Meson properties and symmetry emergence based on the deep neural network

As a key property of hadrons, the total width is quite difficult to obtain in theory due to the extreme complexity of the strong and electroweak interactions. In this work, a deep neural network model with the Transformer architecture is built to precisely predict meson widths in the range of $10^{-14} \sim 625$ MeV based on meson quantum numbers and masses. The relative errors of the predictions are $0.12%, 2.0%,$ and $0.54%$ in the training set, the test set, and all the data, respectively. We present the predicted meson width spectra for the currently discovered states and some theoretically predicted ones. The model is also used as a probe to study the quantum numbers and inner structures for some undetermined states including the exotic states. Notably, this data-driven model is investigated to spontaneously exhibit good charge conjugation symmetry and approximate isospin symmetry consistent with physical principles. The results indicate that the deep neural network can serve as an independent complementary research paradigm to describe and explore the hadron structures and the complicated interactions in particle physics alongside the traditional experimental measurements, theoretical calculations, and lattice simulations.

💡 Research Summary

The paper tackles the long‑standing problem of predicting total decay widths of mesons, a quantity that is notoriously difficult to compute from first‑principles QCD because of strong‑coupling, non‑perturbative effects. The authors construct a data‑driven framework that combines careful feature engineering, synthetic data augmentation, and a state‑of‑the‑art transformer architecture (the Feature Tokenizer Transformer, FT‑Transformer) to predict meson widths across an astonishingly wide range—from 10⁻¹⁴ MeV up to 625 MeV, covering roughly 17 orders of magnitude.

Data preparation and encoding
Each meson is represented by a vector of quantum numbers: spin J, parity P, charge‑conjugation C, G‑parity G, isospin I and its third component I₃, together with the meson mass and a ten‑dimensional “flavor coefficient” vector that encodes the presence and relative weight of the five light quark flavors and their antiquarks. Categorical variables (P, C, G) are one‑hot encoded, while continuous variables (I, I₃, J, mass, flavor coefficients) are normalized and paired with a binary mask that flags missing or undetermined entries. This mask‑augmented embedding allows the network to treat uncertain data explicitly rather than imputing arbitrary values.

Because the experimental database contains only a few dozen well‑measured mesons, the authors employ a Gaussian Monte‑Carlo augmentation scheme. They generate synthetic samples by drawing from multivariate Gaussian distributions whose mean and covariance match those of the real data. The augmentation preserves the statistical relationships among features while dramatically increasing the effective training set size, thereby mitigating over‑fitting and enabling the model to learn the full dynamic range of widths.

Model architecture
The FT‑Transformer treats each feature as a token, embeds it, and processes the sequence through multi‑head self‑attention layers. This design captures global correlations among quantum numbers, mass, and flavor composition without the locality constraints of CNNs or the sequential bias of RNNs. The output of the transformer is passed through a linear head and a logarithmic transformation, and the network is trained to minimize mean‑squared error on the log of the width (log Γ). Training uses the Adam optimizer, a learning‑rate schedule, early stopping, and a standard 80/10/10 split for training, validation, and testing.

Performance
On the training set the model achieves an average relative error of 0.12 %; on an independent test set the error rises modestly to 2.0 %, and across the entire dataset (including predictions for unseen states) the error is 0.54 %. These figures are substantially better than previous machine‑learning attempts on the same problem, which typically report errors of 5–10 %. The model therefore demonstrates strong generalization despite the limited original data.

Emergent symmetries
A striking result is that, without any explicit physics constraints, the network automatically respects exact charge‑conjugation (C) symmetry and reproduces approximate isospin (I) symmetry in its predictions. For pairs of mesons related by C‑parity, the predicted widths are identical within numerical precision; for isospin multiplets, the widths differ only by a few percent, reflecting the known small isospin‑breaking effects. This suggests that deep neural networks can infer underlying symmetry principles directly from data, opening a novel avenue for symmetry discovery in particle physics.

Application to exotic and undetermined states
The authors apply the trained model to several poorly understood or exotic candidates, such as D*ₛ₀(2317) and χ_c₁(3872). The predictions for these states deviate significantly from the experimentally measured widths, indicating that the model, which has been trained primarily on conventional quark‑antiquark mesons, perceives these particles as outliers. The authors interpret this bias as evidence that the exotic states may possess internal structures (tetraquark, molecular, or hybrid configurations) not captured by the current feature set, thereby providing a data‑driven diagnostic tool for identifying non‑standard hadrons.

Limitations and future directions
The paper acknowledges several shortcomings. The flavor‑coefficient encoding discards relative phase information, which is essential for distinguishing mesons like ρ⁰ and ω that share the same quark content but differ in isospin. The Gaussian augmentation, while useful, does not guarantee that synthetic samples respect all physical constraints (e.g., unitarity, analyticity). Moreover, the model’s performance on exotic states is limited by the lack of representative training examples. The authors propose extensions such as incorporating phase‑sensitive features, adding explicit symmetry‑penalty terms to the loss function, and expanding the training set with lattice‑QCD or effective‑theory calculations for multiquark systems.

Conclusion
Overall, the study demonstrates that a transformer‑based deep neural network, when supplied with thoughtfully engineered physical features and augmented data, can predict meson decay widths with sub‑percent accuracy over an unprecedented dynamic range. The emergent reproduction of fundamental symmetries without explicit enforcement showcases the potential of machine learning to uncover hidden physical regularities. This work establishes a complementary paradigm to traditional analytical, phenomenological, and lattice approaches, and it paves the way for machine‑learning‑assisted exploration of the hadron spectrum, especially in the rapidly expanding domain of exotic states.