{"id":"fe9b920a-23df-4e6d-832c-59c380b672f8","arxiv_id":"2509.03835","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A neural image codec trained with a BiLSTM channel model fitted to a physical free-space optical link transmits images over strong atmospheric turbulence about 3.5 dB better in MS-SSIM than a codec trained under AWGN.","lead":"Free-space optical links carrying images through turbulent air are hard to keep stable. This paper trains a neural image codec using a learned model of the real laser channel, and shows about 3.5 dB better image quality under strong turbulence than a codec trained assuming simple Gaussian noise.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BiLSTM channel model is trained on pre-fine-tuning symbols but held fixed while the encoder shifts; if surrogate accuracy degrades, the claimed 3.54 dB gain may not stem from faithful channel modeling.","rationale":"The reader's weakest_assumption correctly identifies the distribution shift between the pre-trained encoder's symbols (used to train the BiLSTM) and the fine-tuned encoder's symbols (used during end-to-end training). This is the most load-bearing concern because the paper's central claim is not merely empirical—it explicitly attributes the performance gain to the learned channel model. If the surrogate is inaccurate for the final encoder distribution, the training objective in stage 3 is mismatched to the real channel, and the reported real-FSO improvement cannot be confidently credited to the proposed channel-modeling approach. The concern is concrete and testable: re-measuring the channel-model fidelity after fine-tuning would directly settle whether the surrogate remains valid. I also note secondary issues—the MS-SSIM 'dB' metric is never defined, Table I and code are missing, and there are inconsistencies (e.g., 'six' vs 'seven' datasets, PAM4/OOK vs PAM4/PAM8)—but these do not displace the distribution-shift problem as the primary threat to the causal argument. The paper's empirical result may still hold, but the mechanism is unproven; a CONDITIONAL verdict with a request for the above validation is appropriate. I agree with the reader's assessment, so the verdict should remain CONDITIONAL rather than change to ACCEPT or REJECT.","tokens_in":10846,"tokens_out":7288,"duration_ms":70810,"concrete_test":"After stage-3 fine-tuning, encode a held-out image set with the final JEM to produce symbols y_f. Transmit y_f through the physical FSO testbed (SLM turbulence) to obtain y_r, and feed y_f through the fixed BiLSTM channel model to obtain y_m. Compute the KL divergence between the amplitude distributions of y_m and y_r using the same procedure as Section IV.A. If the KL divergence rises from the reported 0.0019 to, say, >0.01 (an order of magnitude worse) or is significantly larger than the pre-fine-tuning value, the surrogate is no longer faithful for the optimized encoder, undermining the causal attribution. A stronger check: retrain the BiLSTM on the fine-tuned symbol pairs and re-run stage 3; if the 3.54 dB gain over TRFSO-AWGN shrinks or disappears, the original gain was partly an artifact of surrogate mismatch.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central causal claim is that training with the BiLSTM channel model—fit to physical FSO data—is what gives TRFSO its 3.54 dB MS-SSIM improvement over TRFSO-AWGN under strong turbulence (Section IV.B). The weakest point is the fixed surrogate assumption in the three-stage training (Section II.D). In stage 2, the BiLSTM is trained on symbol pairs (y', y'') where y' comes from the pre-trained AWGN encoder (stage 1). In stage 3, the channel model parameters are frozen and the JEM/JDD are fine-tuned end-to-end. Nothing in the paper checks whether the BiLSTM remains accurate for the new distribution of encoder outputs. The channel model is trained with MSE (Eq. 2) to predict the conditional mean of the real channel; if the fine-tuned symbols fall in a different region of the input space, the surrogate can become systematically biased. The training trick of randomly sampling multiple turbulence-level channel models does not address input distribution shift—it only changes which fixed surrogate is used. As a result, the gradients in stage 3 may be computed against an incorrect channel, and the measured 3.54 dB improvement on the real FSO link may be driven by other factors (e.g., the PAPR regularizer, the Swin Transformer capacity, or implicit robustness from multi-level training), not by the channel model's fidelity. The paper's conclusion explicitly credits the channel model ('finding validates the critical role of accurate channel modeling'), so this unsupported link is load-bearing. The empirical improvement could be real, but the causal mechanism is not established without validating the surrogate on the fine-tuned encoder's output distribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TRFSO, a turbulence-robust end-to-end FSO image transmission system that combines JSCCM based on Swin Transformer blocks with a BiLSTM-based differentiable channel model. The channel model is trained on input/output symbol pairs collected from a physical 30 Gbaud IM/DD FSO link under SLM-emulated turbulence. The system is trained in three stages: (1) pre-training the JEM/JDD under AWGN, (2) training the BiLSTM channel model on the pre-trained encoder's symbols, and (3) fine-tuning the full system end-to-end with the channel model frozen. The authors report that under strong turbulence TRFSO achieves an average 3.54 dB MS-SSIM improvement over a same-architecture model trained only under AWGN, and they attribute this gain to the accuracy of the BiLSTM channel model, which achieves KL divergence as low as 0.0019 against measured amplitude distributions. They also show favorable results under varying zenith angles and ground-station altitudes.","tokens_in":11250,"tokens_out":3664,"duration_ms":36556,"significance":"The paper addresses a timely and practical problem: making JSCCM robust to non-Gaussian, time-varying FSO channels. The experimental setup is a real 30 Gbaud IM/DD FSO link with SLM-emulated turbulence, which is a notable strength compared with simulation-only studies. If the 3.54 dB improvement is genuinely attributable to the learned channel model, it would provide strong evidence for the value of data-driven channel surrogates in end-to-end optical communication. However, the central causal claim rests on an unvalidated assumption about the frozen channel model's accuracy after encoder fine-tuning, and the quantitative metric is reported in dB without a defined conversion. These issue need to be resolved before the result can be considered established.","major_comments":[{"comment":"The BiLSTM channel model is trained on symbols y' from the AWGN-pre-trained encoder (stage 2) and then held fixed while JEM and JDD are fine-tuned end-to-end (stage 3). As the encoder weights change, the distribution of transmitted symbols shifts, potentially outside the region where the surrogate was trained. The paper neither re-validates the channel model on fine-tuned symbols nor adapts it. Since the conclusion explicitly credits accurate channel modeling for the 3.54 dB gain (Section IV.B), the authors should provide evidence that the frozen surrogate remains accurate during and after stage 3. A concrete test is to compare the BiLSTM output against actually measured received symbols using the fine-tuned encoder; if mismatch is significant, retraining or iterative refinement is needed. Without such a check, the improvement could stem from other components (e.g., the PAPR regularizer,","section":"II.D (three-stage training) and IV.B"},{"comment":"All quantitative comparisons are reported in dB (e.g., 'MS-SSIM improvement of 3.54 dB', 'MS-SSIM values remain above 21 dB'), but the mapping from the MS-SSIM index to dB is never defined. Without a formula, such as MS-SSIM_dB = -10*log10(1-MS-SSIM) or equivalent, the numerical values are not interpretable or reproducible. This is particularly load-bearing because the main headline improvement is in dB. Please specify the exact conversion and apply it consistently.","section":"IV.B (all MS-SSIM claims)"},{"comment":"The channel model is trained on data from the same physical FSO testbed used for the final performance evaluation. While the evaluation bypasses the channel model and therefore is not circular, it does limit generalization: the measured improvements may be specific to this link's SLM-emulated turbulence characteristics. Please state whether the evaluation turbulence realizations are disjoint from the training realizations, and discuss how the results would transfer to different links or real atmospheric paths. At minimum, this should be acknowledged as a limitation.","section":"III (experimental setup) and IV"}],"minor_comments":[{"comment":"Table I is referenced but not included in the manuscript. Please include it, as the fairness argument (comparable number of transmitted symbols) depends on it.","section":"III, Table I"},{"comment":"Section II.D says 'six different noise levels' for channel-model training, while Section IV.A says 'seven datasets'. Please make the counts consistent.","section":"II.D vs IV.A"},{"comment":"Section III lists PAM4 and PAM8 as the modulation formats for the traditional baseline, but Section IV.B and Fig. 5 refer to 'PAM4 and OOK'. Clarify which baselines were actually used.","section":"III and IV.B"},{"comment":"The abstract states '3.5 dB improvement', while Section IV.B reports '3.54 dB'. Please use one consistent rounded value.","section":"Abstract and IV.B"},{"comment":"The reshaping of the symbol sequence into a 2D tensor of shape N×M is not fully defined. Specify which dimension is the sequence length and which is the feature dimension, and relate M to the BiLSTM input size.","section":"II.C"},{"comment":"The image distortion loss L_d(x, x̂) in Eq. (1) is not defined. Please specify whether it is the MS-SSIM loss, MSE, or a combination.","section":"II.B, Eq. (1)"},{"comment":"The conclusion states MS-SSIM levels 'exceeding 21.3 dB' under moderate turbulence, but this value does not appear in the results section. Please reconcile the numbers or cite the specific figure.","section":"V"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound in its experimental infrastructure, but the key attribution claim needs strengthening. The most important issue is the fixed channel model under encoder fine-tuning, which could invalidate the surrogate and make the reported gain an artifact of other training factors. The MS-SSIM dB conversion is an easily fixable but fundamental reporting gap. I recommend major revision, not rejection, because the core idea is promising and the issues can be addressed with additional experiments (e.g., validation of the channel model after fine-tuning, possibly retraining) and clearer metrics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely useful experimental paper—30 Gbaud IM/DD FSO link with SLM-emulated turbulence, and a head-to-head comparison showing that training the same JSCCM network with a learned channel model beats AWGN training under strong turbulence by ~3.5 dB MS-SSIM. That result is worth having.\n\nWhat's new: the BiLSTM channel model itself follows [22], and the JSCCM architecture builds on the authors' prior work, but the contribution is the physical integration — real FSO hardware, six turbulence levels, 120k training symbols per level, and the three-stage training recipe. The KL divergence of 0.0019 on amplitude distributions is evidence the surrogate captures marginal statistics. Also, they show that the AWGN-trained version wins under AWGN, which is a sensible sanity check.\n\nThe main soft spot is the fixed-surrogate shift. The BiLSTM is trained on symbols from the stage-1 AWGN encoder, then frozen while the encoder is fine-tuned. The paper never checks whether the surrogate remains accurate for the fine-tuned encoder's output distribution. If the encoder moves to a region where the BiLSTM is biased, the stage-3 gradients are wrong, and the 3.54 dB gain could come from other ingredients (PAPR regularization, multi-level sampling, Swin capacity). That makes the concluding claim—'validates the critical role of accurate channel modeling'—stronger than the evidence supports.\n\nAlso: evaluation is on the same physical setup that generated the channel-model training data; there's no independent testbed. MS-SSIM reported in dB is never defined. Table I is referenced but not present. Minor inconsistency: six vs. seven channel datasets. These are fixable but need attention.\n\nThe paper is not conceptually wrong. The central idea is sensible, and the measurements are plausible. But the causal story needs one additional experiment: retrain the BiLSTM on the fine-tuned encoder's symbols (or at least validate its accuracy after fine-tuning). That would close the loop.\n\nWho it's for: people working on end-to-end learning for optical channels, and anyone building learned surrogates for physical layers. It deserves a serious referee. I'd send it to review with the request that the authors address the distribution-shift check and provide the missing table and code. If those checks land, the result is solid.","headline":"Solid experimental demonstration of turbulence-robust FSO image transmission, but the causal case that the BiLSTM channel model is the source of the gain is not fully closed.","tokens_in":11747,"tokens_out":2382,"would_cite":true,"duration_ms":22266,"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":"The paper argues that the main reason end-to-end learned FSO links fail in the field is that they are trained against a generic Gaussian noise channel; a BiLSTM surrogate of the real channel closes that gap by an average 3.54 dB in MS-SSIM","keywords":["free-space optical communication","end-to-end learning","joint source-channel coding","BiLSTM channel model","atmospheric turbulence","image transmission","MS-SSIM","Swin Transformer"],"falsifier":"Recompute the amplitude-distribution KL divergence between the frozen BiLSTM surrogate and the physical link using symbols from the fine-tuned TRFSO encoder (not the pre-trained one) under the same turbulence levels. If the divergence rises materially above the 0.0019 value quoted for training-time symbols, the surrogate is stale; equivalently, retrain the surrogate on the fine-tuned encoder's symbols and check whether the 3.54 dB MS-SSIM advantage over TRFSO-AWGN persists or shrinks.","tokens_in":10794,"feed_emoji":"📡","tokens_out":6320,"duration_ms":59173,"temperature":0.7,"pith_summary":"This paper aims to show that the main reason learned end-to-end optical links fail under real atmospheric turbulence is that they are trained against a generic additive Gaussian noise channel. The authors build TRFSO, an image-transmission system whose encoder and decoder are trained end-to-end through a bidirectional LSTM (BiLSTM) surrogate of the actual physical FSO channel, learned from symbol pairs measured on a 30 Gbaud intensity-modulation/direct-detection link with controllable turbulence. The surrogate matches the real channel's amplitude statistics to a KL divergence as low as 0.0019, and the system trained through it beats its AWGN-trained twin by an average 3.54 dB in MS-SSIM under strong turbulence while adding no deployment-time cost. If this holds, accurate differentiable channel modeling -- not only network architecture -- is what buys robustness in turbulence-impaired optical links.","feed_headline":"Real-world channel model improves FSO image quality by 3.5 dB","feed_subtitle":"Training through a learned physical-link model beats AWGN training when turbulence is strong.","key_machinery":"The load-bearing mechanism is the frozen BiLSTM channel surrogate: two stacked bidirectional LSTM layers (hidden size 30, dropout 0.15), a two-layer fully connected projection, and a power-normalization layer, trained by MSE on (transmitted, received) symbol pairs from the physical link. It acts as a differentiable substitute for the real FSO channel during end-to-end training, so gradients from the reconstruction loss flow through realistic non-Gaussian, temporally correlated distortions. Its parameters are fixed in the final training stage, and turbulence robustness is instilled by randomly drawing from the fitted surrogate variants corresponding to different turbulence levels in each trai","core_discovery":"The central claim is that replacing the AWGN channel layer with a BiLSTM channel model trained on real FSO measurements makes a joint source-channel coding and modulation (JSCCM) system turbulence-resilient without changing the deployed encoder/decoder. The authors train Swin Transformer-based joint encoding/modulation and decoding/demodulation networks, first pre-train them under AWGN, collect transmitted/received symbol pairs over an IM/DD FSO link with SLM-generated phase screens, train the BiLSTM surrogate on those pairs, then freeze it and fine-tune the whole system end-to-end while randomly sampling turbulence levels. Under strong turbulence (Cn2 = 10^-13 m^-2/3), TRFSO sustains most f","pith_inferences":["If the surrogate's fidelity is the real driver, the same three-stage recipe should transfer to other channels that are non-differentiable or analytically intractable, such as underwater optical or RF fading with memory.","A sharper test would measure the KL divergence of the frozen surrogate on symbols emitted by the fine-tuned encoder: if it climbs well above the 0.0019 training-time value, the reported 3.54 dB gain could be partly an artifact of optimizing against a stale channel.","The paper averages MS-SSIM in dB; looking at per-realization distributions, the system-level benefit may be larger than the average because TRFSO avoids total failures that separate schemes suffer under strong turbulence, where the mean alone underweights catastrophic outliers.","One testable extension is to iterate the pipeline by re-collecting physical pairs from the fine-tuned encoder and retraining the surrogate, which could either widen the gain or reveal a ceiling imposed by surrogate accuracy."],"forward_implications":["The reported gain is attributable to the channel model alone: TRFSO and TRFSO-AWGN share identical encoder/decoder architecture and training hyperparameters, so the 3.54 dB improvement isolates the effect of training through the learned surrogate.","Training costs increase only at training time; deployment uses the same encoder and decoder, so turbulence robustness comes without extra hardware such as adaptive optics.","Because training samples across turbulence levels, a single fixed encoder/decoder maintains graceful degradation across zenith angles and altitudes, with MS-SSIM still 14.88 dB at a 60-degree zenith angle.","On turbulence-free links approximated by AWGN, the AWGN-trained system remains the better choice, implying the training channel model should match the deployment scenario."],"supporting_citations":[{"why":"Supplies the discrete-time analog transmission scheme and the PAPR-aware loss strategy that the proposed system builds on.","marker":"[11]"},{"why":"Prior end-to-end FSO learning with a BiLSTM channel model that this paper extends to physical-link data collection and JSCCM integration.","marker":"[22]"},{"why":"Provides the Swin Transformer backbone used for the joint encoding/modulation and decoding/demodulation networks.","marker":"[23]"},{"why":"Phase-screen generation method used to emulate atmospheric turbulence on the SLM for data collection and evaluation.","marker":"[25]"},{"why":"JPEG2000 source coder used in the separate-coding baseline that TRFSO is compared against.","marker":"[26]"},{"why":"LDPC channel code used in the separate-coding baseline; defines the fairness comparison in transmitted symbols.","marker":"[27]"},{"why":"MS-SSIM metric used to quantify image reconstruction quality for all compared schemes.","marker":"[28]"},{"why":"DIV2K training set for the JSCCM encoder/decoder.","marker":"[29]"},{"why":"RSOD remote-sensing evaluation set used to test image transmission quality.","marker":"[30]"}],"fun_headline_variants":["BiLSTM channel model boosts FSO image quality by 3.5 dB","Real-world channel modeling beats AWGN for turbulent FSO links","Turbulence-robust FSO transmission with learned channel model","Learned channel model improves FSO robustness by 3.5 dB","FSO training on real turbulence data beats AWGN by 3.5 dB"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The frozen BiLSTM channel model is trained on symbols emitted by the pre-trained AWGN encoder, and the whole fine-tuning step assumes that same surrogate stays accurate for the different symbol distribution the encoder produces after training.","fun_headline_variants_meta":{"raw":{"variants":["BiLSTM channel model boosts FSO image quality by 3.5 dB","Real-world channel modeling beats AWGN for turbulent FSO links","Turbulence-robust FSO transmission with learned channel model","Learned channel model improves FSO robustness by 3.5 dB","FSO training on real turbulence data beats AWGN by 3.5 dB"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1152,"prompt_tokens":747,"completion_tokens":405,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":308}},"tokens_in":491,"tokens_out":405,"duration_ms":3732,"temperature":1.0,"reasoning_tokens":308,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:37:06.117218+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the amplitude-distribution KL divergence between the frozen BiLSTM surrogate and the physical link using symbols from the fine-tuned TRFSO encoder (not the pre-trained one) under the same turbulence levels. If the divergence rises materially above the 0.0019 value quoted for training-time symbols, the surrogate is stale; equivalently, retrain the surrogate on the fine-tuned encoder's symbols and check whether the 3.54 dB MS-SSIM advantage over TRFSO-AWGN persists or shrinks.","supporting_citations":[{"cited_title":"Atmospheric Turbulence-Immune Free Space Optical Communication System based on Discrete -Time Analog Transmission","cited_arxiv_id":null,"evidence_quote":"Supplies the discrete-time analog transmission scheme and the PAPR-aware loss strategy that the proposed system builds on."},{"cited_title":"End -to-End Learning for Free Space Optical Communication with BiLSTM -based Channel Model,","cited_arxiv_id":null,"evidence_quote":"Prior end-to-end FSO learning with a BiLSTM channel model that this paper extends to physical-link data collection and JSCCM integration."},{"cited_title":"Simulation of wave propagation in three-dimensional random media,","cited_arxiv_id":null,"evidence_quote":"Phase-screen generation method used to emulate atmospheric turbulence on the SLM for data collection and evaluation."},{"cited_title":"Low-density parity-check codes,","cited_arxiv_id":null,"evidence_quote":"LDPC channel code used in the separate-coding baseline; defines the fairness comparison in transmitted symbols."},{"cited_title":"Elliptic Fourier transformation- based histograms of oriented gradients for rotationally invariant object detection in remote-sensing images,","cited_arxiv_id":null,"evidence_quote":"RSOD remote-sensing evaluation set used to test image transmission quality."}],"review_version":1}