{"id":"5a28fbf9-3aeb-4151-9bb9-cc61ea9e606b","arxiv_id":"2607.20810","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ErMnO3's relaxing photocurrent works as a physical reservoir, lifting recognition of the previous light pulse from ~33% (random) to ~93%.","lead":"This paper shows that the light-induced electrical current in the ferroelectric semiconductor ErMnO3 can act as a physical reservoir, letting a simple linear readout recognize the previous light pulse with ~93% accuracy (vs. 33% random). It offers a tunable, light-driven material platform for energy-efficient temporal information processing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"93% past-pulse accuracy may not require the 10-node reservoir: a single current sample at pulse onset could already encode u(k−1); the missing n=1 control leaves the central reservoir claim under-supported.","rationale":"The reader's verdict is CONDITIONAL, and I do not propose moving away from that. However, the reader's weakest_assumption — that the square-root baseline drift may carry or destroy pulse-history information — is not the most load-bearing issue. For a random pulse sequence, a slow monotonic drift is essentially uncorrelated with the immediately preceding label u(k−1), so it cannot alone explain a jump from 33% to 93%. The more direct threat to the central claim is the absence of a single-node control. The experimental design compares the full 10-node reservoir state only against raw input intensities, not against a minimal physical readout. Given the pulse duration (5 s) relative to the relaxation times (20 ms–several s), the photocurrent at the start of pulse k is already a memory-bearing quantity encoding u(k−1). If that single sample is sufficient, the 'reservoir transformation' is not doing the work attributed to it, and the evidence for high-dimensional projection — a key component of the reservoir-computing claim — collapses to a demonstration of one-step memory in a scalar current. Appendix C omitting n=1 makes this possibility impossible to evaluate from the paper as written. I therefore flag this as the load-bearing concern, while noting that it does not necessarily falsify the broader idea that ErMnO3 photocurrents could serve as a reservoir; it does mean the current evidence is conditional on this missing control. The reader's baseline concern is real but less decisive; I disagree that it is the single weakest assumption.","tokens_in":12479,"tokens_out":15074,"duration_ms":183247,"concrete_test":"Re-run the Appendix B readout pipeline for the past-pulse task using only the first reservoir node i=0 — i.e., x(k) = [1, r(tk)]^T with n=1 — under the identical 1000 random-split protocol used for Fig. 4, and also report the n=1 point in the Fig. 6 saturation curve. If the n=1 accuracy is ≥90%, the claim that the 10-node time-multiplexed reservoir transformation is the source of the memory performance is not supported by the presented task; if n=1 accuracy is near chance (~33%), the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the time-multiplexed reservoir state — 10 virtual nodes sampled across the 5 s pulse window — is what enables past-pulse recognition, lifting accuracy from ~33% to ~93%. But the paper's only 'without reservoir' baseline (Fig. 4, dashed yellow) is trained on the raw input intensities u(k), not on a single photocurrent readout. Since each pulse lasts 5 s and the reported relaxation times range from 20 ms to several seconds, the photocurrent at the onset of pulse k (node i=0) is essentially the tail/steady-state response to the immediately preceding pulse u(k−1). Thus a one-dimensional state [1, r(tk)] may already separate the three past intensity levels with high accuracy. This possibility is not tested: Appendix C's accuracy-vs-n curve starts at n=2, and the text only says the past-pulse accuracy 'rises from n=2 and saturates by about n=8.' No n=1 point is reported. If n=1 accuracy is already near 93%, then the time-evolving, high-dimensional reservoir transformation is not responsible for the demonstrated improvement; the experiment would show short-term memory in the raw photocurrent, but not the reservoir-computing projection claimed in the abstract and introduction. The reader's baseline-drift concern is secondary in comparison: a smooth sqrt drift is nearly independent of u(k−1) in a random pulse sequence, so it is unlikely to fabricate 93% from chance; at worst it could remove some genuine memory and make the reported accuracy a lower bound. The missing single-node control directly targets what the 93% is evidence for.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports an experimental demonstration of photo-induced currents in the ferroelectric semiconductor ErMnO3 as a physical reservoir for a temporal classification task. A random sequence of light pulses with three intensity levels is applied, and the resulting photocurrent is sampled at 10 virtual nodes within each 5 s pulse window. A linear readout, trained by least squares, classifies the present and immediately preceding pulse intensities. The authors report ~99% accuracy for the present pulse and ~93% for the past pulse, compared with ~33% for a readout trained on the raw input intensities. They also present a circuit-model simulation that reaches ~100% accuracy, and they show that the photocurrent relaxation timescale can be tuned via contact type and annealing.","tokens_in":12869,"tokens_out":7895,"duration_ms":83475,"significance":"If the central claim holds, the work is a useful experimental demonstration of short-term memory in a light-driven ferroelectric-semiconductor device, with strengths including an open data/code repository, 1000 random train/test splits, and a fair comparison to a trivial no-memory baseline. However, the specific attribution of the improvement to time-multiplexed reservoir computing (the high-dimensional projection) is not fully established because the natural single-node control is missing. The empirical baseline-drift model also rests on an unvalidated assumption. The paper is of interest to the physical-reservoir-computing community, but the experimental evidence needs additional controls to support the reservoir-computing interpretation.","major_comments":[{"comment":"The 'without reservoir' baseline in Fig. 4 is trained on the raw input intensities u(k), which trivially cannot contain information about u(k−1) in a random sequence. The relevant control is a single photocurrent sample (n=1), such as r(t_k) at pulse onset. Figure 2(c,d) shows that the current at the onset of the present pulse already differs substantially for different past pulses, so n=1 may already achieve high accuracy. The accuracy-versus-n curve in Fig. 6 starts at n=2 and the text does not report n=1. Please provide the n=1 accuracy (and ideally a baseline using a single measured current readout) to show that the time-multiplexed, high-dimensional reservoir state is actually responsible for the ~93% accuracy. Without this, the conclusion in Section II that 'the reservoir transformation, not the raw input, supplies the information about u(k−1)' is under-supported.","section":"Section II, Fig. 4, Appendix C"},{"comment":"The baseline drift is modeled as a fitted square-root function I_bg(t)=a+b√(t−t0) and is assumed to be 'uninformative about the immediate pulse history.' This is an empirical assumption. If the drift contains history-dependent components (e.g., persistent photocurrent accumulation), the background subtraction could remove part of the memory the reservoir is supposed to exploit, or impose an artificial trend. Although the random pulse sequence makes it unlikely that the drift alone fabricates 93% accuracy, a robustness check (e.g., repeating the pipeline without background subtraction or with an alternative detrending) would support the assumption and strengthen the central claim. Please add such an analysis or explicitly discuss the sensitivity.","section":"Appendix A, Eq. (A1)"},{"comment":"The circuit-model simulation is fitted to a single experimental transient (Fig. 7(c)) and then applied to the pulse sequence to obtain ~100% past-pulse accuracy. Since the model parameters are derived from the same material's response, this simulation is not independent evidence that the photocurrent waveform 'already encodes enough' for perfect recognition. The statement in Section II attributing the experimental ~7% shortfall to 'measurement non-idealities' is therefore partly speculative. Please soften this claim or provide a more controlled simulation study (e.g., varying the memory timescale and showing the accuracy degrades when τ1, τ2 are shortened).","section":"Section II, Appendix D"}],"minor_comments":[{"comment":"The equation is written as 'I_bg(t) = a+b√t−t0', which is ambiguous; the square root should clearly cover (t−t0), i.e., a+b√(t−t0).","section":"Appendix A, Eq. (A1)"},{"comment":"The accuracy-versus-n curve starts at n=2. Please state explicitly why n=1 is not included, and if possible add the n=1 point (even as an open symbol) to the figure.","section":"Appendix C, Fig. 6"},{"comment":"The comparison of current values at t=25–30 s and t=150–155 s is described as showing dependence on preceding light pulses, but this example chiefly illustrates the slow baseline drift. Rephrase to avoid confusing drift with memory from u(k−1).","section":"Section II, paragraph after Fig. 2(a)"},{"comment":"The phrase 'developed by intention pictorially' is unclear and should be rewritten. Also, the choice R4 = R3 A2/A1 is not explained; please provide the rationale or a reference.","section":"Appendix D"},{"comment":"When using scikit-learn's LinearRegression with a constant column already in X, the default fit_intercept=True makes the intercept redundant. Please specify fit_intercept=False (or equivalently mention that the constant column supplies the intercept).","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid experimental report with open data and a careful training/evaluation procedure. The main issue is the missing n=1 control, which is essential to support the reservoir-computing interpretation. If the authors add this control and adjust their claims accordingly, the paper could become acceptable. The baseline-drift concern is secondary but should be addressed with a robustness check. I recommend major revision rather than rejection because the experimental data appear sound and the missing control is within the scope of the existing dataset."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper does something real and useful: it shows that the photocurrent response of ErMnO3 carries memory of the previous light pulse and that this memory is tunable via contact engineering. The measurement methodology is careful — 1000 random train/test splits, a fair 33% baseline, clean figures. The ~93% past-pulse accuracy is a legitimate experimental result; the material does have short-term memory.\n\nThe soft spot is that the paper attributes that accuracy to the reservoir transformation — the 10 time-multiplexed virtual nodes — but never runs the control that would justify that attribution. The baseline it compares against is the raw input intensities u(k), which trivially contain no information about u(k−1). The right baseline is a single photocurrent sample, r(t_k), at the onset of the pulse. Given the reported relaxation times (20 ms to several seconds) and the 5 s pulse window, r(t_k) is essentially the tail of the previous pulse's response, so a one-dimensional state may already separate the three intensity classes. The accuracy-versus-n curve in Appendix C starts at n=2; n=1 is missing. If n=1 already gives ~93%, then the time-multiplexed reservoir is not doing the work — the memory is in the raw current, and the 'reservoir computing' framing is an overstatement. If n=1 were poor and the accuracy jumped at n=2 or 3, the claim would hold. Right now we don't know.\n\nThe baseline-drift subtraction (Appendix A) is a lesser concern: a smooth square-root drift is nearly orthogonal to a random pulse sequence, so it's unlikely to fabricate 93%; at worst it may remove some genuine long-term memory. The equivalent-circuit simulation (Appendix D) is fitted to the same transient it is then used to explain, which is circular but only used to attribute the 7% gap to noise, so it doesn't affect the central result. Data and code are stated to be on Zenodo but no link or hash appears in the text; minor workflow issue.\n\nBottom line: this is an incremental but sound experimental contribution, and the authors are honest with their methods. But the central claim needs the n=1 control before the reservoir-computing interpretation is justified. If I were the editor, I'd send it to review, and the referee should ask for that single experiment.","headline":"Solid experimental demonstration of photocurrent memory in ErMnO3, but the missing n=1 control leaves the reservoir-computing claim overreaching.","tokens_in":13357,"tokens_out":2951,"would_cite":false,"duration_ms":32417,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["72.40.+w"],"model":"deepseek-v4-flash","headline":"The time-evolving photocurrent of the ferroelectric semiconductor ErMnO3 is a physical reservoir whose fading memory lets a simple linear readout identify the preceding light pulse with ~93% accuracy.","keywords":["reservoir computing","physical reservoir","photo-induced current","short-term memory","ferroelectric semiconductor","ErMnO3","hexagonal manganite","virtual nodes"],"falsifier":"A decisive test: train the readout on the raw photocurrent without square-root background subtraction, and separately on the baseline-corrected signal, and check that past-pulse accuracy stays near 93% under both treatments. A complementary control is to shuffle the past-pulse labels against the measured traces and confirm accuracy collapses to ~33%; and to fit the baseline on a reversed or reshuffled pulse sequence — if the fitted a+b√(t−t0) changes with pulse order, the drift is not history-independent and the subtraction is not neutral.","tokens_in":12352,"feed_emoji":"💡","tokens_out":9201,"duration_ms":71190,"temperature":0.7,"pith_summary":"This paper tries to establish that the photo-induced current of the ferroelectric semiconductor ErMnO3 is, on its own, enough to perform the memory-dependent step of reservoir computing: recognizing which input pulse came immediately before the current one. Under white-light illumination the photocurrent rises nonlinearly and, after the light is turned off, relaxes over timescales from tens of milliseconds to seconds; sampling that time trace at ten points turns one slow physical signal into a high-dimensional reservoir state. Because the relaxation outlasts the inter-pulse interval, the state carries a fading trace of the previous pulse — directly evidenced by a paired-pulse-facilitation ratio above 100%. A linear readout trained in a single least-squares step then identifies the present pulse at ~99% accuracy and the immediately preceding pulse at ~93%, whereas the same readout on the raw input intensities scores only ~33%, the chance level. The authors further show the relaxation timescale is tunable through contact engineering and reducing-atmosphere annealing, which is what would let such a reservoir be matched to the timescale of a given task.","feed_headline":"Photocurrents lift past-pulse recall from 33% to 93%","feed_subtitle":"A ferroelectric crystal's slow photocurrent acts as a reservoir that remembers the pulse before the current one.","key_machinery":"The key mechanism is time-multiplexed virtual-node sampling applied to a single physical variable. The photocurrent response to each light pulse is sampled at n=10 points spaced by a time interval τ across the 5 s response window, forming the reservoir state vector x(k) = [1, r(tk), r(tk+τ), …, r(tk+(n−1)τ)]^T, the leading 1 acting as readout bias. Two least-squares-trained weight vectors map this same state onto predictions of the present and previous pulse intensities, with the continuous outputs thresholded at the midpoints between the three intensity levels. The physical substrate — nonlinear current rise plus slow trap-governed relaxation — supplies the nonlinearity and the fading memor","core_discovery":"The central discovery is that the dynamic photocurrent of ErMnO3 is not a nuisance signal to be read at steady state but a usable physical reservoir. The shape of the current during a 5 s light pulse depends on both the present intensity and the intensity of the preceding pulse: averaging over all two-pulse configurations yields three distinct current profiles for each present-pulse level, one per possible past pulse. Sampling each response window at ten uniformly spaced times produces the reservoir state that feeds a linear readout. Trained by ordinary least squares, the readout labels the present pulse with ~99% accuracy and the immediately preceding pulse with ~93% accuracy, where the sam","pith_inferences":["Editorial inference: the same protocol could be run as a quantitative short-term-memory benchmark (for example, memory capacity of a linear readout on the reservoir state), which would allow ErMnO3 to be compared directly, on a task-independent scale, with memristive, spintronic, and photonic reservoirs.","Editorial inference: the task only probes one step of memory (the immediately preceding pulse). Since the relaxation is double-exponential with decay constants near 66 ms and 468 ms, the state likely carries information about pulses further back; testing recognition of u(k−2) and u(k−3) would measure the effective memory depth and could clarify whether the experimental shortfall is a memory-depth ","Editorial inference: the background-subtraction step deserves a stress test. A control that trains the readout on the raw photocurrent without the square-root baseline fit, or fits the drift with a history-dependent model and checks whether past-pulse accuracy moves, would show whether the 93% figure is robust to how the slow drift is treated.","Editorial inference: a single-node control — training the same readout on the photocurrent sampled at one optimized time instead of ten — would isolate whether the benefit comes from the temporal trajectory shape or merely from the current level carrying the previous pulse's tail."],"forward_implications":["A single ferroelectric semiconductor device, with no recurrent network in software, can carry the memory-dependent part of a temporal classification task: a linear readout on sampled photocurrent lifts past-pulse recognition from chance level (33%) to about 93%.","The reservoir's memory timescale is an engineering parameter — Schottky contacts give the fastest response, as-grown Ohmic contacts an intermediate one, and annealing in a reducing atmosphere prolongs relaxation to seconds — so the same material can be matched to tasks with different characteristic timescales.","Because the present and past predictions are read from the same reservoir state through different trained weights, switching to a new task requires retraining only the linear readout, not re-running the physical reservoir.","The noiseless circuit model reaches ~100% past-pulse accuracy through the same training pipeline, which the paper treats as evidence that the photocurrent waveform encodes full information and that the experimental gap is noise and baseline drift.","Several relaxation timescales could in principle be combined in one device to raise reservoir dimensionality, and the ferroelectric domain and domain-wall structure offers further, locally distinct responses for downscaling."],"fun_headline_variants":["Photocurrents turn a crystal into a temporal memory bank","Light pulses train a ferroelectric crystal to recall the past","93% past-pulse recall from photocurrents in ErMnO3","Ferroelectric crystal remembers light history via photocurrents","Photocurrent-based reservoir computing lifts past recall to 93%"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the slow, monotonic baseline drift of the photocurrent over the 33-minute run carries no information about the pulse history and can be cleanly removed by fitting a square-root function; if that drift is actually a long-term accumulation of the photo-response, the subtraction could either erase part of the memory the reservoir is supposed to exploit or imprint an artificial trend that inflates past-pulse accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Photocurrents turn a crystal into a temporal memory bank","Light pulses train a ferroelectric crystal to recall the past","93% past-pulse recall from photocurrents in ErMnO3","Ferroelectric crystal remembers light history via photocurrents","Photocurrent-based reservoir computing lifts past recall to 93%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000163,"raw_usage":{"total_tokens":1059,"prompt_tokens":701,"completion_tokens":358,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":282}},"tokens_in":445,"tokens_out":358,"duration_ms":3508,"temperature":1.0,"reasoning_tokens":282,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T09:17:04.026739+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test: train the readout on the raw photocurrent without square-root background subtraction, and separately on the baseline-corrected signal, and check that past-pulse accuracy stays near 93% under both treatments. A complementary control is to shuffle the past-pulse labels against the measured traces and confirm accuracy collapses to ~33%; and to fit the baseline on a reversed or reshuffled pulse sequence — if the fitted a+b√(t−t0) changes with pulse order, the drift is not history-independent and the subtraction is not neutral.","supporting_citations":[],"review_version":1}