Pith. sign in

REVIEW 4 major objections 5 minor 3 references

Predicting Organic-Inorganic Halide Perovskite Photovoltaic Performance from Optical Properties of Constituent Films through Machine Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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…

desk verdict 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. read the letter →

arxiv 2412.09638 v1 pith:HXBA4BT6 submitted 2024-12-06 cond-mat.mtrl-sci cs.LGphysics.comp-ph

Ruiqi Zhang , Brandon Motes , Shaun Tan , Yongli Lu , Meng-Chen Shih , Yilun Hao , Karen Yang , Shreyas Srinivasan
show 2 more authors
Moungi G. Bawendi Vladimir Bulovic
This is my paper · ORCID
classification cond-mat.mtrl-scics.LGphysics.comp-ph
keywords perovskitesolarcellsmachinelearningopticalcharacterizationphotoluminescencetransmissionspectroscopyneuralnetworkregressiondeviceperformancepredictionthermaldegradationclassification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

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.

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 (4)
  1. [Neural Network Regression Predicting Results (pp. 7-9) and Methods (pp. 15-16)] 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.
  2. [Methods, Device Fabrication and Characterization (p. 15)] 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.
  3. [Connecting Neural Network Predictions to Physics Properties (pp. 10-11 and SI-Fig.S23)] 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.
  4. [Abstract and Neural Network Regression Predicting Results (pp. 8-9)] 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.
minor comments (5)
  1. [Abstract] 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.
  2. [Fig. 3d and Fig. 2d] 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.
  3. [Connecting Neural Network Predictions to Physics Properties (p. 11)] The notation 'C!"#"' appears to be a garbled binomial coefficient; it should be written as C(20,10) or a standard binomial coefficient expression.
  4. [Data availability (p. 18)] 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.
  5. [Throughout] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: ML predictions are held-out empirical fits over independently measured optical inputs and J-V targets.

full rationale

The derivation chain is an empirical supervised-learning pipeline, not an analytic derivation whose output is assumed in its inputs. The targets (Voc, Jsc, FF) are independently measured J-V characteristics obtained with a Keithley 2400 source meter under AM1.5 illumination, while the inputs are separately measured optical spectra (%T, SrPL, TrPL); neither the inputs nor the outputs are defined in terms of one another. Reported accuracies are computed on held-out samples after training, including repeated random-selection procedures and a pressure test with increasing test-to-train ratios, so the quoted R2 values are not the training fit itself. The linear-regression weight analysis is interpretive post-hoc analysis, not used to define the targets, and the extracted physical parameters are fitted to predict the same independently measured JV values; the claim that SrPL FWHM correlates with PCE is checked against the data rather than assumed. Self-citations support fabrication protocols and prior device results, but no load-bearing uniqueness claim or excluded alternative rests on a self-citation. The paper's own caveat that FF depends on series resistance and fabrication artifacts, and the separate concern about random splits over fresh and degraded cells, are generalization and data-leakage validity issues rather than circularity. No step can be exhibited in which a fitted parameter is renamed as a prediction or in which Eq. X equals Eq. Y by construction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the sufficiency of three film-level optical measurements, on the representativeness of random splits, and on several hand-set preprocessing and threshold choices. There are no physically invented entities; the main uncharged assumptions are distributional and informational.

free parameters (4)
  • Neural network weights (all layers) = Not enumerated in manuscript; trained via L1/L2 minimization for 15,000-20,000 epochs
    The core regressor is a black-box fit; no architecture parameter count, regularization, or weight values are reported.
  • Linear regression coefficients for 20 extracted physical descriptors = Normalized weights shown in Fig. 5b; not tabulated in text
    Fitted on the 368-sample dataset to predict Voc/Jsc/FF and to rank descriptor importance.
  • Training hyperparameters (learning rate, epoch count, 250-point spectral downsampling) = Learning rate 0.0001; 15,000 epochs (220-set) and 20,000 epochs (368-set); 250 points per spectrum via 3-neighbor…
    Chosen by observing convergence and error, not by independent validation; preprocessing choice sets input dimensionality.
  • Classifier category thresholds and product-score rubric = Good: Voc>900 mV, Jsc>20 mA/cm2, FF>60%; bad: Voc<700 mV, Jsc<17 mA/cm2, FF<45%; score = product of 0/1/2 categories
    Hand-set boundaries define the classification labels and the aggregated 'cell performance score' used to claim >90% accuracy.
assumptions (4)
  • domain assumption The three measured optical film spectra are informationally sufficient to determine device Voc, Jsc, and FF
    This is the load-bearing premise of the entire study; the paper itself notes in the NN regression results section that FF also depends on series resistance and fabrication artifacts, so the premise is only partially satisfied even by the authors' own account.
  • domain assumption Random splits from the same 368-cell fabrication campaign estimate performance on new devices
    All test sets are random subsets of the same population; no separate batch, run, or laboratory validation is performed, so distribution shift is not measured.
  • domain assumption The measured J-V values are accurate ground truth without significant instrument drift or systematic error
    Methods describe a Keithley source meter and AM1.5 simulator with a shadow mask, but no uncertainty or reproducibility statistics for the J-V measurements are reported.
  • ad hoc to paper The aggregated product-score rubric is a meaningful classification target
    The 'cell performance score = product of per-parameter category scores' rule is introduced by the authors and used in Fig. 6e to report >90% classification accuracy; it can hide errors in individual parameters and is not a standard metric.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Predicting Organic-Inorganic Halide Perovskite Photovoltaic Performance from Optical Properties of Constituent Films through Machine Learning." pith.science (2026). https://pith.science/paper/HXBA4BT6

@misc{pith2026241209638,
  author       = {Pith},
  title        = {Pith review of: Predicting Organic-Inorganic Halide Perovskite Photovoltaic Performance from Optical Properties of Constituent Films through Machine Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HXBA4BT6}},
  note         = {Machine review of arXiv:2412.09638}
}
read the original abstract

We demonstrate a machine learning (ML) approach that accurately predicts the current-voltage behavior of 3D/2D-structured (FAMA)Pb(IBr)3/OABr hybrid organic-inorganic halide perovskite (HOIP) solar cells under AM1.5 illumination. Our neural network algorithm is trained on measured responses from several hundred HOIP solar cells, using three simple optical measurements of constituent HOIP films as input: optical transmission spectrum, spectrally-resolved photoluminescence, and time-resolved photoluminescence, from which we predict the open-circuit voltage (Voc), short-circuit current (Jsc), and fill factors (FF) values of solar cells that contain the HOIP active layers. Determined average prediction accuracies for 95 % of the predicted Voc, Jsc, and FF values are 91%, 94% and 89%, respectively, with R2 coefficients of determination of 0.47, 0.77, and 0.58, respectively. Quantifying the connection between ML predictions and physical parameters extracted from the measured HOIP films optical properties, allows us to identify the most significant parameters influencing the prediction results. With separate ML-classifying algorithms, we identify degraded solar cells using the same optical input data, achieving over 90% classification accuracy through support vector machine, cross entropy loss, and artificial neural network algorithms. To our knowledge, the demonstrated regression and classification work is the first to use ML to predict device photovoltaic properties solely from the optical properties of constituent materials.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages

  1. [1]

    pressure test

    Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, 02139, US 2. Research Laboratory of Electronics (RLE), Massachusetts Institute of Technology, Cambridge, MA, 02139, US 3. Department of Chemistry, Massachusetts Institute of Technology, Cambridge, MA, 02139, US 4. Department of Aeronautics and ...

    work page

  2. [20]

    Nature Communications, 2021

    Zhao, Y., et al., Discovery of temperature-induced stability reversal in perovskites using high-throughput robotic learning. Nature Communications, 2021. 12(1): p. 2191. 21. Howard, J.M., et al., Quantitative Predictions of Moisture-Driven Photoemission Dynamics in Metal Halide Perovskites via Machine Learning. The Journal of Physical Chemistry Letters, 2...

    work page 2021

  3. [38]

    ACS Applied Materials & Interfaces, 2019

    Neukom, M.T., et al., Consistent Device Simulation Model Describing Perovskite Solar Cells in Steady-State, Transient, and Frequency Domain. ACS Applied Materials & Interfaces, 2019. 11(26): p. 23320-23328. 39. Ren, X., et al., Exploring the Way To Approach the Efficiency Limit of Perovskite Solar Cells by Drift-Diffusion Model. ACS Photonics, 2017. 4(4):...

    work page 2019

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.