Pith. sign in

REVIEW 3 major objections 5 minor 6 references

Multi-Output Convolutional Neural Network for Improved Parameter Extraction in Time-Resolved Electrostatic Force Microscopy Data

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read A multi-output CNN recovers bi-exponential kinetics from time-resolved electrostatic force microscopy data.

desk verdict Genuine step beyond their own single-exponential network, with shipped code and a real voltage-pulse ground truth, but the real-data reconstruction metric is partly self-consistent. read the letter →

arxiv 2502.03572 v1 pith:2MTH3R3V submitted 2025-02-05 cond-mat.mtrl-sci cond-mat.dis-nn

classification cond-mat.mtrl-scicond-mat.dis-nn
keywords time-resolvedelectrostaticforcemicroscopyconvolutionalneuralnetworkbi-exponentialkineticsparameterextractioncantileverphysicssignalreconstructionhalideperovskitephysics-informedmachinelearning
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

Time-resolved electrostatic force microscopy (trEFM) records how a vibrating cantilever's frequency shifts when a laser excites a material, but the cantilever's own physics is tangled with the material's transient kinetics, so recovering the kinetics from the signal is hard. This paper claims that a multi-output convolutional neural network, fed the frequency trace together with the cantilever's spring constant, quality factor, and resonance frequency, can extract the two time constants and the mixing coefficient of a bi-exponential surface-potential response. On simulated, labeled voltage-pulse, and unlabeled perovskite data, the network reconstructs the measured signal more accurately than the earlier single-exponential feedforward network (average R2 of 0.97 versus 0.79 on an artificial image, and 0.89 versus 0.71 on real perovskite data). The practical payoff is that trEFM imaging could characterize multi-exponential ion and carrier dynamics in materials such as halide perovskites without assuming a single time constant.

What carries the argument

The central mechanism is the multi-branch convolutional architecture that learns the multi-timescale structure of the trEFM frequency trace: small kernels read the fast initial frequency shift where the kinetic information is concentrated, larger kernels read the slow cantilever-governed relaxation, and a dedicated branch learns the mixing weight $A$. The cantilever parameters $k$, $Q$, and $\omega$ are concatenated into the dense layers so the network can factor out cantilever physics from the kinetics. A custom loss function weights the squared error of $\tau_1$ most heavily, reflecting that $\tau_1$ produces the subtlest signal changes.

What would settle it

Run the trained CNN on trEFM data from a sample whose transient kinetics are independently known to be tri-exponential or stretched-exponential; if the network still reports high reconstruction R2 while its extracted τ1, τ2, and A disagree with the true kinetic moments, then the CNN is fitting the assumed bi-exponential form rather than the actual dynamics.

Watch

Extended reading notes

Core claim

The paper's central claim is that a multi-branched, multi-output CNN trained on simulated trEFM traces and fine-tuned on labeled voltage-pulse experiments can invert the cantilever-obfuscated frequency signal to recover the parameters ($\tau_1$, $\tau_2$, $A$) of an underlying bi-exponential surface potential perturbation $y = A e^{-t/\tau_1} + (1-A) e^{-t/\tau_2}$. The network takes the normalized $\Delta\omega(t)$ trace through three convolutional branches with different kernel sizes to capture fast and slow features, concatenates the latent features with the physical cantilever parameters ($k$, $Q$, $\omega$), and regresses the three parameters through dense output branches. The authors show that this model separates fast and slow time constants with mean errors of 1.01 $\mu$s and 35.87 $\mu$s on fine-tuned experimental data, extracts $A$ with a parity R2 of 0.97, and reconstructs experimental signals with R2 0.89 on unlabeled perovskite images, substantially better than the single-exponential feedforward baseline. They take this as evidence that bi-exponential assumptions and physics-informed inputs make CNN-based parameter extraction practical for trEFM.

Load-bearing premise

The forward model that converts a bi-exponential surface-potential decay into a simulated cantilever frequency trace must be a faithful description of real trEFM experiments, because that model generates all of the training data and is also reused to score how well the extracted parameters reconstruct the measured signal.

Editorial extensions

If this is right

  • trEFM data can be analyzed under a bi-exponential model without empirical cantilever calibration, using only the cantilever parameters already measured in the experiment.
  • The CNN keeps reconstruction R2 above 0.9 at signal-to-noise ratios down to about 7, where the single-exponential network degrades to 0.43-0.8.
  • The extracted parameter maps on perovskite films reproduce known physics: grain boundaries show slower surface potential equilibration than grain interiors.
  • SHAP analysis confirms the network relies on the initial frequency shift, not the cantilever relaxation, for the kinetic parameters, and learns that high-Q cantilevers correspond to faster underlying dynamics.
  • The approach is claimed to be generalizable to other underlying functional forms and other time-resolved scanning probe techniques.

Reading between the lines

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

  • A natural extension would be to relax the bi-exponential assumption to tri-exponential or stretched-exponential kinetics; the same multi-branch architecture should transfer, but the training data would need to come from a forward model that includes those forms.
  • Because the network's inputs include cantilever parameters, it may generalize across cantilevers and tip states without retraining, a testable claim the paper does not directly demonstrate.
  • The reconstruction-R2 metric is computed through the same forward model used to generate training data, so high R2 on experimental data is necessary but not sufficient evidence that the extracted parameters are physically correct; an independent kinetic measurement would close that gap.
  • A deployment risk is that the network will confidently fit a bi-exponential to data that are not bi-exponential; adding a model-mismatch output or an uncertainty estimate would let users know when the assumed form is inadequate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper presents a multi-output convolutional neural network (CNN) for extracting bi-exponential surface-potential dynamics (parameters τ1, τ2, and A) from time-resolved electrostatic force microscopy (trEFM) frequency traces. The network is trained on simulated traces generated with the FFTA forward model, fine-tuned on a small set of experimentally labeled voltage-pulse traces, and then evaluated on simulated test data, held-out voltage-pulse data, an artificial image, and real unlabeled perovskite trEFM data. The authors report accurate parameter extraction on the held-out voltage-pulse tests (R2 = 0.74, 0.79, 0.97 for τ1, τ2, A), improved signal reconstruction R2 compared to a previous single-exponential feedforward network (0.97 vs 0.79 on an artificial image; 0.89 vs 0.71 on real perovskite data), and SHAP-based analyses that confirm physically intuitive feature importance. The core claim is that this network enables more accurate and noise-robust extraction of multi-exponential dynamics from trEFM data than existing single-exponential approaches.

Significance. If the claims hold, this work extends quantitative trEFM analysis from single-exponential to bi-exponential dynamics and provides an open-source, reproducible pipeline. The voltage-pulse experiment supplies an external ground truth that partially validates the parameter extraction, which is a genuine strength. However, the real-data reconstruction metric is computed with the same FFTA forward model used to generate the training data, so the improved reconstruction on unlabeled data may reflect improved fitting to the forward model rather than to true experimental physics. Cross-cantilever generalization is also not directly demonstrated. These gaps currently limit the significance of the practical claims, although the methodological contribution and code availability are valuable.

major comments (3)
  1. [Section 3.5 and Methods 2.3] The reconstruction R2 on real perovskite data (0.89) is computed by simulating Δω(t) from the extracted parameters using the FFTA forward model, which is the same model used to generate all simulated training data. This evaluation is self-consistent: the network is trained on FFTA-simulated traces and the 'reconstructed' signal is simulated with the same FFTA model. Consequently, the comparison against the single-exponential network (0.89 vs 0.71) may indicate which set of parameters better reproduces the forward model's output, rather than which more accurately represents the true experimental signal. Please provide an independent validation that does not rely on the training forward model, for example by comparing reconstructed raw cantilever deflection traces against measured ones, or by using labeled voltage-pulse data collected with a different cantilever and tip state.
  2. [Section 3.3 and Supporting Information Fig. 6] The held-out voltage-pulse test set is obtained by randomly splitting the 500 fine-tuning traces into 350/50/100 for training/validation/testing, but the paper does not report whether the test traces were collected on a different cantilever, tip, or session from the fine-tuning set. If the test traces share the same cantilever transfer function and noise floor as the fine-tuning traces, the reported parity R2 values (0.74, 0.79, 0.97) partly reflect in-distribution memorization rather than generalization to a new cantilever. Please report the cantilever/session identity of the test traces or add a cross-cantilever validation to support the claimed generalizability to new experimental conditions.
  3. [Section 3.2, Eq. (1), and Supporting Information Fig. 2] The training and evaluation are confined to τ1 in [1,10] μs, τ2 in [50,500] μs, and A in [0,1]. The paper does not discuss expected behavior for real data with dynamics outside these ranges, and the perovskite data are not reported to lie within the training domain. Because the network regresses parameters within a bounded, normalized output, extrapolation to values outside the training ranges is unlikely to be reliable. Please state this limitation explicitly and, if possible, report the extracted parameter ranges for the perovskite data to confirm they fall within the training distribution.
minor comments (5)
  1. [Section 3.4] The main text references 'Supporting Information Fig. 9' for SHAP distributions of cantilever parameters, but the supporting information lists Fig. 9 as the noise-evaluation figure and Fig. 10 as the SHAP cantilever-parameter figure; the reference numbers appear to be swapped.
  2. [Section 2.3] The description of the simulation dataset split (7,000/1,500/1,500 out of 10,000 traces) is clear, but it would be helpful to state the analogous split for the 500 voltage-pulse traces in the same paragraph or in Section 2.4 for consistency.
  3. [Supporting Information Fig. 4 caption] The caption gives the FFTA repository as 'https://github.com/rajgiriUW/ffta' while the main text and Data Availability give 'https://github.com/GingerLabUW/FFTA'; please standardize the URL.
  4. [Supporting Information Fig. 13 caption] The caption refers to 'shown in main text and shown in Figure 5' and later to 'Figure 6', but the main text figure is Figure 5; the second reference appears to be a typo.
  5. [Section 3.3] The statement 'we believe this level of accuracy should be sufficient for many if not most imaging applications' is an assertion without a quantitative criterion; consider providing a concrete example of an imaging application and the acceptable error level for that application.

Circularity Check

1 steps flagged · score 4.0 of 10

Parameter extraction is externally anchored by voltage-pulse labels, but the real-data reconstruction R2 is computed with the same FFTA forward model used to generate the CNN's training data, making the reconstruction claim partly self-consistent by construction.

  1. fitted input called prediction [Section 3.5 (Signal Reconstruction on Unlabeled Experimental Data); SI Fig. 4 caption; Methods 2.3]
    "Given these parameters, we simulated Δω(t) traces for each pixel to evaluate how well the bi-exponential parameters reconstructed the experimental data. Supporting Information Fig. 14 shows the map of R2 values comparing the experimental and simulated data, showing an average R2 score of 0.89."

    The CNN is trained on Δω(t) traces generated by the FFTA forward model (Methods 2.3; SI Fig. 4: 'All simulations performed with the FFTA code package'). Section 3.5 then evaluates the extracted parameters by simulating Δω(t) with that same FFTA model and computing R2 against the experimental trace. The network learns an inverse of the FFTA map, and the reconstruction applies that same forward map to the inverse output, forming a closed loop. A high reconstruction R2 therefore primarily indicates that the real traces lie in the FFTA training manifold; it is not an independent check of the extracted τ1, τ2, and A values.

full rationale

The central parameter-extraction claim is not circular: it is anchored by held-out voltage-pulse traces with known τ1, τ2, and A values (Methods 2.4; SI Figs. 6 and 7), providing an external ground truth. The paper also openly acknowledges that collecting sufficient labeled data on multiple cantilevers would take too long (Methods 2.4), which limits cross-cantilever generalization claims but does not make the fine-tuning test circular. The circular component is confined to the reconstruction R2 metric used in Sections 3.3 and 3.5: the CNN is trained on FFTA-simulated Δω(t) traces, and the reconstructed signals are then simulated with the same FFTA model. This closed loop makes the reconstruction R2 a self-consistency measure rather than an independent physical validation, so the 'improved reconstruction' claim is partly forced by the shared forward model. No load-bearing uniqueness theorem or self-citation chain is invoked, and the voltage-pulse parity data provide genuine external support for parameter extraction. Score 4 reflects that the central claim still has independent content while one advertised prediction (real-data reconstruction) is partially circular by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim depends on the bi-exponential model, the FFTA forward simulator, the voltage-pulse equivalence assumption, and several hand-tuned training choices. The network's learned weights are not listed because they are fitted model parameters, but the hyperparameters and training ranges above are the load-bearing hand-chosen inputs that shape what the network can extract.

free parameters (4)
  • Loss-function weights (τ1, τ2, A MSE multipliers) = 2.5, 1.75, 1
    Chosen by sweeping on validation data (SI Note 1); the reported test errors are conditional on these hand-tuned weights.
  • Training ranges for τ1, τ2, A = τ1: 1-10 μs; τ2: 50-500 μs; A: 0-1
    These ranges define the label space and constrain what the CNN can output; real dynamics outside these ranges would be mis-estimated.
  • Cantilever parameter sampling distributions = Distributions in SI Fig 1 (k, Q, ω)
    The simulated training set samples cantilever parameters from these distributions; if real cantilevers fall outside them, the network may extrapolate poorly.
  • Fine-tuning split of voltage-pulse traces = 350 train / 50 validation / 100 test
    The small labeled dataset (500 traces, likely from one cantilever setup) bounds the statistical power of the fine-tuning evaluation.
assumptions (5)
  • domain assumption The underlying surface potential dynamics are described by a bi-exponential function (Eq 1) with A in [0,1] and two distinct time constants.
    Introduced in Section 3.1 and Methods 2.3; the entire output space of the network is built on this model.
  • domain assumption The FFTA forward model accurately maps a surface potential perturbation to the measured cantilever frequency shift Δω(t).
    Inherited from refs 17 and 31; used to generate all simulated training data and to compute reconstruction R2 in Section 3.5.
  • domain assumption Voltage pulses applied to a conductive substrate reproduce the same cantilever response as photoexcitation-induced surface potential changes.
    Methods 2.4 asserts this for the labeled fine-tuning data; differences in force distribution or sample response are not characterized.
  • ad hoc to paper Adding Gaussian noise to the simulated oscillation, then demodulating, captures the dominant experimental noise sources.
    SI Note 3 describes this noise model; thermal, photodiode, and trigger jitter are not separately modeled.
  • ad hoc to paper The custom loss weights (2.5, 1.75, 1) are a workable training objective.
    SI Note 1 states the weights were found by sweeping; no principled derivation is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Output Convolutional Neural Network for Improved Parameter Extraction in Time-Resolved Electrostatic Force Microscopy Data." pith.science (2026). https://pith.science/paper/2MTH3R3V

@misc{pith2026250203572,
  author       = {Pith},
  title        = {Pith review of: Multi-Output Convolutional Neural Network for Improved Parameter Extraction in Time-Resolved Electrostatic Force Microscopy Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2MTH3R3V}},
  note         = {Machine review of arXiv:2502.03572}
}
read the original abstract

Time-resolved scanning probe microscopy methods, like time-resolved electrostatic force microscopy (trEFM), enable imaging of dynamic processes ranging from ion motion in batteries to electronic dynamics in microstructured thin film semiconductors for solar cells. Reconstructing the underlying physical dynamics from these techniques can be challenging due to the interplay of cantilever physics with the actual transient kinetics of interest in the resulting signal. Previously, quantitative trEFM used empirical calibration of the cantilever or feed-forward neural networks trained on simulated data to extract the physical dynamics of interest. Both these approaches are limited by interpreting the underlying signal as a single exponential function, which serves as an approximation but does not adequately reflect many realistic systems. Here, we present a multi-branched, multi-output convolutional neural network (CNN) that uses the trEFM signal in addition to the physical cantilever parameters as input. The trained CNN accurately extracts parameters describing both single-exponential and bi-exponential underlying functions, and more accurately reconstructs real experimental data in the presence of noise. This work demonstrates an application of physics-informed machine learning to complex signal processing tasks, enabling more efficient and accurate analysis of trEFM.

Figures

Figures reproduced from arXiv: 2502.03572 by the authors.

Figure 5
Figure 5. 30 Supporting Information [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

6 extracted references · 4 canonical work pages

  1. [3]

    (14) Shi, Y .; Rojas-Gatjens, E.; Wang, J.; Pothoof, J.; Giridharagopal, R.; Ho, K.; Jiang, F.; Taddei, M.; Yang, Z.; Sanehira, E

    https://doi.org/10.1038/s41560-024-01660-1. (14) Shi, Y .; Rojas-Gatjens, E.; Wang, J.; Pothoof, J.; Giridharagopal, R.; Ho, K.; Jiang, F.; Taddei, M.; Yang, Z.; Sanehira, E. M.; Irwin, M. D.; Silva-Acuñ, C.; Ginger, D. S. (3-Aminopropyl)Trimethoxysilane Surface Passivation Improves Perovskite Solar Cell Performance by Reducing Surface Recombination Veloc...

  2. [4]

    We demonstrate that the trained model extracts parameters which accurately reconstruct labeled and unlabeled experimental data

    Conclusion We present a multi-branched, multi-output CNN trained on simulated data and fine tuned on labeled experimental data that efficiently and accurately extracts parameters that describe the underlying physical dynamics measured by trEFM. We demonstrate that the trained model extracts parameters which accurately reconstruct labeled and unlabeled exp...

  3. [5]

    (30) Fawaz, H

    https://doi.org/10.1038/s41524-019-0148-5. (30) Fawaz, H. I.; Forestier, G.; Weber, J.; Idoumghar, L.; Muller, P.-A. Deep Learning for Time Series Classification: A Review. Data Min Knowl Discov 2019, 33, 917–963. https://doi.org/10.48550/arXiv.1809.04356. 16 (31) Karatay, D. U.; Harrison, J. S.; Glaz, M. S.; Giridharagopal, R.; Ginger, D. S. Fast Time-Re...

  4. [6]

    Channel Normalization in Convolutional Neural Network avoids Vanishing Gradients

    (b) Map of R2 values comparing experimental data with signals simulated with single-exponential feedforward neural network3 parameters, evaluating how well the single-exponential network reconstruct the experimental data. 32 References (1) Dai, Z.; Heckel, R. Channel Normalization in Convolutional Neural Networks Avoids Vanishing Gradients. 2019, 1–13. ht...

  5. [49]

    (26) Yaman, M

    https://doi.org/10.1038/s41524-019-0186-z. (26) Yaman, M. Y .; Kalinin, S. V .; Guye, K. N.; Ginger, D. S.; Ziatdinov, M. Learning and Predicting Photonic Responses of Plasmonic Nanoparticle Assemblies via Dual Variational Autoencoders. Small 2023, 19, 2205893. https://doi.org/10.1002/smll.202205893. (27) Jesse, S.; Kalinin, S. V . Principal Component and...

  6. [2024]

    Sub-diffraction Imaging of Carrier Dynamics in Halide Perovskite Semiconductors: Effects of Passivation, Morphology, and Ion Motion

    https://doi.org/https://arxiv.org/abs/2412.04423. (5) Jiang, F.; Pothoof, J.; Muckel, F.; Giridharagopal, R.; Wang, J.; Ginger, D. S. Scanning Kelvin Probe Microscopy Reveals That Ion Motion Varies with Dimensionality in 2D Halide Perovskites. ACS Energy Lett. 2021, 6, 100–108. https://doi.org/10.1021/acsenergylett.0c02032. (6) Giridharagopal, R.; Cox, P....

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.