{"id":"697134ae-05b5-4da3-a8f1-c38f6f15f62f","arxiv_id":"2412.09638","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Machine learning predicts perovskite solar cell voltage, current, and fill factor from transmission and photoluminescence spectra of the absorber film alone, with moderate R2 values and no released code or external validation.","lead":"This paper trains neural networks on optical measurements of perovskite films (transmission, photoluminescence spectra and lifetimes) to predict the voltage, current and fill factor of completed solar cells, then uses linear regression to identify which optical features matter. If it holds up, it offers a fast, non-destructive screening route for perovskite solar cell manufacturing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Random splits over a dataset that mixes fresh and thermally degraded cells may place the same physical device in both training and test, inflating R2; a device- or batch-grouped split is needed before the generalization claim can be accepted.","rationale":"The reader's weakest assumption correctly identifies that random splits within the same 368-cell campaign may not estimate generalization to future devices, and that non-optical factors such as series resistance can affect FF. My stress-test sharpens this into a specific, load-bearing mechanism: the expanded dataset of 368 samples is composed of 220 initial cells plus 148 thermally degraded cells, and the manuscript does not state whether the degraded cells are distinct devices or remeasurements of the same cells. If they are remeasurements, random splits can leak device identity into the training set, inflating all reported R2 and accuracy figures. This is a more concrete and more damaging failure mode than a generic 'non-optical factors' caveat, because it directly explains why R2 for Jsc jumps from 0.06 to 0.77 when degraded samples are added: the model may be recognizing repeated devices rather than learning from a broader distribution. A second concern, that the optical inputs are measured on completed solar cells rather than constituent films, further undermines the abstract's phrasing, though it is secondary to the leakage issue. The reader's conditional verdict remains appropriate: the paper is a plausible and clearly written application, but its central claim cannot be accepted at face value until the grouped-split validation is performed and the measurement protocol is clarified. I therefore recommend no change to the reader's verdict, but with a stronger and more specific condition for acceptance.","tokens_in":11945,"tokens_out":5960,"duration_ms":67338,"concrete_test":"Ask the authors to rerun the entire expanded-dataset evaluation using a grouped split: assign all spectra and JV records from the same physical device, including any fresh and post-heat measurements of that device, and all cells from the same fabrication batch, to the same fold; train on the remaining devices/batches and evaluate only on held-out devices/batches. If the R2 values for Voc, Jsc, and FF drop substantially (for example, Voc R2 falls below about 0.2) or the 95% accuracy intervals widen markedly, the random-split result is inflated by sample leakage and the film-only generalization claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that optical spectra of the active film predict JV of unseen devices. The supporting evaluation uses random splits of 368 samples, but this set is not necessarily 368 independent devices: it is 220 initial cells plus 148 thermally degraded cells, and the text does not establish that the degraded cells are distinct from the initial ones. If any degraded cell is a remeasurement of a cell already in the 220, or shares a batch with training cells, a random split can place two measurements of the same device/batch in training and test. Because fresh and degraded spectra and JV of the same device are strongly correlated, the network can exploit device identity rather than learn a general optical-to-electrical mapping. The reported R2 values (0.47, 0.77, 0.58) and the '95% within 91/94/89%' accuracy would then overstate generalization to truly new cells. A compounding issue is that the Methods state optical measurements are taken on 'fabricated solar cells' (full stack including Spiro-OMeTAD and Au), not on isolated constituent films, so the film-only framing of the title and abstract is not directly tested by the reported protocol. These concerns do not require new algorithm development; they require an experimental and data-handling validation that the current manuscript does not provide.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript trains neural-network and linear-regression models to predict open-circuit voltage (Voc), short-circuit current (Jsc), and fill factor (FF) of 3D/2D halide perovskite solar cells from three optical spectra of the cell: optical transmission, steady-state photoluminescence, and time-resolved photoluminescence. The models are trained on 220 initial cells and then on an expanded 368-cell dataset that includes 148 thermally degraded cells. On held-out random samples from the expanded dataset, the authors report R2 values of 0.47 (Voc), 0.77 (Jsc), and 0.58 (FF) and state that 95% of predictions fall within 91%, 94%, and 89% of the measured values. A linear-regression analysis over 20 physically motivated spectral descriptors is used to identify transmission-edge steepness and SrPL linewidth as the dominant predictive features, and a set of classifier models (SVM, cross-entropy, neural-network classifier) is used to categorize cells as good, OK, or bad, with claimed classification accuracy above 90% under an aggregate scoring rubric.","tokens_in":12214,"tokens_out":4171,"duration_ms":45875,"significance":"If the reported predictive mapping holds on truly new devices, the approach could serve as a rapid, non-destructive screening tool for perovskite solar cells, and the physical-descriptor analysis is a useful step toward interpretable ML for photovoltaics. The manuscript has concrete strengths: a comparatively large experimental dataset (368 samples), measurement of three complementary optical spectra, comparison of several regression and classification algorithms, and an explicit attempt to connect model weights to material properties. The central claim, however, is currently supported only by random splits within a single fabrication campaign, and the input spectra are measured on complete device stacks rather than on isolated constituent films. These two issues directly affect the generality and the conceptual framing of the result, but they are addressable through grouped cross-validation and a clarified measurement protocol; no circularity in the ML fitting is present.","major_comments":[{"comment":"The expanded 368-sample dataset is composed of 220 initial cells and 148 thermally degraded cells, but the text does not state whether the degraded cells are distinct devices or re-measurements of devices already in the 220-cell set. If a degraded spectrum is a re-measurement of the same physical device, or if all degraded cells come from the same batches as the fresh cells, a random split can place strongly correlated measurements of the same device or batch into both training and test, inflating the reported R2 values. The generalization claim requires a device-grouped or batch-grouped cross-validation, for example training on some batches and testing on held-out batches, and reporting those results.","section":"Neural Network Regression Predicting Results (pp. 7-9) and Methods (pp. 15-16)"},{"comment":"The Methods state that %T, SrPL, and TrPL are measured on 'fabricated solar cells', i.e., on the full device stack including Spiro-OMeTAD and Au, whereas the title and abstract describe predictions from 'optical properties of constituent films'. If the optical data are acquired through the entire stack, the input contains information about transport layers, electrode reflections, and interfaces, so the experiments as reported do not demonstrate prediction from film-level properties alone. The authors should clarify the exact measurement geometry and either provide film-level measurements or re-frame the claim to state that the inputs are optical spectra of the complete device.","section":"Methods, Device Fabrication and Characterization (p. 15)"},{"comment":"The linear-regression weight analysis identifies parameters 2, 3, and 4 (the steepness of the %T band-edge transition) as dominant predictors, but SI-Fig.S23 is said to show no significant dependence of PCE on the %T slope. This inconsistency weakens the physical interpretation that a steeper %T edge is a key driver of device performance. Because the 20 descriptors are mutually correlated, the raw regression weights are not a reliable measure of feature importance. Please provide permutation importance or ablation results for the %T-edge and SrPL-FWHM parameters and reconcile the weight analysis with the direct correlation test in SI-Fig.S23.","section":"Connecting Neural Network Predictions to Physics Properties (pp. 10-11 and SI-Fig.S23)"},{"comment":"The headline metric 'average prediction accuracies for 95% of the predicted Voc, Jsc, and FF values are 91%, 94% and 89%' is not defined in the text; it appears to denote one minus the mean absolute percentage error over the middle 95% of predictions, which is a different quantity from classification accuracy and does not directly convey the moderate R2 values (0.47, 0.77, 0.58) reported in the same sentence. For a regression task, the paper should report standard error metrics such as MAE, RMSE, and prediction intervals for each parameter, and the authors should avoid using the term 'accuracy' without a precise definition.","section":"Abstract and Neural Network Regression Predicting Results (pp. 8-9)"}],"minor_comments":[{"comment":"The phrase 'average prediction accuracies for 95 % of the predicted Voc, Jsc, and FF values are 91%, 94% and 89%, respectively' is grammatically ambiguous; please reword to make clear whether the 91/94/89% figures refer to the fraction of predictions within a stated error band or to a separate accuracy metric.","section":"Abstract"},{"comment":"The number of test predictions reported in the text varies ('100 tested data', '225 ML predicted values', '200 testing samples') without a consistent explanation; please state the exact number of independent test draws used in each figure and whether samples are drawn with replacement.","section":"Fig. 3d and Fig. 2d"},{"comment":"The notation 'C!\"#\"' appears to be a garbled binomial coefficient; it should be written as C(20,10) or a standard binomial coefficient expression.","section":"Connecting Neural Network Predictions to Physics Properties (p. 11)"},{"comment":"The statement 'Selected codes will be available on GitHub later' is not sufficient for a reproducibility-focused manuscript; the code and processed dataset should be deposited at submission time, with raw data available upon reasonable request.","section":"Data availability (p. 18)"},{"comment":"Several SI references (e.g., SI-Fig.S4a, SI-Fig.S10, SI-Fig.S23) are cited in the main text but the SI figure numbering is not fully consistent with the cited labels; please verify all cross-references.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is a data-driven application study; its main value is the relatively large experimental dataset and the systematic comparison of regression and classification models. The two load-bearing issues are the independence of the train/test split and the mismatch between the 'constituent films' framing and the actual measurement on complete device stacks. Both are fixable with additional analyses and a revised narrative. The claim of being 'first to use ML to predict device photovoltaic properties solely from optical properties' should be softened in revision, as prior ML studies on perovskite optical properties exist (ref. 27) and the current protocol does not fully isolate film properties. If the grouped-split results degrade significantly, the authors should consider reporting the method as a within-campaign screening tool rather than a general predictor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nShort version: this is a plausible applied-ML paper with a genuinely new input-output pairing — three optical spectra of the perovskite layer (transmission, steady-state PL, time-resolved PL) mapped to device Voc/Jsc/FF — and a decent-sized dataset (220 then 368 cells). The novelty claim holds as an application, and the paper is clearly written.\n\nWhat's good: the specific mapping does not appear in the cited prior work, so first-use claims are reasonable. The paper includes useful ablations (one/two spectra vs three), a pressure test varying train/test ratio, and a linear-regression weight analysis that identifies transmission-edge steepness and PL linewidth as the dominant physical descriptors. That interpretation is cross-checked against a simple PCE correlation, which is a nice sanity check. The expanded dataset with thermally degraded cells broadens the input distribution and improves Jsc prediction.\n\nSoft spots: First, the title and abstract repeatedly say \"constituent films,\" but the Methods state that optical measurements were taken on \"fabricated solar cells\" — the full device stack including Spiro-OMeTAD and Au. That is a real mismatch. The optical signal is likely dominated by the perovskite, but the paper either needs film-level measurements or more careful claims. Second, the evaluation uses random splits within one campaign. For the expanded dataset, the text does not establish that the 148 degraded cells are distinct devices from the initial 220. If any degraded cell is a remeasurement of a cell already in the 220, a random split can place the same device in both training and test, inflating R2. A device-level or batch-level split is needed before generalizing to new cells. Third, the \"91/94/89% accuracy for 95% of predictions\" metric is non-standard and should be clarified or replaced. Fourth, no code or data is released yet; \"selected codes will be available later\" is not enough for reproducibility. The classifier product-score rubric is ad hoc but secondary.\n\nVerdict: the core concept is sound, and the physical-interpretation section is a genuine plus. The claim of accurate prediction, however, is only partially supported by the reported R2 values (0.47 Voc, 0.77 Jsc, 0.58 FF on the expanded set). This deserves a serious referee, but a conditional one that asks for a grouped split, a precise accuracy metric, and either film-level measurements or revised claims.\n\nRecommendation: send to peer review. It is a solid candidate for a device/applications journal, not a methods breakthrough.","headline":"A useful, clearly written applied-ML paper with a genuinely new input-output pairing, but the 'accurately predicts' claim is weakened by moderate R2, a train/test split that may leak device identity, and a mismatch between 'constituent films' and the actual full-stack measurements.","tokens_in":12817,"tokens_out":3363,"would_cite":false,"duration_ms":34480,"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 the current-voltage behavior of 3D/2D halide perovskite solar cells can be predicted from three optical spectra of the active film alone, and demonstrates a neural network that does so with $R^2$ of 0.47, 0.77, and…","keywords":["perovskite solar cells","machine learning","optical characterization","photoluminescence","transmission spectroscopy","neural network regression","device performance prediction","thermal degradation classification"],"falsifier":"Train the model on one batch of cells and test it on a second batch fabricated with the same recipe but intentionally higher series resistance; if the fill-factor predictions lose accuracy while Voc and Jsc predictions hold, the optical input is not informationally sufficient for device-level performance, and the model is learning batch-specific correlations.","tokens_in":11711,"feed_emoji":"🔆","tokens_out":7174,"duration_ms":63323,"temperature":0.7,"pith_summary":"The paper tries to establish that the photovoltaic performance of a finished solar cell can be read off from optical measurements of its perovskite film alone, before any electrical testing. The authors train a neural network on several hundred 3D/2D (FAMA)Pb(IBr)$_3$/OABr cells, feeding it three spectra per cell: optical transmission, spectrally-resolved photoluminescence, and time-resolved photoluminescence. The network predicts open-circuit voltage, short-circuit current, and fill factor with reported $R^2$ values of 0.47, 0.77, and 0.58 on an expanded 368-sample set, and the same inputs give classifiers that flag degraded cells at over 90% accuracy. If this is right, high-throughput optical screening could replace slow J-V testing when optimizing perovskite recipes and processes.","feed_headline":"Perovskite solar cell performance predicted from film spectra alone","feed_subtitle":"A neural network reads three film spectra and forecasts voltage, current, and fill factor before the device is tested.","key_machinery":"The central machine is a feedforward neural network that takes each spectrum as a 250-point vector, passes the three vectors through separate three-layer networks that compress them to 3×50 and then 3×3, and concatenates the result into a three-output vector for Voc, Jsc, and FF; training minimizes L1 or L2 loss against measured J-V data. Alongside it, a linear-regression model over 20 physically extracted parameters (transmission edge positions, PL peak and width, TRPL decay descriptors) provides interpretable weights, and support-vector-machine, cross-entropy, and neural-network classifiers convert the same features into good/OK/bad categories.","core_discovery":"The central claim is that film-level optical spectra carry enough information to predict device-level J-V parameters, making it possible to screen perovskite formulations before fabricating complete devices. On the expanded dataset of 368 cells (220 fresh plus 148 thermally degraded), the neural network predicts Voc, Jsc, and FF with $R^2$ values of 0.47, 0.77, and 0.58 respectively, with 95% of predicted values within 91%, 94%, and 89% accuracy. A linear regression over 20 physically meaningful spectral parameters reproduces much of this accuracy, and its weights identify the steepness of the transmission edge at the bandgap and the full-width at half-maximum of the photoluminescence peak as the dominant physical descriptors. The paper also reports that simpler inputs degrade performance, and that classifiers built on the same inputs separate good, OK, and bad cells with over 90% accuracy under a product-score rubric.","pith_inferences":["A natural next test the paper does not run is cross-batch transfer: training on one fabrication run and predicting cells from another, where series resistance differs. If FF predictions degrade, the optical inputs are not sufficient by themselves.","The low $R^2$ for Voc (0.47) alongside high percentage accuracy (91% for 95% of predictions) suggests the practical error is small but the metric is compressed by the narrow range of Voc values; both numbers are needed to judge real utility.","Because the same optical inputs already encode thermal degradation through TRPL lifetime shifts, the pipeline could plausibly be extended to predict long-term stability, but the paper stops at immediate J-V parameters.","The classifiers' >90% accuracy under a product-score rubric implies the raw spectra could be used as an industrial pre-sorter for films before they are coated into full devices, a consequence the paper mentions only as device-level screening."],"forward_implications":["A perovskite lab could measure three spectra per film and get immediate estimates of Voc, Jsc, and FF without building the full solar cell stack.","The same optical inputs can flag degraded cells with over 90% classification accuracy, supporting in-line quality control.","Because the linear-regression weights emphasize transmission-edge steepness and PL linewidth, process teams can target more crystalline films without waiting for device J-V tests.","Adding degraded cells to the training set lifted Jsc $R^2$ from 0.06 to 0.77, showing that broadening the performance distribution improves learning more than adding similar samples."],"supporting_citations":[{"why":"Supplies the chemical-bath-deposition SnO2 electron-transport layer method used to fabricate every device in the dataset.","marker":"[3]"},{"why":"Establishes the improved carrier-management 3D/2D device architecture that defines the solar cell stack being predicted.","marker":"[13]"},{"why":"The prior demonstration that ML predicts halide-perovskite optical behavior with >90% accuracy, the step this work extends from spectra to device J-V parameters.","marker":"[27]"},{"why":"Provides the three-step experimental workflow (sample preparation, algorithm development, model evaluation) that structures the study.","marker":"[31]"},{"why":"Sample-size guidance the paper uses to justify matching the number of spectral data points to the number of measured cells.","marker":"[41-43]"}],"fun_headline_variants":["AI predicts perovskite solar performance from film spectra","Neural net forecasts solar cell output from raw spectra","Film spectra alone enable ML prediction of perovskite cells","Spectral fingerprints predict perovskite device output","Machine learning reads films, predicts solar cell metrics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the three film-level optical spectra contain enough information about the finished device's electrical behavior, particularly fill factor, which the paper itself notes is also affected by series resistance and fabrication artifacts that are not optical.","fun_headline_variants_meta":{"raw":{"variants":["AI predicts perovskite solar performance from film spectra","Neural net forecasts solar cell output from raw spectra","Film spectra alone enable ML prediction of perovskite cells","Spectral fingerprints predict perovskite device output","Machine learning reads films, predicts solar cell metrics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1410,"prompt_tokens":1022,"completion_tokens":388,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":329}},"tokens_in":638,"tokens_out":388,"duration_ms":4671,"temperature":1.0,"reasoning_tokens":329,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:50:28.026823+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the model on one batch of cells and test it on a second batch fabricated with the same recipe but intentionally higher series resistance; if the fill-factor predictions lose accuracy while Voc and Jsc predictions hold, the optical input is not informationally sufficient for device-level performance, and the model is learning batch-specific correlations.","supporting_citations":[],"review_version":1}