{"id":"572e2731-c4e4-4040-b315-12031193731c","arxiv_id":"2501.00046","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A DRL-assisted Jacobian-Free Newton-Krylov method finds 303 fixed points of the 2D Kuramoto-Sivashinsky equation on a 20x20 periodic domain.","lead":"The authors use deep reinforcement learning to supply better starting guesses to a Newton-Krylov solver, enabling it to find fixed points of the 2D Kuramoto-Sivashinsky equation where the solver alone fails. They report 303 fixed points and demonstrate controlled navigation between two of them.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 303-distinct-fixed-point claim is not verifiable from the published tables: only a few Fourier magnitudes are given, phases are omitted, and E27/E28 are listed identically, so the count could include duplicates or non-solutions.","rationale":"The reader's conditional verdict is appropriate. The strongest claim needs both genuineness and distinctness of 303 equilibria. The paper provides independent support for the method (public code, a convergence comparison, validation against one literature steady state), but the published tables are too coarse to establish the catalog's content. The E27/E28 duplicate row is the sharpest realization of this gap. The missing data is easily fixable by releasing full fields and residuals, so the correct disposition remains CONDITIONAL rather than rejection. My emphasis is on evidential sufficiency (phases and residuals) rather than on the discretization fidelity itself, so I mark partial agreement with the reader's weakest-assumption statement.","tokens_in":21744,"tokens_out":5164,"duration_ms":56073,"concrete_test":"Run the public repository's JFNK/DRL pipeline to regenerate all 303 converged states, store every 64×64 complex Fourier field and the relative residual reported by JFNK, then compute (i) the PDE RHS residual for each field, and (ii) the pairwise L2 distance between all 303 fields. If any field has residual above the claimed tolerance (or above a threshold of, say, 10−6 in the RHS norm) or any pair of fields has distance below 10−8, the count of 303 distinct genuine fixed points is not supported. At minimum, extract the full fields for E27 and E28 and compare them exactly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Tables 1–2 list 303 distinct fixed points. What would have to be true: each listed state is an equilibrium of the 2D KSE to the claimed tolerance, and the 303 states are pairwise different. The paper does not supply the data needed to check either condition. Table 1 gives only |b̂(0,1)|, |b̂(1,1)|, |b̂(1,0)|, and Table 2 gives seven more magnitudes for E1–E17 only. Complex Fourier phases are never reported, so two physically distinct fields with the same magnitudes but different phases would be indistinguishable in the tables; two identical fields would also be indistinguishable. This is not a hypothetical edge case: rows E27 and E28 in Table 1 are numerically identical (0, 0, 6331.2) and no Table 2 entries are provided for them. Either they are a typo, a duplicate, or two genuinely different solutions whose phase information was suppressed; the published record cannot tell which. In addition, no residual norms or full 64×64 fields are released, so convergence to the stated 10−12 relative residual cannot be checked, nor can the effect of the dt=0.05 time discretization on the reported equilibria be assessed. The claim '303 distinct, previously unreported fixed points' therefore rests on unverifiable numerical artifacts.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a hybrid method in which deep reinforcement learning (DDPG) is used to generate improved initial guesses for a Jacobian-Free Newton-Krylov (JFNK) solver, applied to finding fixed points of the two-dimensional Kuramoto-Sivashinsky equation (2D KSE) on a 20-by-20 periodic domain with 64x64 Fourier collocation points. The authors report 303 distinct fixed points, claim that these are new and not previously reported, and also demonstrate DRL-based navigation between known fixed points. The numerical methods are standard: spectral spatial discretization, exponential time-differencing/RK4 time stepping, and a JFNK solver with stated tolerances. The core empirical claims are the 303-point catalogue in Tables 1 and 2 and the improved convergence of JFNK when initialized by DRL.","tokens_in":22006,"tokens_out":4667,"duration_ms":51686,"significance":"If the catalogue is correct, the paper provides a substantial numerical resource for studying equilibria of the 2D KSE in a chaotic regime, and it demonstrates a practical use of DRL as a preconditioning/initialization strategy for Newton-type solvers in high-dimensional dynamical systems. The public release of code is a positive feature, as is the use of a residual-based reward that avoids circular fitting of the fixed points. The main significance, however, depends entirely on the verifiability of the 303 distinct equilibria. As submitted, the tables do not supply enough information to verify either convergence or pairwise distinctness, so the central claim is not yet supported by the published record.","major_comments":[{"comment":"The central claim that Tables 1 and 2 list 303 distinct fixed points cannot be verified from the published data. For each point, Table 1 gives only the absolute values of three complex Fourier coefficients, and Table 2 gives seven additional magnitudes for E1-E17 only. Complex phases are never reported, so two physically different fields with identical magnitudes are indistinguishable in the tables, and two identical fields are also indistinguishable. The text even acknowledges this degeneracy ('in some cases the values of be(0,1), be(1,1), be(1,0) appear the same for distinct fixed points'), but the published record does not resolve the degeneracy for the full catalogue. To support the 303-count claim, the authors should deposit the full 64x64 Fourier fields (or all complex coefficients) for every listed point, together with a script that computes pairwise distances and verifies distinctness.","section":"Section 4(b), Tables 1 and 2"},{"comment":"Rows E27 and E28 are numerically identical in all three reported magnitudes: (be(0,1), be(1,1), be(1,0)) = (0, 0, 6331.2). Table 2 provides no additional coefficients for these two rows. Since the paper itself notes that distinct fixed points can share these three magnitudes, the published information cannot distinguish between a typo, a duplicate entry, and two genuinely different solutions whose phase information was suppressed. Because the count of 303 rests on such distinctions, this specific case must be resolved explicitly in a revision.","section":"Table 1, rows E27 and E28"},{"comment":"No residual norms are reported for the individual fixed points. The paper states a JFNK relative error tolerance of 10^-12 and Appendix A says successful cases reduce the tolerance below 10^-10, but Tables 1 and 2 do not list the final relative residual for any of the 303 states. Without these data, the reader cannot confirm that each listed state is actually an equilibrium of the discretized 2D KSE to the claimed tolerance. The revision should either tabulate the final residuals or provide them in the public repository along with a verification script.","section":"Section 3(a) and Appendix A"},{"comment":"The grid-independence and time-step-independence checks are reported for a 25-time-unit integration, and the comparison with Kalogirou et al. [35] is made for a single steady state. This does not establish that all 303 converged states, which span a wide range of amplitudes and symmetry properties, are converged solutions of the continuous PDE. The authors should demonstrate for a representative subset of the catalogue, including high-amplitude and asymmetric states such as E135, that the fixed points are insensitive to increasing the resolution to 128x128 and decreasing the time step to 0.025, using the same JFNK convergence criterion.","section":"Section 3(a), validation of discretization"}],"minor_comments":[{"comment":"The sensor location set xs_i in {0,4,8,16...60} appears to be missing the value 12; it should probably read {0,4,8,12,...,60}. Please clarify.","section":"Section 3(a), sensor locations"},{"comment":"The main text states a relative error tolerance of 10^-12, while Appendix A says successful cases reduce the tolerance to below 10^-10. These two statements should be reconciled or explained.","section":"Appendix A"},{"comment":"Figure 2 compares the authors' result with a figure from Ref. [35], but the comparison is only visual. A quantitative error measure between the two steady states would strengthen the validation.","section":"Section 3(a), Figure 2"},{"comment":"The caption says values are 'the absolute value of the first three complex-valued Fourier coefficients' and are not normalized, but it is not stated whether these values are raw magnitudes of the unnormalized FFT or scaled in some other way. Please define the normalization convention precisely in the caption or text.","section":"Table 1 caption"},{"comment":"There are several small reference formatting errors, for example 'Sivashinsk G' instead of Sivashinsky, and inconsistent capitalization in the Physical Review E entries. These should be cleaned up.","section":"References"},{"comment":"Figure 8 states that E1 and E48 exhibit symmetry in x and y, while E81 and E135 are asymmetric. It would be helpful to quantify symmetry (for example, by reporting the residual under the relevant reflection operations) rather than relying on visual inspection.","section":"Section 4(b), Figures 8 and 9"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a valid problem and the DRL-JFNK idea is interesting, but the central numerical claim is not verifiable from the submitted manuscript. The missing data — full Fourier fields, phases, and per-point residuals — are precisely the kind of information that should be part of a computational paper claiming a catalogue of new solutions. The E27/E28 duplicate-looking rows are a concrete red flag that the distinctness claim needs careful checking. I would encourage the editor to request a revision in which the authors supply the data and verification scripts; if they cannot or do not, the paper should not be accepted as evidence for the 303 distinct fixed points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my take. The paper extends the DRL-assisted JFNK idea from 1D KSE (Zeng & Graham, Bucci et al.) to the 2D case, which is a meaningful step because the state space is much larger and the control problem harder. The authors validate their spectral code against Kalogirou et al., make the code public, and show a clean comparison where DRL-enhanced guesses cut the average JFNK iteration count from 38.2 to 27.4. That part is solid and useful.\n\nThe weak spot is exactly where the reviewer and the stress-test note put it: the central claim—303 distinct, previously unreported fixed points—is not something a reader can check from the manuscript. Table 1 gives only the absolute values of three Fourier coefficients per solution, and phases are never shown. Two physically different fields could share those magnitudes, and two identical fields would also be indistinguishable. The paper itself acknowledges this and points to Table 2 for extra coefficients, but Table 2 covers only E1–E17. E27 and E28 are listed with the identical triple (0, 0, 6331.2) and no further data, so the count of distinct solutions is unverifiable. There are also no residual norms anywhere, so we cannot confirm that all 303 entries actually satisfy the fixed-point equation to the stated 10^-12 tolerance. The 64x64 spatial discretization and dt=0.05 time step are validated for one steady state, but not for the whole catalog, so some entries could be artifacts of the temporal scheme (though I think that risk is modest).\n\nNone of this means the results are wrong. The catalog might be genuine and would be a nice contribution to the 2D KSE literature. But a referee needs the full solution data—residuals, phases, and ideally the 64x64 fields—to verify the count. This should be a condition for acceptance, not a reason for outright rejection. The method comparison and the navigation control section are fine. I'd send this to review, but I'd tell the authors to release the data and fix the E27/E28 ambiguity.","headline":"A plausible method and a potentially valuable catalog, but the 303 fixed points claim is under-evidenced by the published tables.","tokens_in":22585,"tokens_out":3543,"would_cite":false,"duration_ms":30931,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65H10","65M70"],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid reinforcement-learning and Newton solver identifies 303 previously unreported fixed points of the two-dimensional Kuramoto-Sivashinsky equation, a canonical chaotic partial differential equation.","keywords":["deep reinforcement learning","Kuramoto-Sivashinsky equation","fixed point","Jacobian-Free Newton-Krylov","DDPG","chaotic dynamical systems","two-dimensional PDE","initial guess generation"],"falsifier":"Recompute a sample of the reported equilibria, especially E27 and E28, at $128\\times128$ Fourier resolution with time step 0.025, and compare the full Fourier spectra; if the high-wavenumber modes merge or the residual does not converge below $10^{-10}$, the stated catalog over-counts or mislabels distinct fixed points.","tokens_in":21511,"feed_emoji":"🌀","tokens_out":7185,"duration_ms":66681,"temperature":0.7,"pith_summary":"This paper tries to establish that deep reinforcement learning can solve a practical bottleneck in finding equilibria of high-dimensional chaotic PDEs: the Jacobian-Free Newton-Krylov (JFNK) method needs a good initial guess or it fails. The authors train a Deep Deterministic Policy Gradient (DDPG) agent to steer the two-dimensional Kuramoto-Sivashinsky equation (2D KSE) toward states that barely change over one time step, then feed the best such state to JFNK. With that division of labor, they report 303 distinct fixed points of the 2D KSE on a $20\\times20$ periodic domain, which they say have not been reported in the literature. If correct, this shows that the chaotic 2D KSE hosts a large catalog of invariant equilibria, and that learning-based initial guesses can make Newton solvers practical in high-dimensional settings.","feed_headline":"Reinforcement learning finds 303 fixed points in 2D KSE","feed_subtitle":"Machine-learned initial guesses rescue a Newton solver that fails on two-thirds of random starts.","key_machinery":"The load-bearing mechanism is the DRL-as-preconditioner loop. The reward function $r = -\\|\\mathrm{FFT}(u_{t+\\Delta t}) - \\mathrm{FFT}(u_t)\\|_2$ acts as a spectral residual: it is zero exactly at a fixed point, where the unforced state reproduces itself after one time step, so states with high reward lie in the basin of convergence of JFNK. The actor-critic DDPG agent, with 256 sensor measurements and 36 Gaussian actuator amplitudes, searches this reward landscape over 500-step episodes, and the single best state per episode is handed to JFNK when its reward exceeds $\\theta_{\\mathrm{th}}=-45$. This matters because JFNK alone converges to a tolerance below $10^{-10}$ in successful cases but plateaus at residuals of order 0.1 to 1 in about two-thirds of 1500 random-start trials.","core_discovery":"On the paper's own terms, the central claim is a numerical discovery: the two-dimensional KSE on the square domain $[0,20]^2$ with periodic boundary conditions has at least 303 distinct steady solutions, all unstable, all previously absent from the literature. The route to this discovery is a two-stage method. First, a deep reinforcement learning agent, using 16x16 local sensors and 36 Gaussian actuators, is rewarded by the negative Fourier-space distance between the current state and its unforced evolution one time step later, namely $r = -\\|\\mathrm{FFT}(u_{t+\\Delta t}) - \\mathrm{FFT}(u_t)\\|_2$; maximizing this reward drives the system toward states close to a fixed point. The best state found in an episode, when its reward exceeds the threshold $-45$, becomes the initial guess for JFNK, which then converges to the equilibrium proper. The paper also claims that the learned controller can navigate the system between known fixed points, stabilizing the trajectory near an unstable target.","pith_inferences":["If the 303 count survives higher-resolution checks, it is a lower bound, not a census; the dense clustering of low-mode signatures near the origin in Figure 9 hints that many more equilibria may await discovery by the same pipeline.","The two entries E27 and E28 share identical values for the three tabulated low-order Fourier coefficients, yet no higher-order coefficients are given to separate them; a recomputation at $128\\times128$ resolution would settle whether the catalog over-counts.","Because the reward is computed in Fourier space with no forcing, the method is essentially a learned globalization strategy for Newton's method; the same idea could apply to finding other exact coherent structures, such as periodic orbits, by rewarding closeness to a recurrence rather than to a fixed point."],"forward_implications":["The 2D KSE in the chaotic regime at $2L=20$ possesses at least 303 invariant equilibria, giving a concrete target set for state-space analyses of spatiotemporal chaos.","The DRL-enhanced initial guesses cut the average JFNK convergence iterations from 38.2 to 27.4 when both start from the same random seeds.","The same reward function can be reused as a cheap fixed-point proximity oracle for other high-dimensional PDEs where Newton solvers fail without good initial guesses.","The trained DRL controller can steer the 2D KSE from a random state to a chosen unstable fixed point and hold it there for at least 50 time units, within roughly 8 time units of actuation."],"supporting_citations":[{"why":"contributes the DDPG actor-critic framework and the distance-based reward that this work adapts from 1D to 2D KSE control.","marker":"[33]"},{"why":"prior DRL active-control study identifying fixed points of the 1D KSE, the approach this paper extends to two dimensions.","marker":"[32]"},{"why":"reference numerical study of the 2D KSE whose published steady-state pattern validates the present simulation code and whose domain-size analysis fixes the chaotic regime.","marker":"[35]"},{"why":"provides the fourth-order exponential time-differencing RK4 stiff-PDE time stepper used for all simulations.","marker":"[34]"},{"why":"the open JFNK implementation that the authors adapt; its convergence behavior sets up the initial-guess problem addressed by DRL.","marker":"[36]"},{"why":"survey of Jacobian-free Newton-Krylov methods that frames the need for good initial guesses in large-scale nonlinear problems.","marker":"[16]"}],"fun_headline_variants":["RL boosts Newton solver to find 303 new KSE fixed points","Machine learning uncovers 303 unstable fixed points in 2D KSE","Reinforcement learning finds 303 new fixed points in 2D KSE","RL-guided Newton solver finds 303 fixed points"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The numerical configuration of 64x64 Fourier modes with time step 0.05 on the $20\\times20$ periodic domain faithfully captures every genuinely distinct fixed point of the continuous 2D KSE, including the high-wavenumber structure that tells E27 and E28 apart.","fun_headline_variants_meta":{"raw":{"variants":["RL boosts Newton solver to find 303 new KSE fixed points","Machine learning uncovers 303 unstable fixed points in 2D KSE","Reinforcement learning finds 303 new fixed points in 2D KSE","RL-guided Newton solver finds 303 fixed points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000633,"raw_usage":{"total_tokens":2906,"prompt_tokens":917,"completion_tokens":1989,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":1913}},"tokens_in":533,"tokens_out":1989,"duration_ms":15622,"temperature":1.0,"reasoning_tokens":1913,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:54:01.174188+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute a sample of the reported equilibria, especially E27 and E28, at $128\\times128$ Fourier resolution with time step 0.025, and compare the full Fourier spectra; if the high-wavenumber modes merge or the residual does not converge below $10^{-10}$, the stated catalog over-counts or mislabels distinct fixed points.","supporting_citations":[{"cited_title":"2019 Control of chaotic systems by deep reinforcement learning","cited_arxiv_id":null,"evidence_quote":"contributes the DDPG actor-critic framework and the distance-based reward that this work adapts from 1D to 2D KSE control."},{"cited_title":"2021 Symmetry reduction for deep reinforcement learning active control of chaotic spatiotemporal dynamics","cited_arxiv_id":null,"evidence_quote":"prior DRL active-control study identifying fixed points of the 1D KSE, the approach this paper extends to two dimensions."},{"cited_title":"2015 An in-depth numerical study of the two- dimensional Kuramoto–Sivashinsky equation","cited_arxiv_id":null,"evidence_quote":"reference numerical study of the 2D KSE whose published steady-state pattern validates the present simulation code and whose domain-size analysis fixes the chaotic regime."},{"cited_title":"2005 Fourth-order time-stepping for stiff PDEs","cited_arxiv_id":null,"evidence_quote":"provides the fourth-order exponential time-differencing RK4 stiff-PDE time stepper used for all simulations."},{"cited_title":"2004 Jacobian-free Newton–Krylov methods: a survey of approaches and applications","cited_arxiv_id":null,"evidence_quote":"survey of Jacobian-free Newton-Krylov methods that frames the need for good initial guesses in large-scale nonlinear problems."}],"review_version":1}