Syn2Co: Dual Applications in Vision & Surface Science

• Syn2Co is an acronym with two distinct interpretations: one for synthetic-to-contrast self-supervised vision pre-training using diffusion-based data and synthetic negatives, and one for a cobalt-catalyzed on-surface polymerization protocol for fluorographdiyne nanosheets.
• In computer vision, the Syn2Co framework blends diffusion-driven synthetic data augmentation with synthetic hard negatives in a momentum-encoder pipeline, improving representation learning with Vision Transformers.
• In surface chemistry, the Syn2Co protocol achieves high coupling efficiency and large-domain, single-layer fluorographdiyne nanosheets on Au(111) via precise cobalt activation and coronene templating.

Syn2Co has been used in two distinct ways in recent arXiv literature. In self-supervised computer vision, it denotes Synthetic-to-Contrast, a framework that combines synthetic data augmentation in image space with synthetic hard negatives in representation space for Vision Transformer pre-training on ImageNet-100 (Giakoumoglou et al., 2 Sep 2025). In surface chemistry and 2D conjugated polymers, it denotes a protocol for selective on-surface 2D covalent polymerization of fluorographdiyne nanosheets on Au(111) through the combination of cobalt catalysis and coronene templating (Shu et al., 30 May 2026). The shared label does not imply a shared methodology; the two usages address unrelated technical problems and operate at different levels of abstraction.

1. Nomenclature and domain-specific meanings

The term is best understood as an acronym with domain-dependent semantics rather than as a single canonical method.

Usage of “Syn2Co”	Domain	Core meaning
Synthetic-to-Contrast	Self-supervised vision	Synthetic data plus synthetic hard negatives in a momentum-encoder InfoNCE pipeline
Syn2Co protocol	Surface synthesis	Cobalt catalysis plus coronene templating for fluorographdiyne growth on Au(111)

In the vision usage, Syn2Co is explicitly expanded as Synthetic-to-Contrast and is designed for contrastive self-supervised learning with DeiT-S and Swin-T backbones (Giakoumoglou et al., 2 Sep 2025). In the surface-science usage, Syn2Co refers to an on-surface polymerization protocol that synthesizes single-layered fluorographdiyne nanosheets up to $60\times 60\ \mathrm{nm}^2$ on Au(111) (Shu et al., 30 May 2026).

A common source of confusion is that acronym-based search can also retrieve the nearby term SyCo, short for Synthetic Coordinate Embedding, a molecular graph generation framework in latent Euclidean space rather than a Syn2Co method (Ketata et al., 2024). This suggests that citation by arXiv identifier is especially important when the acronym alone is ambiguous.

2. Syn2Co in self-supervised vision

In the computer-vision literature, Syn2Co targets the standard contrastive objective of learning an encoder $f:X\to\mathbb{R}^d$ such that positive pairs are nearby in feature space while negatives are far apart. Its defining move is to combine two “faking” strategies: synthetic data augmentation in image space and synthetic hard negatives in feature space, both inserted into a momentum-encoder, queue-based InfoNCE pipeline for Vision Transformers (Giakoumoglou et al., 2 Sep 2025).

The synthetic-data component uses a class-conditional diffusion model, Relay Diffusion [22], trained to clone ImageNet-100’s class distribution. The resulting synthetic set is $X_s \approx 130{,}000$ images, stated as one per original ImageNet-100 sample. Pre-training then mixes real and synthetic images by sampling a batch in which a fraction $\alpha\in[0,1]$ comes from real data $X$ and the remainder $1-\alpha$ from $X_s$. The boundary cases are explicit: $\alpha=1.0$ is pure real pre-training and $\alpha=0.0$ is pure synthetic pre-training.

The synthetic-negative component starts from a standard momentum-encoder pair $(f_q,f_k)$, temperature $f:X\to\mathbb{R}^d$0, and a queue $f:X\to\mathbb{R}^d$1 of size $f:X\to\mathbb{R}^d$2. For each query $f:X\to\mathbb{R}^d$3, cosine similarities $f:X\to\mathbb{R}^d$4 are computed for all $f:X\to\mathbb{R}^d$5, after which the top-$f:X\to\mathbb{R}^d$6 hardest real negatives are selected. Synthetic negatives are then produced by applying an overview function $f:X\to\mathbb{R}^d$7 to these hard negatives. The paper lists six synthesis modes from SynCo [12], including interpolation, extrapolation, mixing, jittering, perturbation, and adversarial variants. Each synthetic negative is normalized as $f:X\to\mathbb{R}^d$8, and the final negative pool becomes the union of real and synthetic negatives.

The resulting loss augments the usual InfoNCE denominator with synthetic negatives: $f:X\to\mathbb{R}^d$9

The architecture scope is narrow and explicit: DeiT-Small (21 M parameters) and Swin-Tiny (29 M parameters). The evaluation protocol is equally explicit: linear probing, with the encoder frozen and a single linear classifier trained for 100 epochs, reporting top-1/top-5 accuracy on the ImageNet-100 validation set.

3. Training dynamics, hyperparameters, and empirical behavior in vision

The training loop samples $X_s \approx 130{,}000$0 from $X_s \approx 130{,}000$1 and $X_s \approx 130{,}000$2 from $X_s \approx 130{,}000$3, applies two augmentations per image, computes query and key embeddings, mines the top-$X_s \approx 130{,}000$4 hard negatives from the queue, synthesizes $X_s \approx 130{,}000$5 feature-space negatives, evaluates $X_s \approx 130{,}000$6, updates the online encoder, performs the momentum update $X_s \approx 130{,}000$7, and refreshes the queue (Giakoumoglou et al., 2 Sep 2025). The key hyperparameters listed are $X_s \approx 130{,}000$8, $X_s \approx 130{,}000$9, $\alpha\in[0,1]$0, $\alpha\in[0,1]$1, $\alpha\in[0,1]$2, $\alpha\in[0,1]$3, batch size $\alpha\in[0,1]$4, and epochs $\alpha\in[0,1]$5. Representative values in the paper include $\alpha\in[0,1]$6, $\alpha\in[0,1]$7, $\alpha\in[0,1]$8, $\alpha\in[0,1]$9, $X$0, batch sizes $X$1, and $X$2.

The reported linear-probe results on ImageNet-100 show that the gains are architecture-dependent.

Method	DeiT-S Top-1	Swin-T Top-1
DINO	79.41	81.78
MoBY	79.36	83.90
Syn1Co (data only, 152 400)	81.86	83.68
Syn1Co-Neg	78.96	84.04
Syn2Co (full, 152 400)	82.12	83.70

The main quantitative highlight is that DeiT-S gains $X$3 Top-1 over MoBY when using both synthetic data and synthetic negatives at 200 epochs. For Swin-T, the paper reports modest gains ($X$4) from synthetic negatives alone, while synthetic data adds limited benefit. The study also reports that a purely synthetic pre-training run ($X$5) is within $X$6 of the fully real regime in linear-probe accuracy, using this as an empirical proxy for diversity.

The paper’s own analysis is cautious. It states that low-resource regimes benefit more from mixing in diffusion-generated images, that extended training extracts more signal from synthetic clones, and that datasets with strong class conditioning are easier to clone. It also states that synthetic negatives are computationally cheap once the queue is built, but that $X$7 and $X$8 must be tuned per architecture because excessively many synthetic negatives or too much hardness can hurt. The practical recommendations are correspondingly conservative: tune $X$9 on a held-out set, use moderate hardness $1-\alpha$0 and synthetic ratio $1-\alpha$1, extend pre-training by $1-\alpha$2 epochs when relying heavily on synthetic data, and monitor representation collapse via inter-class cosine-similarity distributions.

4. Syn2Co as a protocol for fluorographdiyne nanosheet synthesis

In the surface-science literature, Syn2Co denotes a selective on-surface 2D covalent polymerization protocol for synthesizing single-layered fluorographdiyne nanosheets on Au(111) through the combination of cobalt catalysis and coronene templating (Shu et al., 30 May 2026). The target material belongs to the broader family of graphdiyne derivatives with sp-sp$1-\alpha$3 hybridized skeletons.

The reagents and setup are specified in procedural detail. The precursor is 1,3,5-tris(chloroethynyl)-2,4,6-trifluorobenzene (tFtCEB), degassed at 333 K. The template is coronene, which sublimes at $1-\alpha$4. The catalyst is Co metal evaporated to 0.05–0.10 ML coverage using an e-beam evaporator. The substrate is an Au(111) single crystal, cleaned by Ar$1-\alpha$5 sputtering/annealing under UHV $1-\alpha$6, with base pressure during depositions $1-\alpha$7.

The synthesis sequence is reported stepwise. First, 0.7 ML tFtCEB is deposited onto Au(111) held at 250 K, forming self-assembled alkynyl–Au–alkynyl dimers. Annealing to 473 K for 5 min yields a large-domain honeycomb sp-MON network with Csp–Au–Csp linkages. Coronene may then be deposited as an optional step at 0.1 ML and 300 K, producing coronene-embedded sp-MON. Next, 0.05–0.10 ML Co is deposited at 250 K onto the (coronene-embedded) sp-MON. A further anneal to 473 K for 5 min produces partial demetallization intermediates, stated as $1-\alpha$8 C–C coupled. Final annealing to 553 K for 5 min yields large-domain fluorographdiyne nanosheets.

The reported outcome metrics are explicit: coupling efficiency $1-\alpha$9 for the Csp–Au $X_s$0 Csp$X_s$1–Csp$X_s$2 conversion, hexagon-formation selectivity of $X_s$3 without coronene and $X_s$4 with coronene, and typical covalent domains up to $X_s$5. These data place the protocol within the larger effort to obtain large-domain, regular, single-layered graphdiyne-type sheets on surfaces, a problem the paper identifies as a longstanding synthetic challenge.

5. Cobalt activation, coronene templating, and reaction energetics

The mechanistic role of cobalt is formulated as $X_s$6–$X_s$7 coordination between Co and the alkynyl $X_s$8 system: $X_s$9 with $\alpha=1.0$0 quantifying Co(d)–C$\alpha=1.0$1C($\alpha=1.0$2) coupling (Shu et al., 30 May 2026). The reported interpretation is that strong $\alpha=1.0$3–$\alpha=1.0$4 coupling transforms a robust Csp–Au bond into a weaker Csp$\alpha=1.0$5–Au bond, facilitating demetallization and C–C coupling.

The bond-conversion step is written as

$\alpha=1.0$6

In the DFT-optimized structures, the sp-hybridized C–C(Au) bond length in IS2 is $\alpha=1.0$7, whereas the Co-activated (sp$\alpha=1.0$8) C–C(Au) bond length in IS3 is $\alpha=1.0$9. The uncatalyzed surface-stabilized pathway is reported as

$\alpha=0.0$0

while the Co-catalyzed route is

$\alpha=0.0$1

The associated reaction-coordinate data at $\alpha=0.0$2 are $\alpha=0.0$3 for the catalyzed route, compared with $\alpha=0.0$4 for the uncatalyzed pathway. The paper summarizes this as cobalt lowering the Csp–Au $\alpha=0.0$5 Csp$\alpha=0.0$6–Au barrier by $\alpha=0.0$7.

Coronene acts as a template that matches the fluorographdiyne lattice and forms up to six C–H$\alpha=0.0$8F hydrogen bonds per coronene, with $\alpha=0.0$9 total. The paper’s mechanistic picture states that coronene “locks” partial oligomer radicals via noncovalent anchors, thereby reducing lateral diffusion and rotation. This favors six-membered motifs and suppresses kinetically trapped 5/7-MR defects. The same section reports 0.2–0.4 eV per oligomer stabilization in the summary statement. Domain-size distributions also shift toward larger domains with coronene: without coronene, the highest fractions lie in the 82–152 and 152–252 $(f_q,f_k)$0 bins; with coronene, weight shifts into the 252–352, 352–452, and larger bins.

6. Atomic-scale characterization and relation to adjacent terminology

The protocol is supported by atomically resolved microscopy and spectroscopy (Shu et al., 30 May 2026). In STM, the sp-MON honeycomb cell constant is $(f_q,f_k)$1, and the final fluorographdiyne lattice retains a hexagonal cell with $(f_q,f_k)$2 periodicity. In nc-AFM with a CO tip, the C$(f_q,f_k)$3C$(f_q,f_k)$4C$(f_q,f_k)$5C diacetylene linkages appear as bright rods. Bond lengths extracted from AFM contrast are given as $(f_q,f_k)$6 for the C$(f_q,f_k)$7C triple bond and $(f_q,f_k)$8 for a bond with double character.

The spectroscopy is similarly specific. At 5 K, the fluorographdiyne valence-band resonance appears at $(f_q,f_k)$9 and the conduction-band resonance at $f:X\to\mathbb{R}^d$00, corresponding to a measured bandgap of $f:X\to\mathbb{R}^d$01. For coronene, the HOMO is at $f:X\to\mathbb{R}^d$02 and the LUMO at $f:X\to\mathbb{R}^d$03, giving a bandgap of $f:X\to\mathbb{R}^d$04. The spatial maps localize the $f:X\to\mathbb{R}^d$05 density on diacetylene linkages and the $f:X\to\mathbb{R}^d$06 density on aromatic rings; the coronene HOMO/LUMO maps match free-molecule LDOS with a six-lobed flower HOMO and ring-shaped LUMO.

Set against the vision usage, these measurements underscore that the two Syn2Co usages are unrelated except for the acronym. In the vision framework, “synthetic” refers to diffusion-generated images and synthetic feature-space negatives (Giakoumoglou et al., 2 Sep 2025). In the surface-science protocol, Syn2Co refers to Co $f:X\to\mathbb{R}^d$07–$f:X\to\mathbb{R}^d$08 activation and coronene-regulated ring formation (Shu et al., 30 May 2026). A plausible implication is that “Syn2Co” should be treated as a context-sensitive label rather than as a stable cross-disciplinary term. The nearby acronym SyCo, introduced as Synthetic Coordinate Embedding for molecular graph generation in latent Euclidean space, reinforces this point: it is adjacent in spelling but methodologically distinct (Ketata et al., 2024).