, and transport gives one critical temperature per film. Surface imaging is fast and sharp, but no one labels it pixel by pixel.

This note proposes to train an image model that first predicts where the parasitic phase sits on a micrograph and then uses that map to decide how bad the film is. The only supervision is the coarse XRD fraction: the model's predicted map, averaged over the X-ray footprint, must reproduce the measured fraction. The architecture comes from deepfake forensics, where networks first predict a manipulation mask and then classify only from the masked region so they cannot rely on background shortcuts. The supervision trick comes from the physics: an area-averaged measurement fixes how much of the mask should be "on" without saying where. The claimed payoff is threefold. The model should produce phase maps no instrument gives cheaply. It should predict film quality on new growth runs and new substrates better than whole-image regressors, which latch onto thickness and position cues. It should also not mistake droplets for phase domains.

Core Claim

Localization-guided classification grounds a global decision in a predicted spatial mask. In forensics it has repeatedly beaten unguided classifiers under distribution shift. On the Detection of Digital Face Manipulation regresses a manipulation attention map and multiplies it into the backbone features, lifting Celeb-DF AUC from 52.2% to 71.2%. Detect and Locate supervises a semantic mask and a noise-residual map that gate classification. Locate and Verify steers a classifier with predicted local attention and an extra noise-residual input stream. These papers supply the method. They all assume either dense ground-truth masks or, in the semi-supervised patch-similarity scheme of Locate and Verify, landmark anchors specific to faces, and the authors acknowledge that the semi-supervised version trails full masks.

Thin-film materials have neither dense masks nor landmarks. They do have something faces lack: a physical measurement of the fraction of the area that is defective. Our claim is that this fraction is sufficient supervision. Combined with mask-gated pooling, it makes the localization branch learn the parasitic phase itself rather than correlated nuisances.

Concretely, let $m_\theta(x)\in[0,1]$ be the predicted secondary-phase probability at pixel $x$. Let $A_j$ be the set of image pixels inside the footprint of XRD measurement $j$, $f_j$ its measured secondary-phase fraction, and $\hat y$ the film-quality prediction made from features pooled with weights $m_\theta$. We train on

$$ \mathcal{L} = \ell(\hat y, y) + \lambda \sum_j \Big|\tfrac{1}{|A_j|}\textstyle\sum_{x\in A_j} m_\theta(x) - f_j\Big| + \mu \,\mathrm{H}(m_\theta), $$

where $\ell$ is the loss on the film-level label $y$ (phase-pure versus segregated, or superconducting versus not) and $\mathrm{H}$ is the mean pixel entropy, which pushes the mask toward near-binary values.

The proportion term alone is degenerate: a uniform map at value $f_j$ satisfies it. Two things break the degeneracy. The entropy term rules out a flat grey map. Overlapping footprints with different fractions across a combinatorial wafer, or across an annealing series, force the model to place domains consistently across measurements. This is learning from label proportions married to mask-gated classification, and as far as the research file shows, neither half has been applied to PLD films.

The physics motivates the target. Influence of oxygen reduction on Sr1−xLaxCuO2 thin films (Influence of oxygen reduction on the structural and electronic properties of electron-doped Sr1−xLaxCuO2 thin films) shows that a parasitic "long-c" phase (c ≈ 3.62 Å) competes with the superconducting phase (c < 3.41 Å) during vacuum reduction. It forms extensively on DyScO3 substrates and narrows the usable annealing window, yet it is characterized only by area-averaged θ–2θ scans. Those films were grown by sputtering. The same diagnosis problem applies to PLD oxides, and that is the regime we target.

Research Program

The work has three parts.

A controlled synthetic test. We render film-like micrographs with known secondary-phase masks, droplets and thickness-dependent contrast gradients. From the known masks we compute footprint fractions and train only on those fractions. This tells us whether proportion supervision recovers masks at all, and how many overlapping measurements it needs, before any furnace time is spent.

A PLD data campaign designed around the proportion constraint. We grow complex-oxide films on two substrates with deliberately varied post-anneal conditions, so that secondary-phase fractions span roughly zero to majority. Each film is imaged by scanning electron microscopy (SEM) and atomic force microscopy on a registered grid. We take microbeam XRD maps with overlapping footprints, plus one film-level functional label.

We also re-image an existing style of combinatorial wafer. Performance Prediction of High-Entropy Perovskites built a composition-spread LSMCF library by combinatorial PLD and mapped XRD, Raman and electrochemistry at 321 points. That gives a ready template for many overlapping measurements whose fractions vary smoothly across a single wafer.

A small held-out subset of regions receives dense phase masks from fine-step microbeam XRD or electron backscatter mapping. These masks are used only for evaluation.

The model. An Xception or ResNet backbone carries a direct-regression attention head in the style of Dang et al. Mask-gated pooling feeds the classifier. A second input stream carries high-pass noise residuals, following Locate and Verify, because fast SEM imaging trades signal-to-noise ratio for throughput, as documented in the SEM defect-inspection review. Droplets get their own cheap label: a few hundred hand-drawn polygons, which are easy to draw because droplets are visually distinct. This lets us test whether the phase mask learns to exclude them.

We expect a tension between local detail and receptive field. SPAN argues that global attention washes out fine boundaries, while the face-forensics models regress masks from deep, wide-context features. We treat local-pyramid versus deep-feature mask heads as an explicit ablation rather than assume an answer.

Baselines

Naive references. The do-nothing reference predicts every film's label and phase fraction from its growth recipe and substrate alone, with no imaging. If imaging adds nothing over the recipe, the project has no reason to exist. A second naive reference is the tabular route of the high-entropy perovskite study: a Random Forest on point-wise XRD and Raman features. That paper reports R² up to 0.98 for composition and 0.97 for performance on its 321-point library. We rerun it on our measurement points with identical splits. It sets the bar for any image model that claims to predict functional outcome.

Strongest unguided competitor. A whole-image convolutional network with the same backbone regresses the footprint fraction and the film label directly, with no mask. This mirrors the unguided Xception baselines in forensics that our guided model must beat.

Two-stage pipeline. We include the two-stage pipeline of Deep Learning-Based Defect Classification and Detection in SEM Images: an unsupervised U-Net denoiser followed by RetinaNet. That pipeline reports 81.6% mAP for a three-backbone ensemble versus 78.8% for ResNet152 alone, but only 52.0% AP on the subtle probable-nano-gap class. Its detector needs box labels, so we give it boxes derived from the dense evaluation masks of a training-only subset. This is a supervision advantage we state openly.

Fairness and ablations. All learned models share the backbone, training images, a 30-trial hyperparameter search and five seeds. The ablations that isolate the key ingredient are three. In the first, the mask is trained with the proportion loss but not used for gating. In the second, gating uses a freely learned attention map with no proportion loss. The third is an oracle trained with dense masks, which bounds how much the weak supervision costs.

What counts as winning. On held-out growth runs and on the held-out substrate, the proposal beats a baseline only if two things hold. Its phase-fraction error and film-label AUC must be better by more than twice the seed-to-seed standard deviation. Its mask intersection-over-union on the densely annotated regions must be clearly above the threshold baseline described below.

Why not just…?

Why not just regress the XRD fraction with a standard CNN? This is the real threat. On in-distribution splits it may well match us. Our bet concerns shift. Film contrast varies with thickness, and the PLD plume's angular spread makes thickness track wafer position. An unguided regressor can therefore score well by reading position cues. The forensics evidence (52.2% versus 71.2% cross-dataset AUC in Dang et al.) shows that guidance pays off under shift, not in-distribution. The held-out-substrate and held-out-run splits test exactly this. If the gap closes there too, guidance is unnecessary and we will report that.

Why not just threshold image contrast and count pixels? If the secondary phase were visibly distinct, simple contrast thresholding would localize it for free. We run Otsu thresholding and k-means on the residual-filtered images, calibrated to the footprint fractions. If these reach mask quality comparable to ours on the dense evaluation regions, the learned localizer is redundant. In that case we would pivot to using thresholded masks as gates and drop the proportion loss. We expect thresholding to fail because droplets and thickness gradients produce contrast similar to phase domains, and the droplet-overlap analysis below measures exactly that.

Why not just hand-label masks and train a segmentation network? Dense labels of irregular, low-contrast domains are the bottleneck everywhere in the file. Even the resist SEM dataset labeled only 1,324 of 5,465 images, and that was with boxes. The dense-mask oracle tells us how much the proportion-only model gives up. If the oracle is far ahead and a few dozen masks close the gap, the honest conclusion is "label a little". We would then report the label-count curve rather than claim weak supervision wins.

Why not just use the tabular Random Forest? It is accurate where full multi-instrument characterization exists. Its own authors note the cost: slow, multi-modal mapping with interpolation between instrument grids. It also gives no spatial information about where a film fails. Our comparison asks whether cheap images plus sparse XRD match its prediction quality while adding maps.

Proposed Experiments and Decisive Checks

Synthetic test. We measure mask intersection-over-union against the number and overlap of footprints per image. The idea is viable only if masks become accurate with footprint densities achievable in a single day of microbeam XRD.

Domain shift. We train on one substrate and test on the other, mirroring the KTaO3 versus DyScO3 contrast in the SLCO study, and separately hold out whole growth runs. We report fraction error, film-label AUC and mask intersection-over-union against all baselines above.

Droplet overlap. This check is specific to PLD. We measure how much of the predicted phase mask lands on annotated droplets. A guided model that has learned the phase should put markedly less mass there than the attention maps of the unguided network.

Public microscopy sanity check. We convert the 32 nm resist SEM boxes into image-level labels and test whether mask-gated classification trained without boxes approaches the detection pipeline's per-class numbers, especially the 52.0% probable-nano-gap AP. This tells us whether localization guidance transfers to microscopy at all, independent of our own films.

Kill condition. The project is falsified if, on held-out substrates, the gated model's advantage over the plain regressor and over thresholding is within seed noise. In that case the area-averaged physics label is not enough to teach the model where defects are.