{"id":"d211f98c-3831-45bf-9498-5137dd966c4f","arxiv_id":"2506.13057","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Reinforcement learning (PPO) is shown to inverse-design a random 2D scattering billiard to achieve three target transmission matrix properties: rank-1 behavior, a degenerate eigenvalue exceptional point, and degenerate transmission eigenvalues.","lead":"A reinforcement learning agent moves small scatterers inside a random microwave billiard until the system's transmission matrix takes a desired shape, such as a rank-1 matrix, a degenerate exceptional point, or equal channel transmission. This is worth reading because it tests whether RL, a standard AI method, can replace slow trial-and-error engineering for complex wave-scattering devices.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Scale-free cost functions in Table 1 are trivially minimized by the zero (or arbitrarily weak) transmission matrix; the paper never reports absolute transmission, so the demonstrated rank-1, EP, and degenerate-τ designs may be vacuous.","rationale":"The paper's central claim is that PPO can inverse-design random media to realize a rank-1 TM, an exceptional point with degenerate eigenvalues, and degenerate transmission eigenvalues. For this claim to be meaningful, the designed systems must be actual power-handling devices, not just matrices with the right normalized properties but negligible overall scale. Table 1's cost functions are homogeneous of degree zero, so they are invariant under T → cT. Consequently, the origin T = 0 is a global minimizer (or a limit point of minimizers) for all three objectives: the rank-1 cost is zero for any matrix with a single nonzero singular value no matter how small; the degenerate-eigenvalue cost is zero for any nilpotent matrix no matter how small; and the degenerate-τ cost is zero for T = 0. The paper asserts that scale-free costs are specifically chosen to prevent the RL from collapsing the matrix scale to zero, but this misunderstands the geometry: a scale-free objective gives no gradient signal along the radial direction, so the scale is not controlled and can drift to arbitrarily small values without changing the reward. The paper reports only ratios (output intensity ratio 4.7, phase difference -2.7, coalescence C=1, eigenvalue degeneracy) and the vanishingly small eigenvalue 1.06e-7, but never the absolute magnitude of the nonzero singular value or total transmitted power. Without this number, the demonstrated rank-1 TM and EP could be realized by a configuration that simply blocks nearly all transmission, which would invalidate the proposed applications (fixed-ratio power conversion, unidirectional mode conversion). This concern is distinct from the reader's focus on optimization reproducibility; even a fully reproducible PPO run would not resolve it if the objective itself permits trivial zero-scale solutions. The concrete test above — comparing the absolute TM strength of the optimized configurations against random baselines — would settle the issue. If the test shows strong transmission, the concern is resolved and the remaining limitation is the lack of statistics; if it shows weak transmission, the central claim fails and the paper would need to add a transmission constraint to the cost functions and re-demonstrate the designs.","tokens_in":9631,"tokens_out":10956,"duration_ms":129505,"concrete_test":"From the open-source repository, extract the three reported optimized scatterer configurations and compute the full 2×2 transmission matrix at 15 GHz (or read the saved singular values). Compare the largest transmission eigenvalue τ_1 and trace(T†T) against the mean over at least 50 random scatterer configurations with the same geometry and source. If the optimized τ_1 is below 1% of the random-configuration mean for any of the three targets, the cost function was trivially satisfied by suppressing transmission. As a second check, re-run the degenerate-transmission-eigenvalue optimization with an additional penalty that activates whenever trace(T†T) falls below the random-mean value; if the target property cannot be achieved while keeping transmission at or above that level, the method does not produce functional devices.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing issue is not the absence of multi-seed statistics but the design of the cost functions themselves. Every objective in Table 1 is scale-free and therefore also minimized by the zero transmission matrix: for the rank-1 target, cost = 1 - τ_1/Στ_i equals 0 whenever τ_2 = 0, regardless of how small τ_1 is; for the degenerate-eigenvalue target, cost = 0 for any nilpotent matrix, including arbitrarily small multiples of a nilpotent matrix; for the degenerate-transmission-eigenvalue target, cost = 0 for T = 0. The paper's statement that a scale-free cost prevents the RL from driving the matrix scale toward zero is incorrect: scale invariance removes the penalty for scale collapse rather than preventing it, leaving the absolute scale of the TM uncontrolled. The paper never reports the largest transmission eigenvalue, the Frobenius norm of the TM, or the total transmitted power for the three designed systems, so the reader cannot tell whether the rank-1, EP, and degenerate-τ results are functional devices or vacuous near-zero-transmission configurations. This is a load-bearing gap: if the optimized T is orders of magnitude weaker than a typical random scatterer configuration, the central claim of successful inverse design is not established.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Using a 2D billiard with two input/output waveguides and up to 20 movable dielectric cylinders, the author applies Proximal Policy Optimization (PPO) to position the cylinders so that the simulated transmission matrix attains one of three target properties: rank-1 structure with a transmission zero, a degenerate eigenvalue (exceptional point with coalescing eigenvectors), or degenerate transmission eigenvalues. The agent observes the normalized scatterer coordinates, applies incremental coordinate shifts, and receives a reward equal to the negative of a scale-free cost function computed from FDTD (Meep). For each target, a single converged design is shown and characterized with eigenchannel, phase-winding, coalescence, or participation-number diagnostics.","tokens_in":9831,"tokens_out":7178,"duration_ms":82802,"significance":"If the demonstrations are robust, the paper would provide a conceptually simple and general RL-based route to inverse-design scattering-matrix singularities in disordered media, complementing existing local-optimization and supervised-learning approaches. The external FDTD simulator is an appropriate ground truth, the code is open-sourced, and the physical diagnostics (phase winding, eigenvector coalescence, perturbation scaling) are independent of the training reward, which is a genuine strength. The principal scientific risk is that none of the designs is characterized by an absolute transmission scale, leaving the possibility that the claimed targets are realized at negligible transmission; the lack of repeated runs further limits the evidence for the methodology's reliability.","major_comments":[{"comment":"The claim that a scale-free cost prevents the RL from driving the matrix scale toward zero is not correct. All three costs in Table 1 are also minimized by a zero or arbitrarily weak transmission matrix: the rank-1 cost vanishes for any rank-1 matrix (including tau_1 -> 0), the degenerate-eigenvalue cost vanishes for nilpotent matrices of arbitrarily small norm, and the degenerate-transmission-eigenvalue cost vanishes for T=0 (where the denominator is epsilon). The manuscript reports only the smallest transmission eigenvalue (1.06e-7) in Fig. 2b and never reports the dominant eigenvalue, the matrix norm, or the total transmitted power. Please report the absolute scale of the optimized transmission matrices, or add an explicit scale constraint, so that the designed systems can be distinguished from vacuous near-zero-transmission configurations.","section":"Table 1"},{"comment":"Each target is demonstrated only once, and the text's phrase 'multiple converged results' in the Results section refers to one example per target. No seed variation, success rates, or error bars are provided, and the training curve in Fig. 1c is described only as 'typical.' Without these, the claim that PPO reliably navigates the non-convex landscape to these inverse-design targets is not statistically supported, nor is there any comparison with simple baselines such as random search or evolutionary optimization. Please provide multiple independent runs (varying initial conditions and PPO seeds) and report the distribution of outcomes, including failure cases and computational cost.","section":"Figures 2-4"}],"minor_comments":[{"comment":"The determinant formula is garbled in the rendered text; please rewrite it as det(T(f)) proportional to a product over transmission zeros divided by a product over poles, and define all symbols including the relationship of M and N to the number of channels and resonant modes.","section":"Eq. (1)"},{"comment":"The cost-function formulas are not legible as typeset; please write them explicitly in LaTeX notation, define tau_i, t_ij, epsilon, and state clearly that epsilon is a fixed regularization constant.","section":"Table 1"},{"comment":"The description of the upper training panel is ambiguous: it says the panel plots the maximum Reward for every 128 steps but also mentions a running max and a -log10(-Reward) transformation; please clarify the exact plotted quantity and axis labels.","section":"Fig. 1c"},{"comment":"The definition of eigenvector coalescence C is garbled; it should be written as a product over pairs of normalized eigenvectors of (1 - |v_i . v_j|^2) with an appropriate normalization, and the normalization constant should be stated.","section":"Degenerate eigenvalues"},{"comment":"The paragraph discussing a second-order transmission zero admits a negative result (phase winding number 0 rather than 4pi, attributed to the statistical rarity of double-zero eigenvalues) without quantitative support. Please report the achieved eigenvalue magnitudes and the number of training attempts, or state that this is an anecdotal observation.","section":"Rank-1 TM"},{"comment":"The manuscript lacks specific PPO hyperparameters, Meep simulation resolution, FDTD run times, and the scatterer boundary-handling rules beyond the 'last object wins' mention; since the code is open-sourced this is not blocking, but a short methods paragraph would improve reproducibility.","section":"Methods"},{"comment":"There are several typographical errors and notation inconsistencies, including 'frequncy', 'propogating', 'orthorgonal', and 'choosen'; a careful proofreading pass is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The core idea is appealing and the physical diagnostics are reasonable, but the empirical evidence is currently too thin: the scale-free cost functions can be trivially minimized at zero transmission, and each target is shown only once. These issues are fixable within the manuscript's scope by reporting absolute transmission values, adding a scale constraint or comparison, and providing repeated-run statistics. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it does something genuinely new: it uses PPO to position scatterers in a 2D billiard so that the transmission matrix has a prespecified singularity—rank-1 with a transmission zero, an exceptional point with degenerate eigenvalues, or degenerate transmission eigenvalues. Second, the central worry from the stress-test is real: the cost functions in Table 1 are scale-free, so the zero (or arbitrarily weak) transmission matrix also minimizes them, and the paper never reports absolute transmission. That is a load-bearing gap, not a quibble.\n\nWhat is good: the work applies RL to a target that is not just a scalar figure of merit but a matrix property, and it grounds the demonstrations in Meep, an external FDTD solver. The code is on GitHub,which is reproducible evidence. The internal checks are genuinely supportive: the 2π phase winding of det(T), the eigenvector coalescence C=1, and the effective channel number N_eff=2 are scale-invariant, non-trivially matched, and not fitted. The author also honestly reports a failed attempt (double zero eigenvalues) and a mismatch between the predicted 0.5 and observed 0.9 exponent, with a reasonable explanation.\n\nThe soft spots are in proportion. The scale-free cost issue is the biggest: the paper's claim that a scale-free cost prevents the RL from driving the matrix scale toward zero is backwards—scale invariance removes any penalty for scale collapse. Without reporting the largest transmission eigenvalue or total transmitted power, the rank-1 and degenerate-τ designs could be vacuous near-zero-transmission solutions. The check that det(T) winds by 2π does show the rank-1 case is not the zero matrix, but the magnitude of the nonzero channel is unknown. Second, each target is demonstrated once, with no repeated runs, no seeds, no error bars, and no success rate; this is a common weakness in RL papers but it matters here because the cost landscape is highly non-convex. Third, there is no comparison to a simpler optimizer (e.g., CMA-ES or gradient-free search), so the claimed advantage of RL over existing methods is not established.\n\nWho is this for: researchers in wave chaos and photonic inverse design who care about scattering-matrix singularities. It deserves serious referee time because the idea is new, the code is available, and the simulation-based evidence is non-trivial. But the author needs to supply absolute transmission strengths, multi-seed statistics, and ideally a baseline comparison. I would engage with it if I were in the area, but I would not cite it as a robust demonstration until those numbers are added.","headline":"A plausible new RL application to transmission-matrix inverse design, but the scale-free cost functions leave the actual transmission strength unreported, which is the main gap to fix.","tokens_in":10379,"tokens_out":3161,"would_cite":false,"duration_ms":38125,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that reinforcement learning can inverse-design the transmission matrix of a random scattering system, demonstrating rank-1, exceptional-point, and degenerate transmission matrices in a 2D billiard cavity.","keywords":["inverse design","transmission matrix","reinforcement learning","PPO","exceptional points","transmission zeros","random scattering","billiard cavity"],"falsifier":"Run the released code from ten random seeds and count how often each seed reaches the stated cost threshold for the rank-1 target within a fixed episode budget; if most seeds fail, or if a greedy local search using the same number of FDTD evaluations matches the performance, the paper's claim that RL reliably solves this inverse-design problem is falsified.","tokens_in":9378,"feed_emoji":"🎛️","tokens_out":7408,"duration_ms":68749,"temperature":0.7,"pith_summary":"This paper claims that a reinforcement-learning agent can perform inverse design on a random scattering system by moving its dielectric scatterers one small step at a time. Working with a two-port 2D billiard cavity at 15 GHz and using Proximal Policy Optimization with an FDTD solver for each move, the author achieves three target transmission matrices: a rank-1 matrix with a zero-transmission mode, a matrix with degenerate eigenvalues and unidirectional mode conversion at an exceptional point, and a matrix with degenerate transmission eigenvalues whose transmitted power is independent of the input pattern. The author argues this matters because random media have many degrees of freedom and strongly oscillatory objective landscapes, so conventional optimization gets trapped, whereas RL's stochastic exploration lifts the baseline of performance over episodes. If the demonstrations hold, the framework offers a general recipe for engineering scattering matrices rather than individual devices.","feed_headline":"RL steers a random cavity to exact transmission targets","feed_subtitle":"Proximal Policy Optimization moves scatterers to create rank-1, exceptional-point, and degenerate transmission matrices.","key_machinery":"The central object is the $2\\times2$ transmission matrix $T$ of a two-port billiard, computed with an FDTD solver, and the central mechanism is the PPO policy network that proposes small scatterer displacements. The state is the normalized coordinate list $[x_1,y_1,\\ldots,x_n,y_n]$, actions are normalized increments scaled by $0.005$, and each step's reward is the negative cost. Three scale-free costs steer the matrix: for rank-1, cost $=1-\\tau_1/\\sum_i\\tau_i$; for degenerate eigenvalues, cost $=((t_{11}-t_{22})^2+4t_{12}t_{21})/(\\sum_{ij}|t_{ij}|^2+\\epsilon)$; for degenerate transmission eigenvalues, cost $=|\\tau_1-\\tau_2|/(\\sum_{ij}|t_{ij}|^2+\\epsilon)$, with $\\epsilon=10^{-8}$. Over 1024-step episodes with a mixed reset strategy that returns to the best-known configuration or starts fresh, the agent improves its episode reward and reaches configurations that satisfy the target.","core_discovery":"On the paper's own terms, the discovery is that the singularities and spectral structure of the transmission matrix of a random system are controllable by treating scatterer positions as the action space of a PPO agent. The agent's reward is the negative of a scale-free cost function, so the optimization cannot cheat by shrinking the overall matrix. The converged configurations show a rank-1 transmission matrix with a lowest transmission eigenvalue of $1.06\\times10^{-7}$ and $2\\pi$ phase winding of $\\det(T)$ around the transmission-zero frequency; degenerate eigenvalues with eigenvector coalescence $C=1$ and asymmetry factor $A=1$ for the Jordan-type mode basis; and degenerate transmission eigenvalues with eigenchannel participation number $N_{\\mathrm{eff}}=2$ at the target frequency. The paper concludes that RL is a general methodology for inverse design of scattering matrices in random media.","pith_inferences":["The paper leaves open whether the same PPO loop beats a simple hill-climbing search with the same number of FDTD calls; if it does not, the benefit of RL would be exploratory rather than computational.","The reported asymmetry exponent of 0.9, versus 0.5 predicted by first-order perturbation theory, suggests that scatterer displacements couple nonlinearly to the transmission-matrix perturbation; deriving that coupling map could turn the empirical fit into a predictive design rule.","The rank-1 and degenerate-transmission targets are global properties of $T$ at one frequency; the same cost logic could be extended to bandwidth-averaged costs to design devices that hold their behavior over a frequency window.","The framework's practical value in experiments will depend on whether the optimized scatterer positions remain effective under fabrication disorder; adding a robustness term to the cost would be a direct extension."],"forward_implications":["A rank-1 transmission matrix acts as a fixed-ratio power splitter: any input wavefront produces the same output speckle pattern up to a scalar, and an input orthogonal to the feature vector gives a transmissionless mode.","Degenerate eigenvalues realize an exceptional point at the target frequency, giving unidirectional mode conversion and enhanced sensitivity to perturbations that scale as a fractional power of the perturbation.","Degenerate transmission eigenvalues make the transmitted power independent of the input speckle pattern, since $T$ becomes proportional to a unitary matrix and $N_{\\mathrm{eff}}=N$.","The RL loop can replace the transmission matrix with any linear response matrix, so the same framework can target focusing inside a random medium or a prescribed density-of-states spectrum.","Because the reward is scale-free and no surrogate model is used, the method is aimed directly at the singularities of the transmission matrix, which auxiliary predictors handle poorly."],"supporting_citations":[{"why":"Supplies the determinant rational-function form and the topological phase-winding interpretation of transmission zeros.","marker":"[3]"},{"why":"Provides the two-channel scattering-singularity context of 50:50 power splitting that motivates rank-1 control.","marker":"[7]"},{"why":"Supplies the PPO algorithm that is the core optimizer of the inverse-design loop.","marker":"[29]"},{"why":"Basis for the Heidelberg-model expression of det(T) as poles and zeros, used to link transmission zeros to eigenvalues.","marker":"[31]"},{"why":"Provides the FDTD solver used to evaluate the transmission matrix at every RL step.","marker":"[34]"},{"why":"Supplies the definition of the asymmetry factor A used to quantify unidirectional mode conversion.","marker":"[36]"},{"why":"Supplies the eigenchannel participation number N_eff used as the metric for degenerate transmission eigenvalues.","marker":"[37]"},{"why":"Identifies the interior field-channel matrix whose replacement for T would extend the framework to focusing inside random media.","marker":"[38]"},{"why":"Links the phase derivative of det(T) to the density of states, motivating inverse design of quasi-normal modes.","marker":"[39]"}],"fun_headline_variants":["RL shapes transmission matrices in random systems","PPO hits transmission targets via inverse design","RL crafts exact transmission matrices in disorder","RL tunes random cavities to exact transmission modes","Reinforcement learning designs scattering matrices directly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that PPO, with this state/action parameterization, these cost functions, and 1024-step episodes, can navigate the highly non-convex and oscillatory objective landscape to a good solution within a reasonable number of episodes; the paper shows single successful examples but no success rate, seed variation, or comparison to other optimizers.","fun_headline_variants_meta":{"raw":{"variants":["RL shapes transmission matrices in random systems","PPO hits transmission targets via inverse design","RL crafts exact transmission matrices in disorder","RL tunes random cavities to exact transmission modes","Reinforcement learning designs scattering matrices directly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001136,"raw_usage":{"total_tokens":4637,"prompt_tokens":786,"completion_tokens":3851,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":402,"completion_tokens_details":{"reasoning_tokens":3787}},"tokens_in":402,"tokens_out":3851,"duration_ms":32703,"temperature":1.0,"reasoning_tokens":3787,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:36:09.432572+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released code from ten random seeds and count how often each seed reaches the stated cost threshold for the rank-1 target within a fixed episode budget; if most seeds fail, or if a greedy local search using the same number of FDTD evaluations matches the performance, the paper's claim that RL reliably solves this inverse-design problem is falsified.","supporting_citations":[{"cited_title":"Kang and A","cited_arxiv_id":null,"evidence_quote":"Supplies the determinant rational-function form and the topological phase-winding interpretation of transmission zeros."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the two-channel scattering-singularity context of 50:50 power splitting that motivates rank-1 control."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Basis for the Heidelberg-model expression of det(T) as poles and zeros, used to link transmission zeros to eigenvalues."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the FDTD solver used to evaluate the transmission matrix at every RL step."},{"cited_title":"Feng, Y.-L","cited_arxiv_id":null,"evidence_quote":"Supplies the definition of the asymmetry factor A used to quantify unidirectional mode conversion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the eigenchannel participation number N_eff used as the metric for degenerate transmission eigenvalues."},{"cited_title":"Cheng and A","cited_arxiv_id":null,"evidence_quote":"Identifies the interior field-channel matrix whose replacement for T would extend the framework to focusing inside random media."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Links the phase derivative of det(T) to the density of states, motivating inverse design of quasi-normal modes."}],"review_version":1}