{"id":"844c9c4d-20f8-48ed-a1cc-9c5999bfc40a","arxiv_id":"2608.06203","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":9,"one_line_summary":"A proposed near-field radiative thermal neural network maps deep-learning operations onto heat exchange between nanoscale surfaces, supported only by a simulated inverse-modeling demonstration.","lead":"The paper proposes a computing scheme in which heat flowing between nanoscale surfaces, not electricity, performs convolutional and recurrent neural-network operations using phase-change materials. It reports simulation results for the idea, but no physical demonstration of a working thermal neural network.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper asserts that Eq. (1) with freely tunable W_ij^rad is realized by passive near-field radiative arrays, but no derivation links these weights to the constraints of fluctuational electrodynamics (positivity, reciprocity, energy conservation).","rationale":"The reader's weakest assumption correctly identifies the tunability and independence of W_ij^rad as a central risk. I agree that crosstalk and phase-change dynamics threaten independent weight setting. My concern is broader: even in the absence of crosstalk, the paper never derives Eq. (1) from the radiative transfer physics it presents in Eqs. (7)–(11). The equations of the T-CNN and T-RNN are standard neural-network equations with arbitrarily assigned activation functions and weight matrices; they are not shown to follow from, or even be consistent with, the near-field heat-exchange Hamiltonian. Passive radiative couplings are constrained by positivity, reciprocity, and energy conservation, so arbitrary signed trainable weights are not physically available. The only numerical result is a conventional recurrent network trained on synthetic data generated by the authors' own model; it does not test the proposed hardware. The paper is an interesting proposal, but the central claim overreaches what is demonstrated. The reader's REJECT verdict is therefore appropriate; no change is needed.","tokens_in":15258,"tokens_out":3568,"duration_ms":41593,"concrete_test":"Use a rigorous many-body fluctuational-electrodynamics solver (e.g., scattering-matrix or N-body Green's function with VO2/GST optical constants) to compute the steady-state heat-flux response matrix A_ij = ∂Q_i/∂T_j for a 4x4 array with 10–100 nm gaps over 310–380 K. Then test whether any assignment of W_ij^rad exists such that Eq. (1) reproduces the computed Q_i across many temperature configurations, and check whether the required W_ij are non-negative, reciprocal, and independent of the gate-temperature schedule. If no such assignment exists, or if tuning one gate moves many W_ij beyond tolerance, the claimed programmable neural network is not physically realizable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the claim of 'the first radiative realization of thermal deep learning,' the paper must show that actual near-field radiative exchange among VO2/GST terminals obeys a programmable network update such as Eq. (1): Q_i(t) = Σ_j W_ij^rad Q_j(t) + Σ_k U_ik^rad Q_k(t−1) + b_i. This is never demonstrated. Eqs. (7)–(11) give the standard two-body flux, but the many-body array equations are simply posited; the 'weights' are never expressed in terms of dielectric functions, gaps, or temperatures, and the gradient update (Eq. 2) is not mapped to any physical gate-control mechanism. More fundamentally, passive radiative transfer is not a linear, independently programmable kernel: near-field couplings are non-negative, satisfy reciprocity, and must respect energy conservation, while Eq. (1) allows arbitrary signed weights and the text invokes 'reverse flux' (−1) states as physical outputs. The only quantitative demonstration (Fig. 5) is a software LSTM/GRU regression on synthetic Q–T curves generated by the authors' own model, not an implementation of the T-CNN/T-RNN. Thus the load-bearing assumption—that W_ij^rad are independently adjustable, unconstrained weights—is unsupported and in tension with the physics stated in the paper itself.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a programmable near-field radiative thermal network that is claimed to emulate convolutional and recurrent neural-network operations using only radiative heat exchange among nanoscale phase-change (VO2/GST) elements. Two architectures are introduced, a Thermal Convolutional Neural Network (T-CNN) and a Thermal Recurrent Neural Network (T-RNN), based on radiative diodes and transistors. The authors write neural-network equations for these architectures, propose a backpropagation-style training rule in which weight updates are supposed to correspond to gate-temperature adjustments, and report an LSTM/GRU inverse-prediction result trained on a synthetic dataset of 50,000 heat-flux--temperature pairs generated from their own forward model. The abstract states that the study 'establishes the first radiative realization of thermal deep learning,' a central claim of the paper.","tokens_in":15702,"tokens_out":4645,"duration_ms":48224,"significance":"If the central claim were established, this would be a conceptually striking result: a passive, contactless thermal platform performing convolutional and recurrent inference without charge carriers. The paper does cite and use standard near-field radiative transfer formalism (Eqs. 7--11), and the idea of exploiting phase-change hysteretic emissivity for thermal memory is plausible as a research direction. The large synthetic dataset and leave-one-out cross-validation protocol are reasonable machine-learning practices for the inverse regression task. However, the paper does not demonstrate physical learning. The T-CNN and T-RNN are neural-network equations that are posited, not derived from or validated against radiative-transfer simulations or experiments. The only quantitative result is a surrogate-model fit to synthetic data generated by the authors' own model, which validates the regressor rather than the proposed thermal network.","major_comments":[{"comment":"The core network equation Q_i(t) = Σ_j W_ij^rad Q_j(t) + Σ_k U_ik^rad Q_k(t−1) + b_i is posited without derivation. The paper presents the standard two-body flux expressions in Eqs. (7)--(11), but never expresses W_ij^rad or U_ik^rad in terms of dielectric functions, gap sizes, or temperatures, nor shows that the many-body near-field exchange among VO2/GST terminals takes this linear, recurrent form. The sentence 'The tunability of ε yields variable radiative coupling coefficients that act as trainable convolutional weights' does not establish that the couplings are independently adjustable or that the linear superposition in Eq. (1) is consistent with fluctuational electrodynamics. Since this equation is the foundation of the entire T-CNN/T-RNN framework, the central physical claim is unsupported.","section":"Thermal Radiation Convolutional Neural Network, Eq. (1)"},{"comment":"Equation (2) is a standard gradient-descent update, and the paper states that 'the weight update physically corresponds to modifying gate temperature distributions across the array.' No physical mechanism is given that maps a computed gradient ∂L/∂W to a specific gate-temperature change, nor is there any analysis of whether such changes can realize the signed weights and 'reverse flux' (−1) logic states that the tri-state scheme requires. Passive near-field radiative couplings are non-negative and satisfy reciprocity, while Eq. (1) and the described +1/0/−1 classification allow arbitrary signed weights and even 'α ≤ 0' states. This internal tension is load-bearing because the claimed computation depends on the ability to set weights independently and with arbitrary sign.","section":"Weight update and tri-state logic, Eqs. (2)--(3) and surrounding text"},{"comment":"The only quantitative result in the paper is the training of LSTM and GRU networks on 50,000 synthetic Q–T pairs generated by the authors' own forward model (100 filling ratios × 500 temperatures). This is a supervised regression on simulated data from a standard near-field calculation; it validates the recurrent regressor as a surrogate model, not the proposed T-CNN/T-RNN architecture. Moreover, the dataset generation 'suppose[s] the phase transition process is completed in just one instant at 341K,' which explicitly ignores the hysteretic memory that is elsewhere claimed to be the physical basis of recurrent computation. The paper therefore does not demonstrate that the radiative thermal network itself performs convolutional or recurrent inference.","section":"Discussion and Fig. 5 (LSTM/GRU inverse prediction)"},{"comment":"The caption to Fig. 4 states that panel (a) is a 'Raw temperature field captured by FLIR camera during gate modulation,' and panel (b) is a 'Post processed map' identified by the T-CNN algorithm. No experimental setup, sample fabrication, measurement protocol, or uncertainty analysis is provided anywhere in the manuscript. If this is meant as experimental evidence of radiative thermal computing, the omission is fatal to the claim of a 'radiative realization.' If it is instead a simulation or illustrative overlay, then describing it as a camera capture is misleading. In either case, the present manuscript does not supply the evidence needed for the paper's headline claim.","section":"Fig. 4 and Discussion text"}],"minor_comments":[{"comment":"Typos: 'ast as trainable' should be 'act as trainable,' and 'semiconductiong' should be 'semiconducting.'","section":"p. 8, near Eq. (1)"},{"comment":"'Theorical Fundamentals' should be 'Theoretical Fundamentals.'","section":"Section heading, p. 18"},{"comment":"The caption contains typos: 'Comparioson' should be 'Comparison' and 'with the the epochs' should be 'with the epochs.'","section":"Fig. 5 caption"},{"comment":"The text refers to 'Fig. 7(c)' for the training-loss curves, but the correct label is Fig. 5(c).","section":"Fig. 5 and text, p. 17"},{"comment":"The text says 'These feedback via κ_{1,2}^T into a_{1,2}[t+1]' after defining the storage states s_1 and s_2, but according to Eqs. (69) and (73) the storage outputs h_S couple through κ^S, not κ^T. This notation should be corrected consistently.","section":"Eqs. (88)--(91) and surrounding text"},{"comment":"Reference 53 is an empty entry, and reference 3 (Donghee Lee et al., Physical Review D 112 (2025)) appears unrelated to thermal logic or neural networks; these should be corrected or removed.","section":"References"}],"recommendation":"reject","confidential_remarks":"The referee agrees with the reader's assessment: the paper's central claim of a 'first radiative realization of thermal deep learning' is not supported by any derivation, simulation, or experiment that connects the proposed neural-network equations to actual near-field radiative transport. The mismatch between the abstract's strong claim and the content (a speculative proposal plus a surrogate-model regression) is substantial. The issues are load-bearing and would require new derivations and validation experiments beyond the scope of a normal revision. There are also presentation problems (rampant typos, an empty reference, an apparently unrelated reference, lack of data/code availability). I would not invite a major revision unless the authors can provide a genuine many-body derivation of the network equations and a numerical or experimental demonstration of the T-CNN/T-RNN operations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know before reading: this paper is not a demonstration of thermal deep learning. It is a forward-looking proposal that maps near-field radiative heat transfer onto convolutional and recurrent network equations, and then trains ordinary LSTM/GRU models on synthetic Q–T curves generated from the authors' own forward model. The load-bearing step—that the coupling coefficients in Eq. (1) are independently programmable signed weights—is never derived from fluctuational electrodynamics, and sits in tension with the positivity, reciprocity, and energy-conservation constraints of passive radiative transfer. The paper's own 'reverse flux' (−1) states as physical outputs are something passive two-body radiative exchange cannot produce.\n\nNow the credit. The paper assembles established ingredients—near-field radiative diodes and transistors, VO2/GST phase-change hysteresis, and neural-network mathematics—into explicit T-CNN and T-RNN architectures with complete forward/backward equations. For a reader who wants a concrete template of how one might try to map a thermal device stack onto a CNN/RNN, the notation is careful and consistent. The inverse-prediction experiment is honest about its setup, even if it only validates a regressor on self-generated data.\n\nThe soft spots are in proportion. The central gap: W_ij^rad are treated as free parameters, but no formula connects them to dielectric functions, gap sizes, or temperatures. The gradient update of Eq. (2) is said to correspond to gate-temperature changes, but no physical control mechanism is shown. Fig. 4 is an extrema-detection overlay on a thermal field, not a network output. Fig. 5 is a standard LSTM/GRU regression, not a T-RNN. The framework also introduces under-defined entities—'thermal screen' and 'secondary regulator'—and a swarm of free parameters (α, γ, β, λ, κ, etc.) that are fitted by software training rather than fixed by the physics.\n\nThe math itself is internally consistent; the neural-network algebra is fine. The citation pattern leans heavily on the authors' own prior work, but those references are directly relevant to the near-field devices they build on, so I see no citation-integrity problem.\n\nWho gets value from this? People tracking speculative neuromorphic proposals in thermal physics may mine it for ideas. But the abstract's claim of 'the first radiative realization of thermal deep learning' is not supported by the evidence presented. I would not publish this as a demonstration. If the venue has room for forward-looking proposals, send it to referees with a mandate to check whether the weights can actually be realized under fluctuational electrodynamics; if that fails, the paper should be reframed as a proposal with no claim of physical learning. My honest recommendation: major revision at best, rejection if the overclaim stays.","headline":"This is a forward-looking proposal, not a demonstration: the paper maps near-field radiative transfer onto neural-network equations, but the load-bearing assumption of freely programmable signed weights is never derived from the physics, and the only quantitative result is a regression on synthetic data.","tokens_in":16158,"tokens_out":3065,"would_cite":false,"duration_ms":31798,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["44.40.+a","84.35.+i","64.70.K-"],"model":"deepseek-v4-flash","headline":"A programmable network of nanoscale heat exchangers can execute convolutional and recurrent neural-network operations using only near-field photon tunneling between phase-change surfaces.","keywords":["near-field radiative heat transfer","thermal neural network","phase-change materials","vanadium dioxide","GST","thermal computing","neuromorphic hardware","thermal memory"],"falsifier":"Fabricate a 2×2 array of VO2/h-BN source-gate-drain units with ~50 nm vacuum gaps, program one gate through its insulator-metal transition, and measure the heat flux between the two unaddressed nodes; if that off-target coupling shifts beyond the network's error tolerance, the independent-weight assumption underlying Eq. (1) fails.","tokens_in":15023,"feed_emoji":"🔥","tokens_out":10884,"duration_ms":97421,"temperature":0.7,"pith_summary":"This paper tries to establish that the core operations of deep learning—weighted summation, nonlinear activation, and memory—can be realized physically by near-field radiative heat transfer between nanoscale surfaces, with no charge carriers and no electrical circuits. It proposes a programmable near-field radiative thermal neural network built from phase-change thermal diodes and transistors, and arranges these into a Thermal Convolutional Neural Network (T-CNN) for spatial pattern recognition and a Thermal Recurrent Neural Network (T-RNN) for sequential inference. The authors then show numerically that a T-RNN can invert the heat-flux–temperature response of a VO2 grating to predict its filling ratio with high accuracy, which they present as evidence of a new computing paradigm in which temperature fields are the carriers of intelligence. A reader should care because, if correct, this opens a route to non-electronic, contactless, energy-conserving computing that could operate in extreme environments and store information in material hysteresis.","feed_headline":"Heat flow alone can run neural-network inference","feed_subtitle":"Phase-change gates turn temperature fields into trainable weights and memory, with no circuits.","key_machinery":"The load-bearing machinery is the near-field radiative heat-transfer formalism together with phase-change gate dynamics. Heat flux $Q_{S\\to D}$ between source and drain is computed from the Dyadic Green's function (or fluctuational electrodynamics), with a spectral transmissivity $\\mathcal{T}_{S\\to D}(\\omega,L)$ that includes both propagating modes and evanescent photon tunneling; the gate's phase-change material (VO2 or GST) changes its dielectric response with temperature, giving the sigmoidal emissivity $f(T)$ that serves as the nonlinear activation. The gain of the radiative transistor is characterized by $\\alpha=Q_{G\\to D}/Q_{S\\to D}$, with $\\alpha>1$ meaning heat amplification, and the hysteretic emissivity loop, with a phase lag of roughly 12 K for VO2, provides the recurrent memory channel through the kernel $M(t)=e^{-t/\\tau_s}$. Finally, the training rule $W_{ij}^{(n+1)}=W_{ij}^{(n)}-\\eta\\,\\delta\\mathcal{L}/\\delta W_{ij}$ is interpreted physically as reprogramming gate temperature distributions, so that weight updates correspond to real thermal control signals.","core_discovery":"The paper's central claim is that photon tunneling between closely spaced phase-change surfaces is sufficient to implement neural-network inference and learning. In this scheme the radiative coupling coefficients $W_{ij}^{\\mathrm{rad}}$ act as synaptic weights, the temperature-dependent emissivity $\\varepsilon(\\lambda,T)=\\varepsilon_m+[\\varepsilon_s-\\varepsilon_m]f(T)$ with the sigmoidal transition $f(T)=1/(1+e^{-(T-T_c)/\\Delta T})$ supplies the nonlinear activation, and the hysteretic insulator-metal transition of VO2 supplies non-volatile memory. From these ingredients the authors construct a T-CNN that performs tri-state spatial classification (+1, 0, –1) and a T-RNN that couples current input $Q(t)$ with stored memory $Q(t-1)$ through gated feedback, with a memory kernel $M(t)=e^{-t/\\tau_s}$ and a thermal storage time $\\tau_s=C_{\\mathrm{eff}}R_{\\mathrm{th}}$ set by material composition and thickness. As a demonstration, a T-RNN trained on 50,000 simulated heat-flux–temperature pairs predicts the filling ratio of a VO2 grating with LSTM MAE of 0.014 and $R^2>0.996$, and GRU MAE of 0.019 and $R^2=0.993$. The authors conclude that this is the first radiative realization of thermal deep learning, where heat, not electrons, performs logic and inference.","pith_inferences":["The learning shown here is offline weight programming, not autonomous in-situ learning; a genuinely self-training thermal network would need a physical mechanism that converts the loss gradient into gate-temperature changes without external control.","If the independent-coupling assumption survives crosstalk tests, the practical niche of radiative thermal networks is likely low-bandwidth, high-robustness inference—space, nuclear, or high-temperature settings—since response is bounded by thermal relaxation times (sub-second for VO2, above a second for GST).","A direct experimental falsifier is to fabricate a small VO2/h-BN array and measure the off-diagonal radiative couplings while tuning one gate; the measured crosstalk matrix decides whether the convolutional weights in Eq. (1) remain independent.","The same T-RNN inverse-prediction approach could be turned around into a general surrogate-design tool for any near-field thermal device whose response curves can be simulated, not just filling-ratio retrieval."],"forward_implications":["A T-CNN can classify spatial heat patterns into tri-state logic (+1, 0, –1) using only localized emissivity modulation and near-field coupling, so feature extraction in convolutional networks has a contactless thermal analog.","A T-RNN can retain the previous thermal state through VO2/GST hysteresis and recurrent radiative feedback, giving sequence recognition and inverse design capability without any electrical readout.","Because the weights are physical gate temperatures, training is a thermal control problem: the network is programmed by imposing a temperature map, and inference is a steady-state or time-dependent heat-transfer problem.","The demonstrated accuracy on inverse prediction (LSTM MAE 0.014, GRU MAE 0.019 on filling ratio) supports the practical use of thermal recurrent models as surrogate solvers for near-field radiative design.","Combining conductive thermal stabilization with near-field radiative neural nodes could yield a complete thermal computing architecture, as the authors suggest for future work."],"supporting_citations":[{"why":"Supplies the fluctuational-electrodynamics basis and the sigmoidal emissivity transition function used to model the VO2 gate's nonlinear response.","marker":"47"},{"why":"Defines the gate-to-drain amplification factor α and the net gate-flux relation that make the radiative transistor and the gain concept work.","marker":"48"},{"why":"Provides the Dyadic Green's function formalism used to compute the near-field radiative heat transfer Q_{S→D} in Eq. (7).","marker":"56"},{"why":"The authors' previously reported near-field radiative thermal diodes, transistors, and multi-terminal logic units on which the NRTNN is built.","marker":"45,46"},{"why":"Experimental VO2 hysteresis data with the ~12 K phase lag that grounds the non-volatile thermal memory of the T-RNN.","marker":"50-52"},{"why":"The LSTM architecture used as one of the T-RNN models for inverse prediction of the filling ratio from heat-flux–temperature curves.","marker":"53"},{"why":"The GRU architecture used as the second T-RNN model, giving comparable inverse-prediction accuracy.","marker":"54"},{"why":"The leave-one-out cross-validation protocol used to assess generalization of the trained thermal recurrent models.","marker":"55"}],"fun_headline_variants":["Heat exchange alone powers neural inference","No circuits: thermal networks learn via photons","Photon tunneling enables thermal deep learning","Thermal networks: inference from temperature fields","Heat-based neural networks without any electrons"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each radiative coupling coefficient in a dense array can be set independently by adjusting gate temperatures, with no thermal crosstalk or phase-change dynamics corrupting the other weights.","fun_headline_variants_meta":{"raw":{"variants":["Heat exchange alone powers neural inference","No circuits: thermal networks learn via photons","Photon tunneling enables thermal deep learning","Thermal networks: inference from temperature fields","Heat-based neural networks without any electrons"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000432,"raw_usage":{"total_tokens":2279,"prompt_tokens":1097,"completion_tokens":1182,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":713,"completion_tokens_details":{"reasoning_tokens":1120}},"tokens_in":713,"tokens_out":1182,"duration_ms":9540,"temperature":1.0,"reasoning_tokens":1120,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:37:21.253074+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fabricate a 2×2 array of VO2/h-BN source-gate-drain units with ~50 nm vacuum gaps, program one gate through its insulator-metal transition, and measure the heat flux between the two unaddressed nodes; if that off-target coupling shifts beyond the network's error tolerance, the independent-weight assumption underlying Eq. (1) fails.","supporting_citations":[{"cited_title":"activation","cited_arxiv_id":null,"evidence_quote":"Defines the gate-to-drain amplification factor α and the net gate-flux relation that make the radiative transistor and the gain concept work."}],"review_version":1}