{"id":"941b1027-7caa-44c1-a40c-a1558a5e6677","arxiv_id":"2411.13850","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"ALKPU selects training configurations for DeePMD using Kalman-filter-based prediction uncertainty, achieving similar accuracy to DP-GEN while training only one model.","lead":"Researchers propose a new active learning method, ALKPU, that uses a Kalman filter's uncertainty estimate to pick which molecular configurations to compute with expensive quantum chemistry and add to the training set of a neural network force field. It trains a single model instead of the four used by the standard DP-GEN method and shows comparable or better accuracy on copper, silicon, aluminum, and nickel at lower training cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.2's proof drops the residual-dependent Hessian term (Eq. 4.9), and this is precisely where ALKPU queries; without measuring that term, the claimed local optimality is unsupported.","rationale":"The reader's verdict and weakest_assumption identify Eq. (4.9) as the load-bearing point, and I agree. The paper's own proof of Theorem 4.2 explicitly relies on the approximation 1/2 grad^2_w ||y - h(w,x)||^2 approx grad h grad h^T, which equals keeping only the first-order term in a Gauss-Newton expansion. The correctness of this approximation depends on either weak nonlinearity of h or small residuals y - h; ALKPU's selection strategy deliberately seeks configurations with poor current predictions, so neither condition is guaranteed. This is not a disagreement with the community consensus; it is an internal correctness risk in the theoretical guarantee. The paper does provide useful independent support for the practical method: the method is tested on four crystal systems, the final models reproduce DFT phonon spectra, and the comparison with DP-GEN shows comparable accuracy with one model instead of four. Those experiments give evidence that ALKPU works in these cases, but they do not validate the theoretical claim that KPU maximization is locally entropy-optimal, because the omitted Hessian term is never computed or bounded. A second, smaller gap is the mismatch between the scalar-output theorem and the averaged multi-component KPUforce used in Algorithm 3; even if Eq. (4.9) held, the paper does not prove that the averaged quantity preserves the monotonicity needed for the local optimality argument. These considerations support keeping the reader's CONDITIONAL verdict: the method is plausible and empirically promising, but the central theoretical claim should be accepted only after the flat-region approximation is validated or replaced by a weaker claim. I therefore recommend no change to the reader's verdict.","tokens_in":18610,"tokens_out":5084,"duration_ms":56680,"concrete_test":"On the Cu-108 and Si-64 systems, evaluate the omitted term in Eq. (4.9) for configurations encountered during ALKPU exploration. For each candidate x, compute the two matrices A = grad h(hat w_t, x) grad h(hat w_t, x)^T and B = (y - h(hat w_t, x)) grad^2 h(hat w_t, x), using the DFT label y obtained in the subsequent labeling step and Hessian-vector products or finite differences of the force gradients. If ||B|| is comparable to or larger than ||A|| for a nontrivial fraction of the high-KPU selected candidates, then Eq. (4.10) is not a reliable estimate of entropy reduction, and Theorem 4.2's local optimality claim is unsupported in the extrapolation regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is that selecting configurations with maximal KPU locally maximizes the Shannon-entropy reduction of the weight posterior (Theorem 4.2). The proof requires Eq. (4.9): it replaces the full Hessian of the squared prediction error by grad h grad h^T, dropping all second derivatives of h(w,x). For a scalar output, retaining the dropped term gives the Laplace-approximation posterior precision hat(P)_t^{-1} + kappa( grad h grad h^T - (y - h(w,x)) grad^2 h ), so the entropy reduction is no longer the simple monotone function 1/2 ln det( I + kappa alpha_t^{-2} KPU_t(x) ) stated in Eq. (4.10). The neglected term is multiplied by the residual y - h(hat w_t, x). ALKPU is explicitly designed to select configurations where the current model is least reliable, i.e., where this residual is expected to be large; in such extrapolation regions the Gauss-Newton approximation used in Eq. (4.9) is least justified. The paper never measures the relative size of the omitted term on any of the four test systems. A secondary gap is that Theorem 4.2 is stated for a scalar output, whereas Algorithm 3 computes KPU as an average over randomly selected force components; no argument is given that maximizing or thresholding this averaged quantity inherits the local optimality result. Thus the theoretical guarantee rests on an unvalidated local-quadratic assumption exactly in the regime the method is designed to target. The experimental results show ALKPU works well on the tested systems, but they do not test the theorem's mechanism, so the central claim remains conditional.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes ALKPU, an active learning method for DeePMD that quantifies model prediction uncertainty through the Kalman Prediction Uncertainty (KPU). KPU is derived from the covariance estimate produced by the RLEKF optimizer and computed for configurations generated during an exploration MD trajectory; configurations with KPU above a lower threshold are selected for labeling, while those above an upper threshold are discarded. The authors provide a theoretical analysis claiming that max-KPU selection locally yields the fastest reduction of model uncertainty (Theorem 4.2), and they report experiments on Cu, Si, Al, and Ni systems comparing ALKPU with DP-GEN in terms of selection statistics, training convergence, final energy/force errors, and phonon spectra.","tokens_in":18948,"tokens_out":4802,"duration_ms":46615,"significance":"If the central claims are established, ALKPU would be a practically valuable single-model alternative to the four-model DP-GEN approach, reducing computational overhead while providing a principled uncertainty signal for neural-network force fields. The paper has concrete strengths: a clear algorithmic specification (KPU_cal, KPUforce), a proof-based analysis connecting the Kalman covariance to Fisher information and entropy reduction, and tests on four physical systems. However, the theoretical guarantee depends on an approximation whose validity in the extrapolation regime is not verified, and the experimental comparison is not fully controlled. These issues must be resolved before the main claims can be accepted.","major_comments":[{"comment":"The proof of Theorem 4.2 replaces the full Hessian of the squared prediction error with grad h grad h^T, dropping the term -(y - h(w,x)) grad^2 h(w,x). ALKPU is designed to select configurations where the current model is unreliable, i.e., where the residual y - h(w,x) is expected to be large, so the omitted term is largest exactly in the regime the method targets. No measurement of the relative size of this term is reported for any of the four test systems. The claim that ALKPU 'locally leads to the fastest reduction of model uncertainty' is therefore unsupported unless the approximation is validated on the candidate configurations or the claim is restricted to configurations near the training manifold.","section":"Section 4, Eq. (4.9) and Theorem 4.2"},{"comment":"Theorem 4.2 is stated for a scalar output (y has dimension 1). In practice, Algorithm 3 computes KPUforce as an average over three force components of n randomly selected atoms and uses this scalar for thresholding and selection. No argument is given that maximizing or thresholding this averaged quantity inherits the entropy-reduction property of the scalar-output theorem. The authors should either extend the theorem to vector outputs or empirically demonstrate that the averaged criterion leads to the same selections as applying the scalar criterion componentwise.","section":"Section 3.2, Algorithm 3 and Theorem 4.2"},{"comment":"The comparison between ALKPU and DP-GEN is not controlled. Table 3 reports ALKPU training with batch size 128 and 100 epochs against DP-GEN training with batch size 1 and 400 epochs, and the thresholds for DP-GEN were changed from the default (0.05, 0.15) to (0.02, 0.2) because the default 'can not work well.' The convergence and RMSE comparisons in Figures 3–4 are therefore comparing different training regimes rather than isolating the effect of the selection strategy. Moreover, all experiments are single-run; no error bars or multiple seeds are reported. Please provide a matched comparison (e.g., same compute budget or same number of optimizer steps, several random initializations) to support the efficiency conclusions.","section":"Section 5, Table 3 and Figures 3–4"},{"comment":"The statement that configurations with KPU above sigma1 'deviate too far from the real AIMD trajectory and thus should be discarded' is a design assumption, not derived from the theoretical framework. The thresholds c0 = 1.25 and c1 = 2.0 are manually chosen, and the paper does not provide a sensitivity analysis for c0, c1, or the fraction frac used in Algorithm 3. Since the selection outcome depends on these free parameters, the robustness of the reported active-learning behavior should be demonstrated.","section":"Section 3.2, threshold sigma1"}],"minor_comments":[{"comment":"The verb 'coverage' in 'it can efficiently coverage the system's configuration space' should be 'cover.'","section":"Abstract"},{"comment":"'Photon spectra' should read 'phonon spectra' throughout the experimental section and the figure caption.","section":"Section 5, Figure 6 and text"},{"comment":"The sentence beginning 'his approach uses the efficient RLEKF optimizer' is missing the initial 'T' and should read 'This approach ...'.","section":"Section 6, first sentence"},{"comment":"In the description of the smoothing function s(r), the conditions are written as 'for x< rcs' and 'for x> rc' while the argument is r; the variable should be consistent.","section":"Section 2, smoothing function"},{"comment":"The phrase 'symmetric semi-definite' should be 'symmetric positive semi-definite' to match the standard definition of the Fisher information matrix.","section":"Equation (4.6)"},{"comment":"The sentence 'a large amount of extra configurations should be select as the temporary increases from 300K to 1400K' contains grammatical errors; it should read 'a large number of extra configurations should be selected as the temperature increases from 300K to 1400K.'","section":"Section 5, text near Table 2"}],"recommendation":"major_revision","confidential_remarks":"The KPU metric is a natural by-product of the RLEKF covariance, but the theoretical result in Theorem 4.2 is close to a restatement of the Gaussian approximation, making the empirical validation the load-bearing part of the paper. The current experimental comparison is not sufficiently controlled, and the key approximation is not checked in the extrapolation regime. A major revision that addresses the controlled comparison and validates or restricts the theoretical claim would bring the paper within the standard for the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What is actually new: using the EKF covariance from the RLEKF optimizer as a one-model uncertainty measure for active learning in DeePMD. The experiments on Cu, Si, Al, and Ni are real and show ALKPU reaches comparable accuracy to DP-GEN while training one model instead of four. That is a practical benefit, and the time-cost table is candid about KPU overhead. The citation pattern is honest, including prior Kalman training work and MacKay's active learning. The theory is the soft spot. Theorem 4.2 is essentially the known MacKay result restated, and the proof is correct under the stated Gauss-Newton approximation. The problem is that the approximation drops the residual-dependent Hessian term, and ALKPU is designed to query points with high model uncertainty, i.e., where residuals are expected to be large. That is precisely where the dropped term is least negligible. The paper never measures the neglected term on any of the four systems, so the 'locally fastest reduction' claim is unsupported as a theorem about the actual selection rule. It is a heuristic rationale, not a guarantee. There is also a mismatch: the theorem is for scalar outputs, but Algorithm 3 averages KPU over randomly selected force components; no argument shows the averaged quantity inherits the optimality property. The empirical comparison is suggestive but not tightly controlled. DP-GEN uses batch size 1 and 400 epochs, ALKPU uses batch size 128 and 100 epochs; thresholds are hand-tuned; there are no error bars or multiple runs; no code or data are released. These are moderate soft spots, not fatal. Net: the method is plausible and the demonstration on four crystals warrants serious referee time, but the theory should be reframed as intuition or the approximation should be validated. A revision with multi-run statistics and a measurement of the omitted Hessian term would make the case much stronger.","headline":"A genuinely useful single-model active learning scheme for DeePMD, with a theory that is conditional in exactly the regime where it matters.","tokens_in":562,"tokens_out":589,"would_cite":false,"duration_ms":29377,"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":"The paper proposes a Kalman-filter-derived uncertainty score, KPU, and proves that labeling configurations with the highest KPU is the locally fastest way to reduce a DeePMD model's prediction uncertainty, enabling active learning with a…","keywords":["active learning","neural network force field","DeePMD","Kalman filter","uncertainty quantification","Kalman Prediction Uncertainty","RLEKF optimizer","molecular dynamics"],"falsifier":"Take a system with a double-well or strongly anharmonic potential, train a partial DeePMD model, and compare the next configurations chosen by KPU with those chosen by an acquisition rule using the exact Hessian of the log-posterior; if the rankings differ substantially, the local-fastest claim of Theorem 4.2 fails in that regime. A cheaper check is to measure the correlation between $KPU_{\\mathrm{force}}$ and the true force RMSE on an extrapolation validation set; Theorem 4.3 predicts a lower bound, so a validation set where low KPU coincides with high error would break the premise.","tokens_in":18417,"feed_emoji":"⚛️","tokens_out":5931,"duration_ms":52044,"temperature":0.7,"pith_summary":"This paper tries to establish that a quantity called the Kalman Prediction Uncertainty, or KPU, reliably measures how unsure a trained DeePMD neural-network force field is about a new atomic configuration. It then argues that an active learning loop which labels the configurations with the largest KPU is locally the fastest possible way to reduce the model's uncertainty, giving the method a theoretical justification rather than only an empirical one. The practical payoff is that ALKPU needs to train only one model, whereas the standard DP-GEN method trains four models to estimate variance, so active learning rounds become cheaper while reaching similar ab initio accuracy. If the claim holds, it makes building accurate neural-network force fields for molecular dynamics substantially less expensive in both labeling effort and compute.","feed_headline":"A single model can drive DeePMD active learning, not four","feed_subtitle":"ALKPU labels the configurations that most shrink model uncertainty, matching DP-GEN accuracy at lower cost.","key_machinery":"The load-bearing object is the Kalman Prediction Uncertainty, KPU, defined through the RLEKF optimizer's block-diagonal weight covariance $P_t$. Because RLEKF keeps a Gaussian approximation of the network weights in Kalman-filter form, the paper can compute the variance of any scalar output by back-propagating its Jacobian $D_w h(\\hat{w}_t,x)$, splitting it according to the reorganized layers, and forming $H_l P_{t,l} H_l^T$ per layer. The active learning loop uses the average of these scores over a random subset of atomic force components, $KPU_{\\mathrm{force}}$, and accepts a trajectory configuration only when this score falls in a band $[\\sigma_0,\\sigma_1]$, rejecting both low-uncertainty redundant points and very-high-uncertainty points too far from the physical trajectory. The same machinery connects the covariance to the Fisher information matrix through Theorem 4.1, which is what turns KPU into an entropy-reduction quantity.","core_discovery":"The central discovery is that the uncertainty of a DeePMD prediction can be quantified as $KPU_t(x) = D_w h(\\hat{w}_t,x) P_t D_w h(\\hat{w}_t,x)^T$, the output variance obtained by linearizing the network around the current weight estimate $\\hat{w}_t$ and using the covariance $P_t$ that the extended Kalman filter maintains for the weights. Under a quadratic approximation of the log-posterior, Theorem 4.2 derives the entropy reduction after labeling $(x,y)$ as $S_t = \\frac{1}{2}\\ln\\det\\left(I + \\kappa\\alpha_t^{-2} KPU_t(x)\\right)$, so ranking candidate configurations by KPU is the same as ranking them by expected uncertainty reduction. This makes ALKPU a locally fastest uncertainty-reduction procedure. The paper also proves that for data the model already covers, the expected squared prediction error is bounded below by the KPU (Theorem 4.3), connecting the score to actual force and energy errors. Experiments on Cu, Si, Al, and Ni show that one-model ALKPU matches the accuracy of the four-model DP-GEN approach while using less training compute.","pith_inferences":["Inference: Since KPU is defined purely from a filter covariance and a Jacobian, it could be attached to any neural network trained by a Kalman-style optimizer; the paper states this generality but demonstrates it only for DeePMD.","Inference: On strongly nonlinear potential-energy surfaces, the flat-region approximation in Eq. (4.9) may break down, so a direct comparison between KPU-ranked acquisition and exact-Hessian entropy reduction would reveal how far the local optimality extends.","Inference: The current rule scores each configuration independently; adding a diversity penalty among selected points could improve batch active learning, a direction the paper does not explore.","Inference: KPU could double as a stopping criterion: when the maximum KPU over newly explored trajectories stays below $\\sigma_0$, the model has covered that region and further labeling rounds would add little information."],"forward_implications":["ALKPU requires training one model instead of a four-model committee, so each active learning round costs less training compute while still reaching comparable final accuracy.","The entropy-reduction result makes 'label the highest-KPU configuration' a principled uncertainty-sampling rule rather than a heuristic, at least within the quadratic log-posterior approximation.","The scalar $KPU_{\\mathrm{force}}$ band $[\\sigma_0,\\sigma_1]$ gives a concrete operational criterion for deciding which explored configurations are undercovered but physical, and which are too far outside the model's reliable range.","The inequality $\\mathbb{E}_w[|y-\\hat{y}_t|^2] \\ge \\alpha_t^{-2} KPU_t(x)$ implies that, for covered data, KPU can serve as a per-configuration lower-bound proxy for the actual force and energy MSE.","Because KPU computation reuses the same back-propagation machinery as training, the active learning overhead is concentrated in one cheap post-training pass instead of multiple full model trainings."],"supporting_citations":[{"why":"Supplies the RLEKF optimizer whose block-diagonal weight covariance is the basis for computing KPU.","marker":"[33]"},{"why":"Defines DP-GEN, the four-model active learning baseline that ALKPU is compared against.","marker":"[72]"},{"why":"Describes DeePMD-kit, which supplies the network configuration, loss setup, and training conventions used in the experiments.","marker":"[64]"},{"why":"Introduces the deep potential molecular dynamics model that the active learning method is built for.","marker":"[69]"},{"why":"Defines the end-to-end symmetry-preserving potential energy model, including the embedding and fitting networks that KPU linearizes.","marker":"[70]"},{"why":"Provides the earlier entropy-reduction-based active data selection result whose flat-region approximation Theorem 4.2 extends.","marker":"[45]"},{"why":"Documents the consistency-error problem of committee-based variance that motivates the single-model ALKPU approach.","marker":"[68]"},{"why":"Provides the Kalman filtering theory underlying the extended Kalman filter training and covariance recursion.","marker":"[40]"}],"fun_headline_variants":["One-model active learning matches four-model accuracy in DeePMD","Kalman uncertainty score drives faster DeePMD training","ALKPU proves fastest uncertainty drop for DeePMD models","Active learning picks data that shrink DeePMD uncertainty most"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof that KPU ranks configurations by entropy reduction assumes the log-posterior of the network weights is quadratic, which drops all second and higher derivatives of $h(w,x)$ except the outer-product term; if strong nonlinearity prevails away from the current weights, that guarantee can fail, and the paper does not validate the assumption on the test systems.","fun_headline_variants_meta":{"raw":{"variants":["One-model active learning matches four-model accuracy in DeePMD","Kalman uncertainty score drives faster DeePMD training","ALKPU proves fastest uncertainty drop for DeePMD models","Active learning picks data that shrink DeePMD uncertainty most"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1577,"prompt_tokens":976,"completion_tokens":601,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":534}},"tokens_in":592,"tokens_out":601,"duration_ms":5698,"temperature":1.0,"reasoning_tokens":534,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:48:15.397819+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a system with a double-well or strongly anharmonic potential, train a partial DeePMD model, and compare the next configurations chosen by KPU with those chosen by an acquisition rule using the exact Hessian of the log-posterior; if the rankings differ substantially, the local-fastest claim of Theorem 4.2 fails in that regime. A cheaper check is to measure the correlation between $KPU_{\\mathrm{force}}$ and the true force RMSE on an extrapolation validation set; Theorem 4.3 predicts a lower bound, so a validation set where low KPU coincides with high error would break the premise.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the RLEKF optimizer whose block-diagonal weight covariance is the basis for computing KPU."},{"cited_title":"Zhang, H","cited_arxiv_id":null,"evidence_quote":"Defines DP-GEN, the four-model active learning baseline that ALKPU is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes DeePMD-kit, which supplies the network configuration, loss setup, and training conventions used in the experiments."},{"cited_title":"Zhang, J","cited_arxiv_id":null,"evidence_quote":"Introduces the deep potential molecular dynamics model that the active learning method is built for."},{"cited_title":"Zhang, J","cited_arxiv_id":null,"evidence_quote":"Defines the end-to-end symmetry-preserving potential energy model, including the embedding and fitting networks that KPU linearizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the earlier entropy-reduction-based active data selection result whose flat-region approximation Theorem 4.2 extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the consistency-error problem of committee-based variance that motivates the single-model ALKPU approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Kalman filtering theory underlying the extended Kalman filter training and covariance recursion."}],"review_version":1}