Machine learning on problems from theoretical physics usually trains on data produced by a random generator: random graphs, random mutations, random polynomials. Speech science has studied the hazard of this for decades. Listeners and models trained or tested on synthetic stimuli, which are easy to control but strip out the structure of natural signals, learn cues that do not carry over to real input. This note brings that lesson to toric duality. In toric duality, several different gauge theories, each drawn as a periodic bipartite graph on a torus called a dimer model or brane tiling, describe the same physics because they sit on the same singular geometry. The question is whether neural networks that learn to recognize dual theories pick up the physics or pick up quirks of the generator that made their training data.
The claim is that the known failure of duality classifiers on unseen duality classes comes largely from unrealistic training data. It can be fixed by two changes. The first is to train on physically realized tilings instead of generator-produced quivers. The second is to supervise the network with invariants that come from the physics: R-charges, which are the charges fixed by the theory's superconformal symmetry and computed by a procedure called $a$-maximization, and the torus shape $\tau_R$ of the tiling drawn with angles set by those charges. Toric geometry suits this test because every label can be computed exactly, so the "natural" data are exact rather than noisy. That lets us measure precisely how far synthetic-data training misleads a model.
Core Claim and Positioning
The motivating negative result is Quiver mutations, Seiberg duality, and machine learning. It trained Naive Bayes and convolutional classifiers on adjacency matrices of quivers produced by random mutation in Sage (classes Q1–Q15). A quiver here is a directed graph whose nodes are gauge groups and whose arrows are matter fields. Binary same-class accuracy exceeded 90% and often reached 100%, and multiclass accuracy approached 100% once gauge ranks were added. Performance fell sharply on unseen mutation classes, and the file gives no numeric value for that regime.
The method comes from INVARIANTS OF TORIC SEIBERG DUALITY. That paper showed three things. The torus parameter $\tau_R$ of the isoradial embedding, a drawing in which every face sits in a circle of equal radius and edge angles encode R-charges, does not change under duality moves. It can be computed from the external points of the toric diagram through a closed "toric formula." The number field generated by the R-charges, together with its degree and rank over the rationals, is also invariant under duality. The two papers have never been combined. The classifiers see only discrete adjacency data. The invariants have never been used as training signal.
Speech work supplies the diagnosis and the evaluation design, not the application. Kitahara et al. show that synthetic voice-onset-time continua remove secondary cues. Reitsema et al. show that linear pitch continua cause artificial perceptual shifts. Lost in Transcription finds that a Whisper model's word error rate rises from 3.1% on clean LibriSpeech to 19.8% on synthetically disfluent speech. Topological Data Analysis for Speech Processing finds single attention heads that separate natural from synthetic speech almost perfectly, which shows that the synthetic/natural boundary is often trivially easy for a model to detect. We reuse that logic as a realism audit. If a simple discriminator can tell generator-produced quivers from physically realized ones, the synthetic training set is unrealistic in a measurable way.
What is new is the combination of three pieces: a benchmark of exact natural and synthetic duality data split so that whole geometries are held out; supervision with physically grounded invariants; and cue-conflict tests, borrowed from phonetics, that show which features a model actually uses.
Research Program
The natural corpus starts from toric diagrams, which are lattice polygons encoding the geometry. We take all such polygons up to a fixed area bound. For each, we obtain one consistent brane tiling, then enumerate its toric phases, meaning the dual tilings of the same geometry, by repeatedly applying the local graph move for Seiberg duality ("urban renewal"). We keep only phases in which every field has positive R-charge. Each tiling is labelled with four things. The first is its toric diagram, recomputed with the established forward algorithm (dimer to toric diagram) as an exact check. The second is its R-charges from $a$-maximization, written in the generalized incidence matrix form of Hanany et al.. The third is $\tau_R$ and the Klein $j$-invariant $j(\tau_R)$, a standard function of the torus shape that is unchanged by relabelling the torus. The fourth is the degree and rank of the R-charge number field.
Non-minimal phases where the isoradial embedding degenerates, such as those of $L^{2,2,2}$, are flagged and kept as a hard slice rather than dropped. The synthetic corpus copies the quiver-mutation generator and is matched to the natural corpus in node count and arrow count.
The model is a graph neural network over the tiling, or over the quiver alone for comparison with prior work. It is trained on duality classification with an auxiliary head that regresses per-edge R-charges and the pair $(\mathrm{Re}\,\tau_R, \mathrm{Im}\,\tau_R)$ up to modular equivalence. The hypothesis is that forcing the representation through quantities that duality leaves unchanged, and that are fixed by the physics rather than by the generator, makes duality-class structure carry over to geometries never seen in training.
The realism audit has three parts. The first is a natural-versus-synthetic discriminator, the analogue of spoof detection. The second is an attribution study asking which graph features predict the classifier's decisions. The third is cue-conflict test sets, modelled on cue-trading experiments. In these sets, generator-correlated surface features, such as mutation depth, maximum arrow multiplicity and field count, are put in opposition to true class membership. For example, dual pairs are built with very different field counts, and non-dual pairs with identical field counts.
Baselines
The naive reference is a majority-class and field-count heuristic that guesses duality from matching node and arrow counts. It tells us how much of any accuracy is available from surface statistics alone.
The strongest existing learned method is the adjacency-matrix CNN and Naive Bayes pipeline of Quiver mutations, Seiberg duality, and machine learning. It reports binary accuracy above 90%, often 100%, and near-100% multiclass accuracy with rank information. It is the right competitor because it is the only published learned approach to this exact task. It will be re-run with the same architecture, trained on the same number of examples as our model, with the same hyperparameter search budget. It will be evaluated on our held-out-geometry split, for which there is no reported result on Lacuna, and on its original within-class split to confirm we reproduce its numbers. Beating it means higher accuracy and area under the ROC curve on held-out geometries. The model must also not lose accuracy on the in-distribution split, and the gap must hold over several random seeds.
The key ablation is the same graph network trained on natural tilings without the auxiliary head. It has three variants. In the first, the auxiliary head regresses a continuous target of the same dimension that is not a duality invariant, such as edge degree statistics. In the second, it regresses a purely combinatorial invariant, the toric diagram's area and boundary-point count. In the third, the true R-charges are randomly shuffled across edges. These variants separate three possible sources of gain: natural data alone, any extra supervision, and supervision that is physically grounded. An exact-invariant lookup, described below, serves as an upper bound rather than a competitor.
Why not just…?
Why not just compute the invariants and skip learning? For tilings this is correct, and we concede it. The forward algorithm decides duality exactly, and $\tau_R$ follows from the toric formula of Hanany et al.. The goal is not a better duality decider. The goal is a clean measurement of how generator-produced training data misleads learned representations, in a domain where the truth is exact. That question matters for the many ML-for-physics settings where no oracle exists. We will also report how often $j(\tau_R)$ together with the number-field degree and rank fails to separate distinct toric diagrams. That completeness test is a small physics result in its own right, and it bounds what any invariant-driven model can achieve.
Why not just train on more synthetic data or use bigger models? The degradation reported by the mutation-classification study appeared as the number of classes grew, which suggests scale is not the missing ingredient. The speech evidence (Lost in Transcription) likewise shows that very large models still fail on off-distribution stimuli. We will settle this directly with a scaling curve of synthetic data size against held-out-geometry accuracy. If the curve keeps rising without levelling off, our thesis is weakened.
Why not just train on natural tilings and drop the physics targets? This is the real threat. Realistic data may explain most of the gain. The no-auxiliary ablation measures exactly this. We will report the result honestly even if the physics targets add nothing beyond natural data, because that would still show that the training data, not the architecture, drove the earlier failure.
Proposed Experiments
The decisive experiment is a two-by-two design crossing training source (synthetic or natural) with supervision (classification only, or classification plus invariant regression). Every cell is evaluated on held-out toric geometries, on the cue-conflict sets and on the degenerate non-minimal slice. The prediction is that synthetic-trained models will perform well in distribution, collapse toward the field-count heuristic on cue-conflict sets, and be separated from natural data by the discriminator with near-perfect accuracy. The prediction for the full model is that it will keep its accuracy on cue-conflict pairs.
A secondary check tests whether the auxiliary head's predicted $\tau_R$ stays constant along chains of duality moves, and whether its error grows on degenerate phases. A final, smaller transfer test applies the same audit to the knot–quiver setting of Kucharski and Noshchenko. There, quivers realized by actual knots are compared against random symmetric quivers, and the permutohedron families of equivalent knot quivers serve as the analogue of dual phases. This tells us whether the synthetic-data gap is specific to toric duality or general.
References
- Quiver mutations, Seiberg duality, and machine learning
- INVARIANTS OF TORIC SEIBERG DUALITY
- Perception of Incomplete Voicing Neutralization of Obstruents in Tohoku Japanese
- Perceptual Learning in Lexical Tone: Phonetic Similarity vs. Phonological Categories
- Lost in Transcription: Identifying and Quantifying the Accuracy Biases of Automatic Speech Recognition Systems Against Disfluent Speech
- Topological Data Analysis for Speech Processing
- Knot-quiver correspondence: a brief review