Machine learning classification of dynamical phases in a driven nonlinear Su-Schrieffer-Heeger lattice
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: We're looking at the paper: "Machine learning classification of dynamical phases in a driven nonlinear Su-Schrieffer-Heeger lattice".
Mira: Here's the gist of the paper. Driven nonlinear photonic lattices can settle into steady, symmetry-broken, periodic, or chaotic states Mapping where each occurs in parameter space is key to device design The authors apply four classifiers to the same coherently driven Su-Schrieffer-Heeger (SSH) chain of Kerr resonators Using a converged heuristic diagram as the reference, the CNN reaches 98.0% held-out accuracy, while the two label-free methods reach adjusted Rand indices (ARIs) of 0.916 and 0.947 without seeing any labels Phase Identification Methods The study compares a hand-built heuristic classifier, a supervised convolutional neural network (CNN), an unsupervised library-building scheme based on augmented dynamic mode decomposition (aDMD), and an unsupervised Gaussian mixture model (GMM) on four dynamical invariants The heuristic classifier serves as the reference for all other methods, requiring explicit rules for each phase The supervised CNN is trained on 70% of the heuristic labels and achieves a held-out accuracy of 97.96% The two label-free methods, aDMD library building and aDMD + GMM, return clusters whose physical meaning must be established afterwards Performance Comparison The comparison shows that the CNN gives the closest match to the reference, but only because it is trained on it The two label-free methods need no prior description of the phases and recover the main boundaries without labels, but their clusters must be interpreted afterwards Both separate the stable and bistable phases and the symmetric from the symmetry-broken oscillations, and both largely merge periodic and chaotic switching The aDMD library separates symmetric from symmetry-broken oscillations, which the heuristic misses, though it cannot distinguish periodic from chaotic switching Method Strengths and Limitations The CNN is efficient for extending known classifications to new parameter points but cannot report a phase that is absent from its training labels The aDMD + GMM method uses four scalar invariants to cluster trajectories, and the floor on the reconstruction error is essential for good performance The library method fragments oscillatory phases unless several discovery orders are combined, whereas aDMD + GMM summarises each trajectory by four numbers and returns a few large clusters Workflow Suggestion The results suggest a workflow in which a label-free DMD pass is used first to locate the main boundaries and to flag regions with unexpected structure, followed by targeted labelling and supervised classification This approach is affordable, with the CNN cost being dominated by training while the library’s cost grows faster than linearly with the number of points The authors also note that a classifier can only be judged against a reference that is itself converged Without any labels, the aDMD library and the aDMD + GMM method reach ARI = 0.916 and 0.947 and recover the main phase boundaries, far better than an autoencoder baseline The library also separates symmetric from symmetry-broken periodic oscillations, a physical distinction that the six-class heuristic misses Looking forward, the same approach could be applied to two-dimensional topological lattices and other nonlinear photonic systems whose phase content is not known in advance Data Availability The data that support the findings of this study are available from the corresponding author upon reasonable request The code used for the simulations and the classification methods is available from the corresponding author upon reasonable request Author Contributions C.P.I performed the simulations, implemented and analysed the classification methods, and wrote the manuscript G.H.A developed the driven Kerr SSH lattice model S.W developed the aDMD-based classification approach on which methods (iii) and (iv) are based and advised on its use S.S.O conceived and supervised the project Competing interests declare no competing interests References
1: S.
Kai: H.
Mira: Strogatz, Nonlinear Dynamics and Chaos, 2nd ed (Westview Press, Boulder, CO, 2015)
2: K.
Kai: Ikeda, Multiple-valued stationary state and its instability of the transmitted light by a ring cavity system, Opt.
Paper summary: Kai: So we're looking at this paper now titled "Machine learning classification of dynamical phases in a driven nonlinear Su-Schrieffer-Heeger lattice." It’s about using machine learning to sort out where these driven nonlinear photonic lattices end up—whether they settle into steady states, periodic motion, or chaotic behavior.
Mira: Exactly. The main thing this paper claims is that mapping where each of those dynamics occurs in parameter space is really important for designing actual devices. They apply four different classification methods to the same coherently driven Su-Schrieffer-Heeger chain of Kerr resonators <ref:2611.08287#pg2>.
Lev: It sounds like they are using a supervised convolutional neural network, or CNN, and comparing it against some label-free methods too. I wonder what kind of data they trained that CNN on <ref:2611.08287#pg2>.
Kai: Right, the paper states that the supervised CNN reaches a held-out accuracy of ninety-seven point nine six percent when trained on seventy percent of the heuristic labels <ref:2611.08287#pg2>. That's a high number, suggesting it’s pretty good at extending known classifications to new parameter points <ref:2611.08287#pg3>.
Mira: But then they look at the label-free methods, like the augmented dynamic mode decomposition or aDMD library building scheme <ref:2611.08287#pg2>. They say these methods return clusters whose physical meaning has to be established afterwards, even though they don't need explicit labels <ref:2611.08287#pg3>.
Lev: Establishing the physical meaning is tough, right? If you get a cluster but don't know if it’s stable or chaotic just by looking at it, you still have to figure out what that cluster actually means for the system <ref:2611.08287#pg2>.
Kai: That’s the core of the comparison here. They compare how well these different approaches match a reference, which is a hand-built heuristic classifier <ref:2611.08287#pg2>. The CNN gets the closest match to that reference only because it was trained on it <ref:2611.08287#pg3>.
Mira: But the label-free methods, like the aDMD library building and aDMD plus Gaussian mixture model, they reach adjusted Rand indices of zero point nine one six and zero point nine four seven without seeing any labels <ref:2611.08287#pg3>. That's quite good performance on their own <ref:2611.08287#pg3>.
Lev: So, what’s the trade-off there? The paper mentions that the aDMD library separates symmetric from symmetry-broken oscillations, which is something the heuristic classifier misses, but it can't distinguish between periodic and chaotic switching <ref:2611.08287#pg3>.
Kai: It seems like the CNN is efficient for extending known classifications to new parameter points, but there’s a limit—it cannot report a phase that wasn't in its training labels <ref:2611.08287#pg3>. That means if you go outside the training data, it just can't guess what happens <ref:2611.08287#pg3>.
Mira: And the authors suggest a workflow: start with a label-free decomposition to find the main boundaries and flag weird structures, then use targeted labeling and supervised classification <ref:2611.08287#pg3>. It’s affordable, but the cost of building a library grows faster than linearly with the number of points <ref:2611.08287#pg3>.
Lev: I noticed they mention a floor on the reconstruction error being essential for good performance in the aDMD plus Gaussian mixture model, specifically a floor of ten to the minus eight <ref:2611.08287#pg4>. Without that, they found that stable and bistable phases split roughly in half based on round-off errors <ref:2611.08287#pg4>.
Kai: So, even with the best methods, there’s a bit of noise from how the data is reconstructed <ref:2611.08287#pg4>. It shows that even with good math, you still need careful tuning for these unsupervised methods <ref:2611.08287#pg4>.
Mira: The overall accuracy is dominated by the large stable and bistable phases, but the small oscillatory phases are recognized less reliably when labels are scarce <ref:2611.08287#pg5>. So the classification power depends on how many examples you have for those specific regimes <ref:2611.08287#pg5>.
Lev: It’s interesting that they found the aDMD plus GMM method reaches an adjusted Rand index of zero point nine four seven using only nine of its thirty available mixture components <ref:2611.08287#pg3>. That suggests that summarizing each trajectory with just four numbers works better than trying to map every possible combination <ref:2611.08287#pg5>.
Kai: So, for the listener, what does this mean? It means we have several tools now to figure out where a photonic system will behave—supervised learning for accuracy, label-free methods for initial structure discovery <ref:2611.08287#pg3>.
Mira: And the paper itself is titled "Machine learning classification of dynamical phases in a driven nonlinear Su-Schrieffer-Heeger lattice" <ref:2611.08287#pg1>, suggesting that machine learning isn't just a tool here, but it’s becoming a way to systematically explore complex dynamical systems <ref:2611.08287#pg3>.
Lev: It points toward a future where we might use these techniques to tackle other nonlinear photonic systems whose phase content we don't know beforehand <ref:2611.08287#pg5>. The authors note that this approach could be applied to two-dimensional topological lattices too <ref:2611.08287#pg5>.
Kai: So, the implication is that classification of these nonlinear systems might become much more accessible through data analysis, even when we don't have perfect prior knowledge of the physics <ref:2611.08287#pg5>.
Mira: Indeed. The paper shows that even without labels, label-free methods can recover the main phase boundaries and distinguish things like symmetric from symmetry-broken periodic oscillations <ref:2611.08287#pg3>. It’s about finding structure in the data itself <ref:2611.08287#pg3>.
Lev: It seems like a solid foundation for using these methods to systematically explore the phase diagrams of these driven lattices <ref:2611.08287#pg3>. We'll be looking at how this translates to building real hardware, which is where my focus is <ref:2611.08287#pg4>.
Kai: So we’ve seen the setup, the comparison between supervised and unsupervised methods, and how they handle things like chaotic switching or bistability <ref:2611.08287#pg3>. That gives us a good picture of what’s possible now <ref:2611.08287#pg5>.
Conclusion: Kai: So we're wrapping up this look at the paper titled "Machine learning classification of dynamical phases in a driven nonlinear Su-Schrieffer-Heeger lattice."
Mira: It’s about taking these complex, driven quantum systems—these Su-Schrieffer-Heeger chains—and using machine learning to sort out what kind of behavior they settle into.
Lev: Basically, the authors are using four different classification methods on the same system to see which one gets you closest to the actual physical rules.
Kai: They compare a supervised neural network against two label-free techniques, and they show that both of those label-free methods can map out where things are happening in parameter space without having any prior labels.
Mira: The paper highlights how these methods, specifically the augmented dynamic mode decomposition and the Gaussian mixture model combination, recover the main boundaries even when you don't give them explicit labels.
Lev: But there’s a catch; they say you still have to interpret what those resulting clusters actually mean for the physics afterward. That’s where the real work is.
Kai: The overall message here is that machine learning isn't just a fancy tool for checking results; it’s becoming a way to systematically explore these kinds of dynamical phases in complex quantum hardware designs.
Mira: It suggests that we might be able to use these techniques to predict what kind of state our photonic devices will settle into before we even run the experiment.
Lev: It opens up a path for designing systems where you can quickly map out all the possible stable and chaotic regions in a parameter space, which is crucial when building real hardware.
Kai: And that's exactly what we need to keep looking at next—how this kind of classification applies to other types of nonlinear photonic systems.
Chukwunaedum Patrick Izediuno, *Sang Soon Oh
School of Physical, Chemical and Environmental Sciences, Cardiff University · Department of Physics, University of Tabuk · Ecole Centrale de Lyon, CNRS, INSA Lyon, Université Claude Bernard Lyon 1 · Bentham Instruments Ltd · Department of Control and Instrumentation Engineering, Korea University · Department of Electro-Mechanical System Engineering, Korea University · Department of Mechanical and Robotics Engineering, Gwangju Institute of Science and Technology
physics.optics
Submitted: 2026-11-05
Updated: 2026-11-05
Comments: 13 pages, 7 figures (3 in the main text, 4 in the appendices)
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 92/100
The gist: Driven nonlinear photonic lattices can settle into steady, symmetry-broken, periodic, or chaotic states Mapping where each occurs in parameter space is key to device design The authors apply four…
Key concepts
- Su-Schrieffer-Heeger (SSH) Chain
- This is a specific type of nonlinear photonic lattice model used in the research. It describes how light interacts within a chain of coupled resonators, which can exhibit complex dynamical behaviors like periodic or chaotic states depending on the input parameters.
- Dynamical Phases
- These are the different stable or unstable states that the system settles into when driven. The paper aims to map out where these phases occur in parameter space, such as stable, bistable, symmetric oscillations, and symmetry-broken oscillations.
- Convolutional Neural Network (CNN)
- A type of deep learning model used for classification. In this study, the CNN was trained on a subset of known phase labels to learn patterns and accurately classify new parameter points in the lattice system.
Terminology
Summary
Driven nonlinear photonic lattices can settle into steady, symmetry-broken, periodic, or chaotic states Mapping where each occurs in parameter space is key to device design The authors apply four classifiers to the same coherently driven Su-Schrieffer-Heeger (SSH) chain of Kerr resonators Using a converged heuristic diagram as the reference, the CNN reaches 98. p. 0 Machine learning classification of dynamical phases in a driven nonlinear
Su-Schrieffer-Heeger lattice
Chukwunaedum Patrick Izediuno,1 Ghada H. Alharbi,2 Stephan Wong,3
Zeeshan Ahmad,4 Sunae So,5, 6 Minkyung Kim,7 and Sang Soon Oh1, ∗
1School of Physical, Chemical and Environmental Sciences,
Cardiff University, Cardiff CF24 3AA, United Kingdom
2Department of Physics, University of Tabuk, Tabuk 741, Kingdom of Saudi Arabia
3Ecole Centrale de Lyon, CNRS, INSA Lyon, Universit´e Claude Bernard Lyon 1,
CPE Lyon, CNRS, INL, UMR5270, 69130 Ecully, France
4Bentham Instruments Ltd, Reading RG2 0NH, United Kingdom
5Department of Control and Instrumentation Engineering,
Korea University, 2511, Sejong-ro, Sejong, 30019 Republic of Korea
6Department of Electro-Mechanical System Engineering,
Korea University, 2511, Sejong-ro, Sejong, 30019 Republic of Korea
7Department of Mechanical and Robotics Engineering,
Gwangju Institute of Science and Technology (GIST), 61005 Gwangju, Republic of Korea
(Dated: November 6, 2026)
Driven nonlinear photonic lattices can settle into steady, symmetry-broken, periodic, or chaotic
states, and mapping where each occurs in parameter space is key to device design. Using a converged heuristic diagram as the reference, the CNN reaches 98. 0% held-out accuracy,
while the two label-free methods reach adjusted Rand indices (ARIs) of 0.Where the paper says it
Where the paper says it
Phase Identification Methods
The study compares a hand-built heuristic classifier, a supervised convolutional neural network (CNN), an unsupervised library-building scheme based on augmented dynamic mode decomposition (aDMD), and an unsupervised Gaussian mixture model (GMM) on four dynamical invariants The heuristic classifier serves as the reference for all other methods, requiring explicit rules for each phase The supervised CNN is trained on 70% of the heuristic labels and achieves a held-out accuracy of 97.96% The two label-free methods, aDMD library building and aDMD + GMM, return clusters whose physical meaning must be established afterwards
Performance Comparison
The comparison shows that the CNN gives the closest match to the reference, but only because it is trained on it The two label-free methods need no prior description of the phases and recover the main boundaries without labels, but their clusters must be interpreted afterwards Both separate the stable and bistable phases and the symmetric from the symmetry-broken oscillations, and both largely merge periodic and chaotic switching The aDMD library separates symmetric from symmetry-broken oscillations, which the heuristic misses, though it cannot distinguish periodic from chaotic switching
Method Strengths and Limitations
The CNN is efficient for extending known classifications to new parameter points but cannot report a phase that is absent from its training labels The aDMD + GMM method uses four scalar invariants to cluster trajectories, and the floor on the reconstruction error is essential for good performance The library method fragments oscillatory phases unless several discovery orders are combined, whereas aDMD + GMM summarises each trajectory by four numbers and returns a few large clusters
Workflow Suggestion
The results suggest a workflow in which a label-free DMD pass is used first to locate the main boundaries and to flag regions with unexpected structure, followed by targeted labelling and supervised classification This approach is affordable, with the CNN cost being dominated by training while the library’s cost grows faster than linearly with the number of points The authors also note that a classifier can only be judged against a reference that is itself converged
Conclusion
Without any labels, the aDMD library and the aDMD + GMM method reach ARI = 0.916 and 0.947 and recover the main phase boundaries, far better than an autoencoder baseline The library also separates symmetric from symmetry-broken periodic oscillations, a physical distinction that the six-class heuristic misses Looking forward, the same approach could be applied to two-dimensional topological lattices and other nonlinear photonic systems whose phase content is not known in advance
Data Availability
The data that support the findings of this study are available from the corresponding author upon reasonable request The code used for the simulations and the classification methods is available from the corresponding author upon reasonable request
Author Contributions
C.P.I performed the simulations, implemented and analysed the classification methods, and wrote the manuscript G. p. 6 Author contributions. performed the simulations, implemented and analysed the classification methods, and wrote the manuscript. performed the simulations, implemented and analysed the classification methods, and wrote the manuscript. developed the
driven Kerr SSH lattice model. developed the
driven Kerr SSH lattice model. developed the
aDMD-based classification approach on which methods
(iii) and (iv) are based and advised on its use.Where the paper says it
Where the paper says it
Where the paper says it
References
[1] S. H. Strogatz, Nonlinear Dynamics and Chaos, 2nd ed (Westview Press, Boulder, CO, 2015)
[2] K. Ikeda, Multiple-valued stationary state and its instability of the transmitted light by a ring cavity system, Opt. p. 6 (Westview Press, Boulder, CO, 2015). Ikeda, Multiple-valued stationary state and its instability of the transmitted light by a ring cavity system,
Opt.Where the paper says it
[3] S. Longhi, Y. Kominis, and V. p. 6 30, 257 (1979). Kominis, and V.Where the paper says it
[4] M. T. M. Woodley, L. p. 6 Kovanis, Presence of
temporal dynamical instabilities in topological insulator
lasers, EPL 122, 14004 (2018). Woodley, L. Woodley, L. Del Bino, G. Del Bino, G. Oppo, and
P. Oppo, and
P. Del’Haye, Self-switching Kerr oscillations of counterpropagating light in microresonators, Phys.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[5] M. A. Bandres et al. p. 6 126, 043901 (2021). Bandres et al.Where the paper says it
[6] G. Harari et al. p. 6 , Topological insulator laser: Experiments, Science 359, eaar4005 (2018). Harari et al.Where the paper says it
[7] A. Tikan et al. p. 6 , Topological insulator laser: Theory, Science 359, eaar4003 (2018). Tikan et al. Tikan et al. , Emergent nonlinear phenomena in a
driven dissipative photonic dimer, Nat.Where the paper says it
Where the paper says it
[8] G. H. Alharbi, S. p. 0 Machine learning classification of dynamical phases in a driven nonlinear
Su-Schrieffer-Heeger lattice
Chukwunaedum Patrick Izediuno,1 Ghada H. Alharbi,2 Stephan Wong,3
Zeeshan Ahmad,4 Sunae So,5, 6 Minkyung Kim,7 and Sang Soon Oh1, ∗
1School of Physical, Chemical and Environmental Sciences,
Cardiff University, Cardiff CF24 3AA, United Kingdom
2Department of Physics, University of Tabuk, Tabuk 741, Kingdom of Saudi Arabia
3Ecole Centrale de Lyon, CNRS, INSA Lyon, Universit´e Claude Bernard Lyon 1,
CPE Lyon, CNRS, INL, UMR5270, 69130 Ecully, France
4Bentham Instruments Ltd, Reading RG2 0NH, United Kingdom
5Department of Control and Instrumentation Engineering,
Korea University, 2511, Sejong-ro, Sejong, 30019 Republic of Korea
6Department of Electro-Mechanical System Engineering,
Korea University, 2511, Sejong-ro, Sejong, 30019 Republic of Korea
7Department of Mechanical and Robotics Engineering,
Gwangju Institute of Science and Technology (GIST), 61005 Gwangju, Republic of Korea
(Dated: November 6, 2026)
Driven nonlinear photonic lattices can settle into steady, symmetry-broken, periodic, or chaotic
states, and mapping where each occurs in parameter space is key to device design. Alharbi, S. Gong, and S. Gong, and S. Oh,
Asymmetrical temporal dynamics of edge modes in SuSchrieffer-Heeger lattice with Kerr nonlinearity, Phys.Where the paper says it
Where the paper says it
Where the paper says it
[9] H. Ohadi et al. p. 6 Research 6, L042013 (2024). Ohadi et al. Ohadi et al. , Spin order and phase transitions in chains
of polariton condensates, Phys.Where the paper says it
Where the paper says it
[10] L. Del Bino, J. p. 6 Woodley, L. Del Bino, G. Del Bino, J. Stebbings, and
P. Stebbings, and
P. Del’Haye, Symmetry breaking of counter-propagating
light in a nonlinear resonator, Sci.Where the paper says it
Where the paper says it
Where the paper says it
[11] S. Malzard and H. p. 6 7, 43142 (2017). Malzard and H. Malzard and H. Schomerus, Nonlinear mode competition and symmetry-protected power oscillations in topological lasers, New J.Where the paper says it
Where the paper says it
[12] G. Carleo, I. Cirac, K. Cranmer, L. p. 6 20, 063044 (2018). Cranmer, L. Vogt-Maranto, and L. Zdeborov´a, Machine
learning and the physical sciences, Rev.Where the paper says it
Where the paper says it
[13] J. Carrasquilla and R. p. 6 91,
045002 (2019). Carrasquilla and R. Carrasquilla and R. Melko, Machine learning phases
of matter, Nat.Where the paper says it
Where the paper says it
[14] P. Zhang, H. Shen, and H. p. 6 13, 431 (2017). Shen, and H. Shen, and H. Zhai, Machine learning topological invariants with neural networks, Phys.Where the paper says it
Where the paper says it
[15] S. J. Wetzel, Unsupervised learning of phase transitions: From principal component analysis to variational autoencoders, Phys. p. 6 120, 066401 (2018). Wetzel, Unsupervised learning of phase transitions:
From principal component analysis to variational autoencoders, Phys.Where the paper says it
[16] E. P. L. van Nieuwenburg, Y. p. 6 E 96, 022140 (2017). van Nieuwenburg, Y. van Nieuwenburg, Y. Liu, and S. Liu, and S. Huber,
Learning phase transitions by confusion, Nat.Where the paper says it
Where the paper says it
Where the paper says it
[17] J. F Rodriguez-Nieva and M. p. 6 13,
435 (2017). Rodriguez-Nieva and M. Rodriguez-Nieva and M. Scheurer, Identifying
topological order through unsupervised machine learning, Nat.Where the paper says it
Where the paper says it
[18] M. S Scheurer and R. p. 6 Rodriguez-Nieva and M. Scheurer, Identifying
topological order through unsupervised machine learning, Nat. Scheurer and R. Slager, Unsupervised machine
learning and band topology, Phys.Where the paper says it
Where the paper says it
[19] P. J Schmid, Dynamic mode decomposition of numerical and experimental data, J. p. 6 124, 226401
(2020). Schmid, Dynamic mode decomposition of numerical
and experimental data, J. Schmid, Dynamic mode decomposition of numerical
and experimental data, J. Fluid Mech.Where the paper says it
Where the paper says it
[20] J. H Tu, C. W Rowley, D. M Luchtenburg, S L Brunton, and J N Kutz, On dynamic mode decomposition: Theory and applications, J. p. 6 Brunton, and J. Kutz, On dynamic mode decomposition:
Theory and applications, J.Where the paper says it
[21] B Kramer, P Grover, P Boufounos, S Nabi, and M Benosman, Sparse sensing and DMD-based identification of flow regimes and bifurcations in complex flows, SIAM J. Appl. Dyn. Syst. 16, 1164 (2017)
[22] S Wong, J Olthaus, T K Bracht, D E Reiter, and S S Oh, A machine learning approach to drawing phase diagrams of topological lasing modes, Commun. Phys. 6, 104 (2023)
[23] S Wong, D E Reiter, and S S Oh, Machine learning for identifying dynamical phases in topological lasers, in Artificial Intelligence and Intelligent Matter: Nanoscience, Soft Matter, Philosophy, Machine Intelligence for Materials Science, edited by M te Vrugt (Springer, Cham) 2026 pp. 167–188
[24] D Zvyagintseva, H Sigurdsson, V K Kozin, I Iorsh, I A Shelykh, V Ulyantsev, and O Kyriienko, Machine learning of phase transitions in nonlinear polariton lattices, Commun. Phys. 5, 8 (2022)
[25] L Pilozzi et al., Machine learning inverse problem for topological photonics, Commun. p. 6 Marcucci, and C. Conti, Machine learning inverse problem for topological photonics,
Commun.Where the paper says it
[26] W. P Su, J R Schrieffer, and A J Heeger, Solitons in polyacetylene, Phys. p. 6 Schrieffer, and A. Heeger, Solitons in
polyacetylene, Phys.Where the paper says it
[27] S Malzard et al., Topological dynamics and excitations in lasers and condensates with saturable gain or loss, Opt. p. 6 Cancellieri, and H. Schomerus, Topological dynamics and excitations in lasers and condensates with saturable gain or loss, Opt.Where the paper says it
[28] H Zhao et al., Topological hybrid silicon microlasers, Nat. p. 6 Schomerus, and L. Feng, Topological hybrid
silicon microlasers, Nat.Where the paper says it
[30] G Benettin, L Galgani, A Giorgilli, and J-M Strelcyn, Lyapunov characteristic exponents for smooth dynamical systems and for Hamiltonian systems, Meccanica 15, 9 (1980)
[31] I Loshchilov and F Hutter, Decoupled weight decay regularization, in Proc. ICLR (2019)
[32] I Goodfellow, Y Bengio, and A Courville, Deep Learning (MIT Press, Cambridge, MA) 2016
[33] A L N Fred and A K Jain, Combining multiple clusterings using evidence accumulation, IEEE Trans. Pattern Anal. p. 7 Jain, Combining multiple clusterings using evidence accumulation, IEEE Trans. Pattern
Anal.Where the paper says it
[34] G McLachlan and D Peel, Finite Mixture Models (Wiley, New York) 2000
[35] D M Blei and M I Jordan, Variational inference for Dirichlet process mixtures, Bayesian Anal. 1, 121 (2006)
[36] G Schwarz, Estimating the dimension of a model, Ann. Stat. 6, 461 (1978)
[37] X Guo, L Gao, X Liu, and J Yin, Improved deep embedded clustering with local structure preservation, in Proc. IJCAI (2017) p. p. 7 Yin, Improved deep embedded clustering with local structure preservation, in
Proc. IJCAI (2017) p.Where the paper says it
[38] L Hubert and P Arabie, Comparing partitions, J. Classif. 2, 193 (1985)
[39] S L Brunton, M Budiˇsi´c, E Kaiser, and J N Kutz, Modern Koopman theory for dynamical systems, SIAM Rev. 64, 229 (2022)
[40] Y Gong et al., Topological insulator laser using valley-Hall photonic crystals, ACS Photonics 7, 2089 (2020)
[41] A Dikopoltsev et al., Topological insulator verticalcavity laser array, Science 373, 1514 (2021)
[42] S Wong et al., Nonlinear topological photonics: capturing nonlinear dynamics and optical thermodynamics, ACS Photonics 12, 2291 (2025)
Appendix A Details
The integration window was chosen by running the heuristic classifier (Appendix B) with three windows: tf = 60 with the first third discarded, tf = 120 with the first half discarded, and tf = 240 with the first half discarded The tf = 120 and tf = 240 diagrams agree on 98.4% of the grid points (ARI 0. p. 3 We instead use the adjusted Rand index (ARI) [38]. For
every pair of grid points, it checks whether the two labellings agree, either by placing both points in the same
group or by placing them in different groups.Where the paper says it
Appendix B Rules
A trajectory is labelled chaotic if it exceeds 5 × 10−3 per step, i.e. 0.25 in units of the decay rate A constant trajectory is labelled bistable if their mean intensities differ by more than 5% of their average Switching is detected from the sign of I1 −I2, which is counted only when I1 −I2 exceeds 10% of the mean intensity
Appendix C Details
The library pipeline follows Refs. [22, 23] The consensus reaches ARI = 0. p. 4 ically selected merging threshold depends on the order,
but their partitions agree well with one another (pairwise ARI 0. The consensus of the three orders
contains 60 regimes and reaches ARI = 0. The consensus of the three orders
contains 60 regimes and reaches ARI = 0. 916 against
the reference, with homogeneity 0. The consensus of the three orders
contains 60 regimes and reaches ARI = 0. 916 against
the reference, with homogeneity 0. 916 against
the reference, with homogeneity 0. 865 and completeness
0. 916 against
the reference, with homogeneity 0. 865 and completeness
0.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Appendix D Details
The floor on the reconstruction error is essential: without it, the stable and bistable phases are each split roughly in half according to round-off-level differences in this error A floor of 10−8, which lies below the round-off values, gives 0.61
Appendix E Details
The overall accuracy is dominated by the large stable and bistable phases, however, and the small oscillatory phases are recognised less reliably when fewer labels are available The CNN is therefore an efficient way to extend a known classification to new parameter points, but it cannot report a phase that is absent from its training labels
Appendix G Details
The cost of the library grows faster than linearly with the number of points, because each trajectory is compared with every basis already in the library The CNN cost is dominated by training The heuristic and aDMD + GMM costs grow linearly and the CNN cost is dominated by training
Table II Summary
The agreement of each method with the heuristic reference on the full 60×60 grid is summarized in Table II The aDMD + GMM method reaches ARI = 0.947, with homogeneity 0. p. 5 3(d)] uses 9 of its 30
available mixture components and reaches ARI = 0. 947,
with homogeneity 0. 3(d)] uses 9 of its 30
available mixture components and reaches ARI = 0. 947,
with homogeneity 0. 947,
with homogeneity 0. 899 and completeness 0. 947,
with homogeneity 0. 899 and completeness 0. ically selected merging threshold depends on the order,
but their partitions agree well with one another (pairwise ARI 0. The consensus of the three orders
contains 60 regimes and reaches ARI = 0. 916 (ARI) 0. 760 60a
(iv) aDMD + GMM 0. The aDMD + GMM method [Fig. 3(d)] uses 9 of its 30
available mixture components and reaches ARI = 0. The aDMD + GMM method [Fig. 3(d)] uses 9 of its 30
available mixture components and reaches ARI = 0.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Figure S4 Summary
The held-out accuracy of the supervised CNN versus training-set size saturates at 98.3% for training sets of 50% and 70% of the grid The overall accuracy is dominated by the large stable and bistable phases The periodic and chaotic phases together make up only 16% and are affected much more by the number of labels
Figure S3 Summary
The Dirichlet-process GMM with 9 clusters reaches ARI = 0.946–0.948 The finite full-covariance GMM selected by BIC gives the same agreement with more clusters, seven of which are small groups of mostly boundary points The aDMD + GMM method [Fig. p. 5 845, which shows how strongly an unconverged reference can understate the agreement. The aDMD + GMM method [Fig.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Supervised Convolutional Neural Networks can be used to extend a known classification to new parameter points by reproducing a known classification quickly and with boundary limited errors, although they cannot report a phase that has not been shown in their training labels.
-
Label free DMD based methods can be used as a first survey of an unexplored nonlinear lattice to locate the main boundaries and flag regions with unexpected structure, followed by targeted labeling and supervised classification.
-
The aDMD library building scheme can separate symmetric from symmetry broken oscillations, which the heuristic misses, providing a physical distinction that is useful when it disagrees with the reference diagram in certain regions.
-
The aDMD plus Gaussian mixture model method can summarize each trajectory by four scalar invariants and cluster them using a Dirichlet process GMM to return clusters whose number is inferred rather than fixed in advance.
Abstract
Driven nonlinear photonic lattices can settle into steady, symmetry-broken, periodic, or chaotic states, and mapping where each occurs in parameter space is key to device design. Brute-force mapping is slow and relies on hand-built criteria, while machine-learning classifiers are usually tested on different systems, so it is unclear which suits a new lattice. We therefore apply four classifiers to the same coherently driven Su-Schrieffer-Heeger (SSH) chain of Kerr resonators: a hand-built heuristic classifier based on Lyapunov exponents and spectral diagnostics, a supervised convolutional neural network (CNN) trained on the heuristic labels, an unsupervised library-building scheme based on augmented dynamic mode decomposition (aDMD), and an unsupervised Gaussian mixture model (GMM) on four dynamical invariants that include an aDMD reconstruction error. Using a converged heuristic diagram as the reference, the CNN reaches 98.0% held-out accuracy, while the two label-free methods reach adjusted Rand indices (ARIs) of 0.916 and 0.947, respectively, without seeing any labels. We show that the aDMD library cleanly resolves the stable and bistable phases and separates symmetric from symmetry-broken periodic oscillations, which the heuristic misses, though it cannot distinguish periodic from chaotic switching. We also find that a poorly converged reference understates the label-free methods' accuracy.
Related papers
- Depth-enhanced molecular imaging with two-photon oblique plane microscopy
- Attosecond soft X-ray pulses generated by chirp-dispersed manipulation in an XFEL reveal nonlinear core-electron dynamics in neon
- Raman optical activity and helicity-dependent phonon splitting in chiral CoSi
- Stability and optical quality of "windmill"-formed 8CB liquid crystal films for replenishable plasma mirrors
- Experimental Scaling of Diffraction Efficiency in Laser-Induced Plasma Gratings
- Photonic Exceptional Topological Bound States in the Continuum