{"id":"766be015-b45d-4013-9a67-108345e7d4fc","arxiv_id":"1908.01244","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"A stacked LSTM trained collectively on several MOSFETs predicts the degradation of an unseen MOSFET from its resistance trajectory, with 8.9% error at the 0.05 ohm failure threshold.","lead":"This paper trains a stacked LSTM neural network on degradation data from four MOSFET power transistors and predicts the aging curve of a fifth, unseen transistor. It reports 8.9% average miss-prediction at a critical resistance threshold, with edge inference in 26 ms and 1.87 W on an NVIDIA TX2 board.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 selects the saved model on the held-out device's test data, so the 8.93% leave-one-device error and the 1.98x/1.77x gains are not clean generalization estimates.","rationale":"The paper's central claim is that a stacked LSTM trained collectively on four MOSFET devices can predict an unseen fifth device's delta Rds(on) with roughly 9% error at the 0.05 ohm threshold, beating Kalman and Particle filters by about 2x, while running on a TX2 edge board. The load-bearing condition is therefore that the reported leave-one-device evaluation is a genuine out-of-sample measurement. Algorithm 1 violates that condition by using the held-out device's test batch to select the checkpoint with minimum test error. This is not a minor implementation detail: the saved model is chosen because it fits the test device, so the reported numbers are a best-case selection result, not an unbiased estimate of prediction error on a new device. The comparison with Kalman and Particle filters is also affected because those baselines do not enjoy the same selection mechanism. I do not see this as a reason to reject the paper: the framework, open-source implementation, real NASA accelerated-aging data, and the TX2 power/delay measurements are genuine contributions, and the accuracy claim may survive a corrected protocol. But the claim as stated is not yet established. The reader's overall CONDITIONAL verdict is right, although I would locate the weakest assumption differently: rather than the input being only resistance history (a transfer limitation conceded in the framework's own description of monitoring voltage/current/temperature), the immediate problem is the test-driven checkpoint selection inside the training algorithm. A re-run with a clean validation protocol and repeated seeds would settle it.","tokens_in":14887,"tokens_out":5370,"duration_ms":53542,"concrete_test":"Using the released Deep_RACE code, rerun the five leave-one-device scenarios under two protocols: (A) exactly as Algorithm 1 is written, saving the checkpoint with minimum test error; and (B) with test-based selection removed, either by holding out a validation split from the four training devices and selecting by validation loss, or by fixing the number of training iterations and scoring the held-out device only once at the end. Keep all Table 1 hyperparameters fixed. Repeat both protocols with at least 10 random initializations per scenario and report mean +/- standard deviation for log(MSE) (Table 3) and for the Eq. (15) absolute error at 0.05 ohm (Table 4). If protocol (B) raises the average error above roughly 12% or brings it within the Kalman/Particle interval, the claimed accuracy is an artifact of test-set selection; if the error stays near 8.93%, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 5.3, Deep RACE is claimed to predict a 'completely new and unknown device' from the other four devices (Table 1 sets m=4). Algorithm 1, however, uses Xtest and Ytest inside the training loop: after each optimization step, lines 12-18 compute the loss on the test batch and save the network only when that test error is a new minimum. The reported Tables 3 and 4 are then produced from this test-selected checkpoint. This is test-set-based model selection: the held-out device's trajectory influences which weights are reported, so the 8.93% figure is an optimistic in-sample selection result rather than an unbiased forecast of an unseen device. A device deployed in the field would have no such selection signal. Since the Kalman Filter and Particle Filter baselines do not receive this selection advantage (or at least none is described), the 1.98x and 1.77x improvements are not yet fairly established. The paper also reports no error bars over random initializations, and Table 4 gives a single number per method, so the size of the bias cannot be assessed from the text. This is the most load-bearing weakness because it directly controls the paper's central quantitative claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Deep RACE, a cloud-edge system for real-time reliability modeling of Si-MOSFET power converters. The algorithm side is a stacked LSTM trained in the cloud on aggregated delta-Rds(on) trajectories from multiple devices, with inference deployed on an NVIDIA TX2 edge node. The authors evaluate the approach on five NASA accelerated-aging MOSFET datasets in a leave-one-device-out protocol, reporting a miss-prediction error of 8.93% at the 0.05-ohm detection point, improvements of 1.98x and 1.77x over Kalman Filter and Particle Filter baselines, and measured edge inference cost of 26 ms and 1.87 W. The paper also describes scalability of the aggregated training and a prototype hardware setup.","tokens_in":15197,"tokens_out":3569,"duration_ms":36041,"significance":"If the reported accuracy holds under a clean evaluation protocol, the work would be a useful demonstration of collective deep-learning-based prognostics for power semiconductors, and the measured edge deployment figures (26 ms, 1.87 W) provide concrete evidence of feasibility. The open-source availability of the code is also a strength. However, the central quantitative claim is compromised by a test-set model-selection step in the training algorithm, so the significance of the result depends on whether the evaluation can be corrected and the conclusions re-established.","major_comments":[{"comment":"Algorithm 1, lines 12-18, uses the held-out device's Xtest and Ytest inside the training loop to select and save the network with the minimum test error. This is test-set-based model selection: the reported Tables 3 and 4 are produced from a checkpoint chosen using the labels of the supposedly unseen device. Because the Kalman Filter and Particle Filter baselines receive no such selection advantage, the claimed 1.98x and 1.77x improvements are not fairly established. The evaluation must be changed so that the test device contributes no information to training or model selection; for example, use a validation set drawn from the training devices for checkpoint selection, or use nested cross-validation.","section":"Algorithm 1, Section 4.2.1"},{"comment":"Table 4 reports a single miss-prediction number per method, with no variance, error bars, or repeated-run statistics. Given the LSTM's random initialization (Algorithm 1, line 5) and the small five-device sample, a single run cannot support the claim of superiority. Report the mean and standard deviation over multiple random seeds and, ideally, a paired statistical comparison with the baselines.","section":"Section 5.3, Table 4"},{"comment":"The text describes the leave-one-device experiment as predicting a 'completely new and unknown device,' but the algorithm actually uses the test device's full trajectory for model selection at every iteration. This contradicts the generalization claim. The experimental protocol must be revised so that no information from the test device reaches the training or model-selection pipeline; otherwise the 8.93% figure is an optimistic in-sample selection result, not an unbiased forecast.","section":"Section 5.3, Fig. 10"}],"minor_comments":[{"comment":"The abstract reports '8.9%' while Table 4 reports '8.93%'; unify the rounding for consistency.","section":"Abstract and Section 5.3"},{"comment":"The notation in Eq. (15), especially the subscript/superscript 'mt5%', is not defined clearly; please define m, t5%, and the meaning of the superscript in the equation or its caption.","section":"Section 5.3, Eq. (15)"},{"comment":"There are several language errors, such as 'infertile' in Section 3.1 (should likely be 'ineffective') and 'patriarchy' in Section 4.1.2 (should be 'hierarchy'). A careful proofreading pass is needed.","section":"Section 3.1 and Section 4.1.2"},{"comment":"Section 4.2.2 states that the edge node continuously monitors voltage, current, and temperature, but the LSTM input size is k=1 (Table 1), so only delta Rds(on) is used. Clarify whether the other signals are used by the reliability model or only by the controller for other purposes.","section":"Section 4.2.2 and Table 1"},{"comment":"The y-axis of Fig. 12 is labeled log(MSE) with negative values; clarify whether the plotted quantity is log10(MSE) and describe the Monte-Carlo averaging procedure more precisely.","section":"Section 5.3, Fig. 12"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to be the arXiv posting of a manuscript already published in the IEEE Internet of Things Journal. If this report is for a re-review, the authors should be made aware that the test-set selection issue in Algorithm 1 is a serious methodological problem that affects the paper's headline accuracy and comparison claims. It is correctable with a revised experimental protocol, so I do not recommend rejection, but the current form should not be accepted as-is."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this is a legitimate engineering application of stacked LSTMs to MOSFET degradation forecasting, with a real edge deployment on a TX2, but the headline generalization claim is not supported as reported. The stress-test note is correct: Algorithm 1 selects the saved network using the held-out device's test batch (lines 12-18). That makes the 8.93% figure an in-sample selection result, not an unbiased forecast of an unseen device. The Kalman and Particle filter baselines don't get that advantage, so the 1.98x/1.77x improvement claims are not fairly established as stated.\n\nWhat's genuinely new: applying collective leave-one-device training to the NASA MOSFET resistance dataset, and demonstrating a working cloud-edge system with 26ms inference at 1.87W on an embedded board. The paper ships code, which helps reproducibility, and the system integration is described concretely. For a five-device dataset, the idea of training on four and predicting the fifth is a reasonable test of unit-to-unit generalization. The learning curve in Fig. 12 (MSE decreasing with more devices per batch) is a nice demonstration of the scalability argument, though it's also affected by the same test-selection issue.\n\nSoft spots, in order: (1) The test-based model selection is load-bearing and needs to be fixed. A proper evaluation would hold out the test device entirely, select the model on training/validation from the four training devices, and only then evaluate on the fifth. (2) No error bars over random initializations. Table 4 gives single numbers. Given only five devices and one initialization per scenario, the 8.93% could easily be the lucky draw. (3) Baselines are thin: the Kalman and Particle filter results are cited from prior work, but the protocol is not fully described here, and it's unclear if they were given the same training data or the same prediction window. (4) The model uses only the resistance trajectory as input (k=1), despite the system monitoring voltage, current, and temperature. The claim that resistance history alone encodes degradation might be true for this dataset, but it's an assumption that should be stated and ideally tested.\n\nThis is not a desk-reject paper. The engineering is real, the application is relevant, and the evaluation flaw is correctable. A serious referee should be assigned, with the request to have the authors redo the leave-one-device evaluation with a clean held-out protocol and report variance over seeds. As it stands, cite with caution; the accuracy number is not yet trustworthy.","headline":"A useful edge-reliability application, but the headline accuracy is compromised by test-set-based model selection in Algorithm 1.","tokens_in":15699,"tokens_out":1715,"would_cite":false,"duration_ms":15276,"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":"A stacked LSTM trained on four MOSFETs' resistance histories can predict a fifth device's degradation trajectory with about 9% error at the 0.05-ohm warning threshold, running in 26 ms on an embedded edge board.","keywords":["power MOSFET reliability","stacked LSTM","edge computing","prognostics and health management","remaining useful life prediction","on-state resistance degradation","cloud-edge IoT","accelerated aging data"],"falsifier":"Run Deep RACE on resistance histories collected from MOSFETs in a high-frequency converter under varying load and temperature. If the average absolute error at the 0.05 ohm threshold exceeds 8.93%, or if feeding voltage, current, and temperature as additional inputs materially reduces the error, then the single-input collective-training claim is not general.","tokens_in":14737,"feed_emoji":"⚡","tokens_out":5428,"duration_ms":53382,"temperature":0.7,"pith_summary":"This paper claims that a deep recurrent network can learn a degradation model shared across power MOSFETs of the same type, then predict the resistance-wear trajectory of a device it has never seen. The proposed system, Deep RACE, trains a stacked LSTM in the cloud on aggregated $\\Delta R_{\\mathrm{ds(on)}}$ histories and runs inference on an embedded edge node next to the converter. In leave-one-device-out tests on five accelerated-aging trajectories, the network reaches about $8.9\\%$ absolute average error at the $0.05\\,\\Omega$ warning threshold, roughly 1.98 times better than a Kalman filter and 1.77 times better than a particle filter, while taking 26 ms and about 1.87 W on the edge board. If correct, this makes fleet-learned, real-time reliability awareness practical for high-frequency converters.","feed_headline":"Fleet-trained LSTM forecasts unseen MOSFET wear at 9% error","feed_subtitle":"Trained on four devices' resistance histories, it beats Kalman and particle filters at the 0.05-ohm warning point.","key_machinery":"The load-bearing mechanism is a stacked LSTM: four LSTM layers with 64 hidden units each, followed by a dense layer that maps the normalized hidden state $h_t$ back to physical $\\Delta R_{\\mathrm{ds(on)}}$ values. Training batches are three-dimensional tensors built from randomly selected sequences of length $\\tau + n = 21 + 104$ from each of $m$ devices, so every gradient update sees multiple devices' degradation patterns. The input size is $k=1$: only the resistance trajectory, not voltage, current, or temperature, is fed to the network. The same computation graph is trained on the cloud and then shipped to the edge node, where inference evaluates the LSTM equations on the embedded processor.","core_discovery":"The paper's central claim is that aggregating degradation knowledge across many devices of the same underlying physics, rather than fitting each device in isolation, is what makes accurate unseen-device prediction possible. For each test, the network is trained from scratch on four devices' $\\Delta R_{\\mathrm{ds(on)}}$ sequences and asked to forecast 104 future resistance samples of the fifth device from the last 21 observed samples. The resulting predictions track the measured trajectories, with the error-distribution plots showing average maximum error below 0.9% and the 0.05-$\\Omega$ detection point showing $8.93\\%$ absolute error. The paper also reports that increasing the number of devices in each training batch lowers prediction MSE at an exponential rate, and that the trained inference runs in real time on a low-power embedded processor.","pith_inferences":["Beyond the paper: the evaluation uses only resistance sequences from an accelerated-aging campaign, so in the field, where load, thermal, and environmental histories vary, the single-input model may need voltage, current, or temperature channels to hold its 8.93% error; a direct field trial is the natural next test.","If resistance history alone suffices, edge nodes could upload only $\\Delta R_{\\mathrm{ds(on)}}$ samples instead of raw multi-channel telemetry, sharply reducing cloud bandwidth.","The exponential improvement with fleet size suggests a continual-learning loop: each edge device's later measurements could be recycled as training data to retrain the shared model, making the fleet collectively smarter over its lifetime.","The same edge-cloud contract could be transferred to IGBT or GaN devices by swapping in their precursor signals, though the paper demonstrates only Si-MOSFETs."],"forward_implications":["A converter entering service can receive a reliability model trained on its predecessors before it accumulates any failure data of its own.","As edge nodes are added and their resistance histories join the training pool, prediction error should keep falling, since the paper measures an exponential drop in MSE with more devices per batch.","The 0.05 ohm detection point can be used as an early-warning threshold to trigger load sharing or maintenance before resistance wear becomes critical.","The same cloud-training/edge-inference structure applies to other semiconductor families as long as a measurable degradation precursor time series exists."],"supporting_citations":[{"why":"Supplies the MOSFET accelerated-aging dataset used for all training and testing and the particle-filter comparison.","marker":"[28]"},{"why":"Provides the Kalman-filter-based baseline on on-state resistance that Deep RACE is compared against.","marker":"[30]"},{"why":"Introduces the LSTM cell architecture that the stacked network is built from.","marker":"[34]"},{"why":"Provides backpropagation through time, the optimization procedure used to train the stacked network.","marker":"[31]"}],"fun_headline_variants":["Edge LSTM forecasts MOSFET wear with 9% error","Fleet-trained AI predicts unseen MOSFET degradation","Edge deep learning for real-time MOSFET reliability","LSTM at edge predicts converter wear from fleet data","Deep RACE: edge AI for fleet-wide MOSFET reliability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed accuracy rests on the premise that the resistance trajectory alone encodes enough degradation information, and that the accelerated-aging trajectories used for training behave like real converter duty cycles; if either fails, the leave-one-device numbers will not transfer to the field.","fun_headline_variants_meta":{"raw":{"variants":["Edge LSTM forecasts MOSFET wear with 9% error","Fleet-trained AI predicts unseen MOSFET degradation","Edge deep learning for real-time MOSFET reliability","LSTM at edge predicts converter wear from fleet data","Deep RACE: edge AI for fleet-wide MOSFET reliability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000874,"raw_usage":{"total_tokens":3790,"prompt_tokens":961,"completion_tokens":2829,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":2755}},"tokens_in":577,"tokens_out":2829,"duration_ms":22729,"temperature":1.0,"reasoning_tokens":2755,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:19:03.852221+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Deep RACE on resistance histories collected from MOSFETs in a high-frequency converter under varying load and temperature. If the average absolute error at the 0.05 ohm threshold exceeds 8.93%, or if feeding voltage, current, and temperature as additional inputs materially reduces the error, then the single-input collective-training claim is not general.","supporting_citations":[{"cited_title":"Prognostics of power MOSFETs under thermal stress accelerated aging using data-driven and model-based methodologies,","cited_arxiv_id":null,"evidence_quote":"Supplies the MOSFET accelerated-aging dataset used for all training and testing and the particle-filter comparison."},{"cited_title":"Remaining useful lifetime estimation for thermally stressed power MOSFETs based on on-state resistance variation,","cited_arxiv_id":null,"evidence_quote":"Provides the Kalman-filter-based baseline on on-state resistance that Deep RACE is compared against."},{"cited_title":"Backpropagation through time: what it does and how to do it,","cited_arxiv_id":null,"evidence_quote":"Provides backpropagation through time, the optimization procedure used to train the stacked network."}],"review_version":1}