{"id":"c7d06075-7ec4-4abe-b29e-c55cb788f2c0","arxiv_id":"2411.17755","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Random forest models trained on acoustic emission features can reproduce the jerky force-time curves of Zn micropillar compression, including avalanche timing and size, with R² around 0.6 to 0.65.","lead":"This paper shows that a machine learning model can predict the force response of compressed zinc micropillars from acoustic emission recordings alone. If reliable, the approach could let engineers monitor internal deformation events in bulk materials using sound signals.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim of predicting 'individual deformation events' is not supported by the evaluation: the target is the total force change in fixed 300 ms windows, and the paper's own waiting-time statistics show multiple avalanches can occur within one window.","rationale":"The reader's weakest assumption concerned transfer-function similarity across pillar sizes. That concern is partly mitigated by the Methods statement that all pillars were milled into the same sample and measured without detaching the AE transducer, and by Supplementary Fig. S5d showing similar power spectra across sizes. The more load-bearing issue is internal: the headline claim about individual event magnitudes is tested only through window-aggregated force increments, and the paper's own waiting-time statistics indicate that multiple events can fall in one window. This is a concrete, checkable mismatch between claim and evaluation. The reader did identify the abstract's overstatement and the coarse-scale use of pillar diameter, which are related, but not the event-level aggregation problem. Therefore the verdict should remain CONDITIONAL: the core reconstruction approach may be valid, but the central claim should be reworded or supplemented with event-level metrics before acceptance.","tokens_in":19217,"tokens_out":5667,"duration_ms":61256,"concrete_test":"Perform event-level evaluation on the held-out force curves. Segment individual force drops using a local-minimum search with a threshold on drop size, then count how many 300 ms windows contain more than one drop. For each window where the model predicts a significant negative ΔF, mark a predicted event at the window center and match it to actual drops within ±Δt. Report precision, recall, and median absolute error between predicted and actual drop magnitudes. If a large fraction of windows contain multiple drops, or if event-level precision/recall is substantially worse than the window-level R², the abstract's 'individual deformation events' claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims the method 'can also predict the magnitude of individual deformation events', but the fine-scale target is the force increment ΔF over a fixed Δt = 300 ms window, not an individual avalanche. The choice of Δt is justified only by drop durations (Suppl. Fig. S1), with no check that at most one force drop occurs per window. Suppl. Fig. S2 shows waiting times down to ~10 ms, far below 300 ms, so multiple avalanches can merge into a single window-level ΔF. Thus the model learns a window-averaged or summed quantity, and the reported R² on all windows, most of which have near-zero increments, can be high even if individual events are missed or mis-sized. Furthermore, 'temporal location' is only evaluated at window resolution, with no precision/recall or event-level localization reported. This is an internal claim/evaluation mismatch, not merely a consensus disagreement, and it directly affects the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a machine-learning framework that reconstructs the force-time response of Zn micropillar compression tests from simultaneously recorded acoustic emission (AE) signals. The AE data in fixed time windows are reduced to statistical moments (frequency-independent features) or to moments of spectrogram slices (frequency-dependent features), and a random forest regressor is trained to predict either the force increment in fine windows (Δt = 300 ms) or the force value in coarse windows (ΔT = 50 s). The two predictions are then combined to produce a full force-time curve. The authors evaluate the method on held-out experiments, report R² scores around 0.60–0.65 for fine-scale prediction for 8-μm pillars, analyse feature importance, and report transferability across pillar diameters of 8, 16, and 32 μm. The central claim is that the approach can predict the temporal location and magnitude of individual deformation avalanches from AE data alone.","tokens_in":19392,"tokens_out":5155,"duration_ms":54345,"significance":"If validated, the approach would be a useful step toward quantitative AE-based monitoring of microscale plasticity, where AE is currently used mostly phenomenologically. The paper's strengths include the use of a leave-one-experiment-out evaluation, the use of publicly available experimental data from Ispánovity et al. (2022), and the comparison of frequency-independent versus frequency-dependent descriptors. The claim is falsifiable and the experimental setup is well matched to the question. However, the significance is substantially tempered by the moderate R² values, the small number of experiments (eight), and the mismatch between the abstract's 'individual deformation events' language and the actual window-averaged regression target. The transferability result, while suggestive, rests on a very small number of held-out tests and is presented more strongly than the evidence supports.","major_comments":[{"comment":"The claim that the method predicts 'individual deformation events' is not supported by the evaluation. The fine-scale target is the force increment ΔF over a fixed 300 ms window, not the size of an individual avalanche. Suppl. Fig. S2 shows waiting times between force drops down to about 10 ms, so multiple drops can and do occur within a single 300 ms window. A window-level sum or average is therefore not the same as an individual-event magnitude. Moreover, R² is computed over all windows, most of which have near-zero increments, so a model that predicts near-zero increments in quiet windows can achieve a moderately high R² even while missing or mis-sizing rare individual events. 'Temporal location' is also only evaluated at window resolution; no precision/recall or event-level localization metric is reported. The authors should either provide an event-level evaluation (e.g., matched-event precision/recall and event-size correlation) or revise the abstract and conclusions to state that the model predicts window-integrated force changes, not individual avalanches.","section":"Abstract; Results ('Relevant time scales...'); Fig. 3(a); Suppl. Fig. S2"},{"comment":"The transferability claim is presented as 'demonstrated', but the evidence is limited. The training sets consist of four experiments and the test set of one experiment; with a total of eight experiments, the number of distinct held-out experiments for any given size composition is very small. Fig. 5 reports violin plots over windows rather than error bars or distributions over different held-out experiments, so the uncertainty of the cross-size R² values is not quantified at the experiment level. In addition, all pillars were milled from the same sample and the sample was not detached from the AE transducer between tests, so changes in contact quality, mounting, or transducer coupling—which affect the transfer function discussed in the Introduction—are not tested. The claim of transferability should be qualified accordingly, or experiment-level repeated held-out evaluation should be provided.","section":"Transferability of the method; Fig. 5; Methods ('Machine Learning at the fine-scale')"},{"comment":"The feature-importance analysis appears to select the best feature subsets on the basis of R² and then reports those R² values, but the Methods do not state that this selection was performed on a validation split separate from the test experiment. The Methods specify that random-forest hyperparameters were frozen after a five-fold cross-validation grid search, but no analogous statement is made for the feature-subset selection in Fig. 4(c),(d),(g),(h). If the same held-out experiment was used both to choose the best subset and to compute the reported R², the claim that a few moments (n = 3 or n = 4) achieve the optimal performance is optimistically biased. The authors should clarify the selection procedure or repeat the selection inside cross-validation folds.","section":"Results ('Machine learning model and feature selection'); Fig. 4; Methods ('Machine Learning at the fine-scale')"}],"minor_comments":[{"comment":"There is a broken sentence near 'application of ML feasible The feasibility is demonstrated'; a punctuation or sentence boundary is missing.","section":"Introduction"},{"comment":"The phrase 'saturating at a were strong anti-correlation for k → ∞' appears garbled and should be rewritten.","section":"Results ('Machine learning model and feature selection')"},{"comment":"It would be helpful to state explicitly how the force increment is computed from the 200 Hz force signal within each 300 ms window (e.g., difference between first and last samples, or mean before/after), since the AE features are computed over the same window.","section":"Methods ('Machine Learning at the fine-scale')"},{"comment":"The term 'predications' appears in place of 'predictions' in the description of the combination procedure; please correct the typo.","section":"Results ('Transferability of the method')"}],"recommendation":"major_revision","confidential_remarks":"The central idea is interesting and the evaluation design is mostly sound, but the paper's strongest claims (individual-event prediction, demonstration of transferability) outrun the evidence. The most important fix is to align the abstract and conclusions with the window-level regression target, and ideally to add an event-level evaluation. The feature-selection procedure also needs clarification to rule out selection on the test set. With those changes the paper could be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—the short version: this paper has a real result hiding behind an overstated abstract. The real result is that a random forest trained on statistical moments of windowed AE can predict the force change in 300 ms windows of Zn micropillar compression, with held-out R² around 0.60–0.65, and it transfers reasonably across the three pillar sizes. That's new for this kind of experiment—prior ML-AE work is mostly classification or damage-mode detection, not quantitative force reconstruction. The feature importance analysis is decent: lower moments and the 100 kHz/250 kHz bands carry the signal, which matches the physics of the transducers and events. The held-out protocol is clean, and the underlying data is on Zenodo, so a motivated reader could reproduce.\n\nThe soft spot that matters is the claim/evaluation mismatch. The abstract promises prediction of 'magnitude of individual deformation events' and 'outstanding prediction for the temporal location of avalanches,' but the fine-scale target is the summed force change per fixed 300 ms window. Waiting times between drops go down to ~10 ms, so multiple avalanches land in one window. The model is learning a window-integrated quantity, and R² on mostly-quiet windows can flatter it. There is no event-level precision/recall or localization test. This isn't fatal to the core idea—the paper shows AE carries quantitative information about deformation intensity on a window scale—but it is a real gap between claim and evidence, and the authors should either segment avalanches or reword the abstract to say 'window-averaged force change.'\n\nA few smaller things: the coarse-scale model adds pillar diameter as a feature and the text still says 'merely acoustic emission data' (fine-scale only, really). The RF hyperparameters from the grid search are not reported. No code or processed features are released, only the raw data from the earlier paper. These are all fixable.\n\nBottom line: worth a serious referee. The evaluation is the main issue; the core result is plausible and the transferability analysis is a useful contribution. I'd tell the editor to send it out with the expectation of major revision, and I'd want the authors to run an event-level analysis or explicitly reframe the claims.","headline":"A real window-level AE-to-force regression result is oversold as 'individual event' prediction; the evaluation gap is the main thing to fix.","tokens_in":19949,"tokens_out":4068,"would_cite":false,"duration_ms":35708,"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":"Acoustic emission alone can reconstruct the full force-time response of compressed micropillars, including the timing and magnitude of individual dislocation avalanches.","keywords":["acoustic emission","dislocation avalanche","micropillar compression","machine learning","random forest","wavelet spectrogram","force-time prediction","transferability"],"falsifier":"Prepare an experiment that keeps the deformation mechanism identical but deliberately alters the acoustic transfer function—for example, re-mount the sample with different contact pressure, change the transducer, or test a bulk specimen—then retrain the same random forest on 8 µm pillars and test on the altered geometry. If the $R^2$ score for force-drop timing and magnitude collapses while the mechanics are unchanged, the claim that the mapping comes from the avalanche source rather than from a fixed transfer function would be refuted.","tokens_in":18993,"feed_emoji":"🔊","tokens_out":7431,"duration_ms":68805,"temperature":0.7,"pith_summary":"Acoustic emission (AE) recordings from compressed single-crystalline zinc micropillars contain enough information to reconstruct the full force-time response, including the timing and magnitude of individual dislocation avalanches. The paper trains a random forest regressor on statistical moments of AE time series and on wavelet spectrogram slices to predict force increments in 300 ms windows and the coarse force level in 50 s windows, then combines both into a jerky curve that tracks the measured one. The authors show that only a few features—the energy-like second moment, the amplitude (maximum), and low-order moments—carry nearly all predictive power, and that the most informative frequencies are around 100 kHz and 250 kHz. They also report that a model trained on pillars of one or several diameters predicts force drops for unseen pillar sizes (8, 16, and 32 µm) with only moderately reduced accuracy, suggesting the method may transfer to bulk samples where direct force measurement is unavailable.","feed_headline":"Acoustic bursts alone predict micropillar force drops","feed_subtitle":"Random forest acoustic model predicts timing and size of dislocation avalanches at unseen pillar sizes.","key_machinery":"The load-bearing objects are two descriptor families extracted from equisized time windows of the AE voltage signal: frequency-independent moments of the form $\\sqrt[k]{\\langle |V|^k \\rangle}$ (mean, standard deviation, fourth root of kurtosis, and the maximum as $k\\to\\infty$) and frequency-dependent moments taken from slices of a continuous-wavelet spectrogram at chosen frequencies. A random forest regressor maps these descriptors to the force increment (fine scale, $\\Delta t=300$ ms) or the force level (coarse scale, $\\Delta T=50$ s); the two predictions are combined by adding a per-window linear slope so the final curve matches the coarse-scale shape while preserving fine-scale serrations. The analysis of single-feature and subset performance carries the argument that the energy ($k=2$), the amplitude ($k=\\infty$), and the $k=1$ moment (the measured area under the rectified signal envelope, MARSE) at frequencies near 100 kHz and 250 kHz are the informative content of the AE signal.","core_discovery":"On the paper's own terms, the discovery is that a machine-learning regressor can invert enough of the acoustic-emission measurement chain—without being given the transfer function—to reproduce the plastic response of a compressed micropillar. The authors model the recorded AE voltage as $A_{\\mathrm{AE}}(t)=V(t)*T(t)$, where $V(t)$ is the source function (the local deformation rate during an avalanche) and $T(t)$ is the transfer function encoding geometry, transducer resonance, and amplification; rather than deconvolving $T(t)$, they train a random forest on descriptors of $A_{\\mathrm{AE}}(t)$ to predict the force increment in 300 ms windows and the force level in 50 s windows. They report that the resulting combined prediction tracks both the coarse hardening and softening shape and the fine serrations of the measured force-time curve, and that a model trained without any experiment of the test pillar's diameter still achieves reasonable $R^2$ scores. The feature analysis grounds the result in physics: the energy-like second moment, the amplitude (maximum), and low-order moments, especially at the 100 kHz and 250 kHz bands, are the features that carry nearly all of the predictive power.","pith_inferences":["A natural extension would be testing whether the same descriptors predict avalanche magnitudes for other deformation mechanisms such as twinning or fracture; the feature-importance analysis suggests the 100 kHz and 250 kHz structure is specific to this basal-slip zinc system, so a general model would likely need training data per mechanism.","The saturation of performance at three to four features implies that low-cost sensors with limited bandwidth, or even single-channel recordings, might suffice for force-drop prediction—a practically testable claim.","The transferability result is consistent with the transfer function being dominated by specimen geometry and transducer resonance; a stronger test would be to vary sample mounting or use a different sample material while keeping the same transducer.","The convolution model $A_{\\mathrm{AE}}=V*T$ implies that any size-dependent change in the transfer function that is not represented in the training set will appear as a distribution shift; monitoring AE power spectra during training could flag when the model is applied outside its learned regime."],"forward_implications":["A model trained on AE descriptors can locate deformation events in time and estimate their size without threshold-based event detection, avoiding the information loss of conventional burst-counting methods.","The energy-like second moment and the amplitude of the AE signal are the most valuable predictors, and combining them with low-order moments reaches near-optimal performance with only about three to four features.","Frequency content matters: the 100 kHz resonance tail and the 250 kHz burst onset carry most of the predictive information, while the 500 kHz noise band is largely uninformative.","Trained models extrapolate to unseen pillar diameters with only moderately reduced accuracy, indicating that AE-based force prediction may work for samples whose plastic response is not directly measurable."],"supporting_citations":[{"why":"Supplies the concurrent AE and force data from in-situ SEM micropillar compression used for training and testing the model.","marker":"[51]"},{"why":"Provides additional concurrent AE microcompression experiments that extend the dataset used for the transferability analysis.","marker":"[56]"},{"why":"Defines the source-transfer-function convolution $A_{\\mathrm{AE}}=V*T$ that frames the inverse problem the machine-learning mapping bypasses.","marker":"[42]"},{"why":"Shows that dislocation avalanches emit strong, measurable acoustic signals, the physical premise for prediction from AE.","marker":"[29]"},{"why":"Demonstrates that statistical moments of acoustic or timeseries data combined with machine learning can predict slip events, supporting the descriptor choice.","marker":"[60]"},{"why":"Presents an independent route to uncovering avalanche sources via acceleration measurements, used as the comparison case for AE-based inversion.","marker":"[43]"}],"fun_headline_variants":["Machine learning decodes acoustic bursts into force drops","Acoustic emission alone predicts micropillar avalanches","Random forest maps acoustic signals to micropillar force response","Acoustic bursts reveal hidden micropillar force drops"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The transferability claim assumes the path from an avalanche to the recorded voltage—sample geometry, sensor resonance, and contact quality—stays similar enough across 8, 16, and 32 µm pillars and across the eight experiments, because the paper supports this only by the similarity of power spectra and does not directly vary contact or mounting.","fun_headline_variants_meta":{"raw":{"variants":["Machine learning decodes acoustic bursts into force drops","Acoustic emission alone predicts micropillar avalanches","Random forest maps acoustic signals to micropillar force response","Acoustic bursts reveal hidden micropillar force drops"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000551,"raw_usage":{"total_tokens":2626,"prompt_tokens":941,"completion_tokens":1685,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":1625}},"tokens_in":557,"tokens_out":1685,"duration_ms":11253,"temperature":1.0,"reasoning_tokens":1625,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:40:20.449389+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Prepare an experiment that keeps the deformation mechanism identical but deliberately alters the acoustic transfer function—for example, re-mount the sample with different contact pressure, change the transducer, or test a bulk specimen—then retrain the same random forest on 8 µm pillars and test on the altered geometry. If the $R^2$ score for force-drop timing and magnitude collapses while the mechanics are unchanged, the claim that the mapping comes from the avalanche source rather than from a fixed transfer function would be refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the concurrent AE and force data from in-situ SEM micropillar compression used for training and testing the model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides additional concurrent AE microcompression experiments that extend the dataset used for the transferability analysis."},{"cited_title":"A., Gou, B., Rooke, S","cited_arxiv_id":null,"evidence_quote":"Defines the source-transfer-function convolution $A_{\\mathrm{AE}}=V*T$ that frames the inverse problem the machine-learning mapping bypasses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that dislocation avalanches emit strong, measurable acoustic signals, the physical premise for prediction from AE."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates that statistical moments of acoustic or timeseries data combined with machine learning can predict slip events, supporting the descriptor choice."},{"cited_title":"& Shilo, D","cited_arxiv_id":null,"evidence_quote":"Presents an independent route to uncovering avalanche sources via acceleration measurements, used as the comparison case for AE-based inversion."}],"review_version":1}