{"id":"ac53be3b-1254-496d-a7c9-65c0d2b4999b","arxiv_id":"2508.00895","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Potential Loss Analysis attributes wafer defects to individual process steps by fitting a monotone value function whose increments serve as non-negative attribution scores, replacing the zeroed-out counterfactuals of partial trajectory regression.","lead":"This paper proposes Potential Loss Analysis, a method that attributes wafer defect densities to upstream manufacturing process steps by comparing predicted outcomes along partial processing routes. It claims that framing the attribution as an optimal control problem and solving a Bellman equation improves both prediction accuracy and interpretability over partial trajectory regression.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The implemented Eq. (21) never optimizes over downstream actions, so F_theta is not the Bellman minimum of Eq. (13); attribution scores are differences of a monotone regressor, not comparisons with best possible outcomes.","rationale":"The reader's weakest assumption correctly identifies that the implemented algorithm never optimizes over downstream routes and that Eq. (21) is a soft-constrained regression on observed routes. My stress-test agrees: the min over x1,x2,... in Eq. (13) disappears entirely from the derivation, replaced by inequalities over observed transitions and a terminal regression loss. This is not a minor implementation detail; it is the difference between 'best possible outcome' and 'fitted monotone prediction'. The paper itself flags a missing proof when it says 'More detailed mathematical analysis shows that the solution must satisfy F(z_{t+1}) >= F(z_t)' without supplying it, and the non-negativity guarantee in Eq. (23) is purely a consequence of the ReLU parametrization rather than of Bellman optimality. Because the central contribution is explicitly the Bellman reduction, this concern directly undercuts the headline claim. The verdict remains CONDITIONAL rather than REJECT because the method is coherent as monotone trajectory regression with Bellman-inspired regularization, and it could be reframed honestly; the empirical improvement in correlation and the qualitative attribution results still have value. The condition should be that the authors either demonstrate genuine optimization over downstream routes (e.g., with the synthetic-MDP test above) or explicitly drop the optimal-control claim and describe PLA as regularized monotone regression.","tokens_in":10886,"tokens_out":5222,"duration_ms":57847,"concrete_test":"Construct a small deterministic MDP (e.g., 10 states, 3 actions per state, known transition costs and terminal rewards), generate training trajectories by a suboptimal policy, train PLA using Eq. (21), and compare the learned F_theta to the exact optimal value function computed by value iteration. If F_theta(z) is not equal to min_x [C(z) + F_theta(z')] for at least one state reachable in the test data, the implemented algorithm does not solve Eq. (13) and the Bellman/optimality claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that alpha_k = F_theta(z_k) - F_theta(z_{k-1}) compares best possible downstream outcomes because F_theta solves the Bellman minimization in Eq. (13). The derivation from Eq. (13) to the implemented objective Eq. (21), however, never introduces an action set, transition probabilities p(z'|z,x), or any optimization over alternative continuations. Eq. (14) is only a necessary inequality for any single observed transition; imposing it on observed (z,z') pairs and maximizing sum F_theta in Eqs. (16)-(18) yields, at best, the greatest monotone function consistent with the training routes, not the minimum over alternative downstream routes. The 'optimal downstream route' in Fig. 1(b) is never constructed; the terminal regression term (y - F_theta(z_L))^2 in Eq. (21) is the same supervised objective as PTR, and non-negativity in Eq. (23) is enforced by the ReLU parametrization of Eq. (22), not by any optimality property. Thus F_theta is a monotone trajectory regressor, and alpha_k is a difference of fitted predictions, not a counterfactual comparison against the best possible outcome. If the training data do not cover the relevant counterfactual continuations, the attribution scores lack the causal semantics claimed in Sec. 4.2. This is load-bearing because the Bellman/optimality framing is the paper's stated contribution over PTR.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes Potential Loss Analysis (PLA), a framework for attributing wafer defect densities to individual process steps in heterogeneous processing routes. PLA models wafer processing as a sequential decision process and defines an optimal expected cumulative loss F* through a Bellman minimization over downstream actions (Eq. (13)). The paper claims that a tractable objective (Eq. (21)) solves this optimization jointly with prediction, and defines the attribution score for process k as α_k = F_θ(z_k) - F_θ(z_{k-1}) (Eq. (23)), guaranteed non-negative through a ReLU-parameterized difference (Eq. (22)). Experiments on 787 wafers from an FEOL fab report improved prediction correlation over PTR (0.87 vs. 0.61) and more interpretable cumulative attribution curves. The main conceptual claim is that PLA removes the arbitrary reference point of Shapley-style methods by comparing best possible downstream outcomes.","tokens_in":11114,"tokens_out":10768,"duration_ms":103230,"significance":"If the Bellman-optimality interpretation were valid, PLA would be a meaningful step: it would provide reference-free, path-dependent attribution and simultaneous prediction from a single fitted value function, and the empirical study on real fab data is a useful point of reference for the semiconductor manufacturing community. The paper is clearly written, and the idea of imposing monotonicity to obtain non-negative, interpretable attribution scores is attractive. However, the connection between the Bellman equation and the implemented objective is not established, and the empirical comparison does not isolate the method's contribution. As presented, the central claim is substantially overstated; the framework reduces to a monotone trajectory regressor whose attribution scores are differences of fitted predictions. The paper's value would be greatly improved by either providing a rigorous derivation of the Bellman reduction with an explicit action set and transition model, or reframing the contribution as a heuristic monotone attribution method and validating it against ground truth.","major_comments":[{"comment":"The derivation of the implemented algorithm from the Bellman equation is a load-bearing gap. Eq. (13) defines F* as a minimum over downstream actions under a stochastic transition model p(z'|z,x), but no action set, transition probabilities, or optimization over alternative continuations is ever introduced. After assuming deterministic transitions, Eq. (15) is only a necessary inequality for a single observed transition. Imposing this inequality on the observed (z,z') pairs and maximizing ΣF_θ (Eqs. (16)–(18)) yields, at best, the largest monotone function consistent with the training routes, not the minimum over alternative downstream routings. The 'optimal downstream route' in Fig. 1(b) is therefore never constructed. Because the claimed advantage of PLA over PTR rests on this optimality interpretation, the central contribution is not supported as stated.","section":"§4.2–4.4, Eqs. (13)–(21)"},{"comment":"The final objective is a supervised regression with a monotonicity penalty, not a Bellman solver. Eq. (21) combines the terminal squared loss (y - F_θ(z_L))^2, which is the same supervised objective used in PTR, with a soft penalty on consecutive observed states. Non-negativity of α_k is enforced by the ReLU parameterization in Eq. (22), not derived from any optimality property. Consequently, α_k = F_θ(z_k) - F_θ(z_{k-1}) is a difference of fitted predictions from a monotone regressor, and its interpretation as a comparison of best possible outcomes is not established. The statement after Eq. (21) that 'more detailed mathematical analysis' implies F_θ(z_{t+1}) ≥ F_θ(z_t) is unproved; it should be either proved or explicitly labeled as an assumption.","section":"§4.4–4.5, Eqs. (21)–(23)"},{"comment":"The empirical evaluation does not provide controlled evidence for the attribution claims. PTR is implemented with a linear model while PLA uses a two-hidden-layer neural network, so the reported prediction-correlation improvement (0.87 vs. 0.61) is confounded by model capacity. More importantly, the attribution scores are evaluated only qualitatively on two wafers, with no ground-truth root causes, ablations, or comparison to alternative attribution methods. The claim that PLA 'pinpoints problematic processes' is thus supported only by visual inspection of two examples. A fair comparison (e.g., PTR with the same network architecture) and a quantitative attribution evaluation would be necessary to support the practical claims.","section":"§5.2, Fig. 4"}],"minor_comments":[{"comment":"Eq. (18) writes R(θ|µ), but the final objective in Eq. (21) involves both µ and µ_i; please clarify the relationship between the two hyperparameters.","section":"§4.4, Eq. (18)"},{"comment":"The sentence following Eq. (21) refers to 'more detailed mathematical analysis' without a proof or a reference; if the monotonicity property is a theorem, include its proof or a pointer to a complete derivation.","section":"§4.4, Eq. (21)"},{"comment":"In Sec. 5.2, 'correlation efficient' should read 'correlation coefficient.'","section":"§5.2"},{"comment":"Fig. 4 uses '(C)' and '(D)' in the caption but the text cites 'Figs. 4 (c) and (d)'; please make the capitalization consistent.","section":"Fig. 4"},{"comment":"References contain malformed author lists, e.g., 'Han, et al., S.' and 'Lvet al.'; these should be corrected.","section":"References"},{"comment":"Eq. (24) uses F_θ(z_0) without defining z_0; if z_0 denotes the zero initial state used in Eq. (8), please state this explicitly.","section":"Eq. (24)"},{"comment":"The derivation depends on the unpublished companion paper Miyaguchi (2025); for journal publication, the present manuscript should be self-contained, especially for the key step from Eq. (13) to Eq. (21).","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The editor may wish to note that the paper's central derivation relies on the companion preprint Miyaguchi (2025), and the present manuscript does not provide enough detail to verify the Bellman reduction. If the authors can supply the missing action model and proof, or alternatively reframe the method as a heuristic, the paper could be publishable after substantial revision. I do not see grounds for rejection on novelty, but the current claims exceed what the method implements."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a plausible engineering method wrapped in a theory claim the implementation doesn't support. The actual contribution is a monotone trajectory regressor that produces non-negative per-step blame scores and better defect-density prediction on one FEOL dataset. Read that way, it's a reasonable applied paper; read as Bellman optimality over downstream actions, it falls short.\n\nWhat's genuinely useful: framing attribution as a difference of a monotone value function is new in wafer defect analysis and avoids the arbitrary baseline that plagues Shapley-style and PTR scores. The same model does regression and attribution. The ReLU-parametrized increment is a clean way to enforce non-negativity. The empirical setting is real: 787 wafers, hundreds of steps, FEOL PLY data. PLA's correlation of 0.87 versus 0.61 for a linear baseline is a real improvement, and the two wafer traces suggest jumps align with long waits at tools.\n\nWhere it gets soft. The central derivation has a load-bearing gap. Eq. (13) defines F* as a minimum over actions; but the algorithm never models an action set, transition probabilities, or alternative continuations. Eq. (21) is a soft-constrained regression over observed transitions. Imposing Bellman inequalities on observed (z,z') pairs and maximizing F gives the greatest monotone function consistent with the training routes—not the best possible outcome over counterfactual downstream routes. Non-negativity is enforced by the ReLU parametrization in Eq. (22), not by any optimality property. The paper leans on Miyaguchi 2025 for the 'more detailed mathematical analysis,' but as presented the theory overstates what is computed. This matters because the Bellman/optimal-route framing is the stated contribution over PTR. The circularity concern is real but secondary: the blame scores and the prediction come from the same fitted model, and there is no independent validation that the attribution is correct. Empirically, one proprietary dataset, no error bars, and qualitative validation on two wafers. The baseline comparison is also odd: the predictive comparison is against linear regression, while PTR is only compared visually.\n\nBottom line: I'd send this to peer review. The method is coherent and potentially useful, but a referee should push either for a demonstration of genuine optimization over downstream routes or for an honest reframing as Bellman-inspired monotone trajectory regression, plus error bars and quantitative attribution validation. The flaws are fixable; the overclaim is not disqualifying on its own.","headline":"Useful but overclaimed: a monotone trajectory regressor dressed as Bellman optimal control, with real data and a fixable gap.","tokens_in":11730,"tokens_out":5725,"would_cite":false,"duration_ms":66326,"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 claims that wafer defect blame can be computed per process step by solving a Bellman equation, giving a non-negative score per step plus a defect-density predictor from one fitted value function.","keywords":["potential loss analysis","wafer defect attribution","cross-process root cause analysis","Bellman equation","partial trajectory regression","process embedding","non-negative attribution","semiconductor manufacturing"],"falsifier":"Run a controlled experiment with two downstream routes after a fixed upstream step: one route that repairs most defects and one that leaves them. Train PLA only on wafers that took the non-repairing route. If the blame score assigned to the upstream step does not decrease when wafers with the repairing route are then added to the training set, the framework is not actually comparing optimal downstream routes; it is only fitting the observed transitions.","tokens_in":10565,"feed_emoji":"🔬","tokens_out":13470,"duration_ms":118234,"temperature":0.7,"pith_summary":"The paper proposes Potential Loss Analysis, a framework that attributes wafer defect density to individual upstream processes by comparing, for each partial processing route, the best possible outcome from that point onward rather than arbitrarily zeroing out later process steps. It claims that identifying this best possible outcome reduces to solving a Bellman equation, and that the solution yields a single fitted value function $F_\\theta(z)$ that simultaneously predicts terminal defect density and assigns each process a non-negative blame score $\\alpha_k = F_\\theta(z_k) - F_\\theta(z_{k-1})$. On 787 real wafers from a front-end-of-line process, the fitted function reaches a correlation of 0.87 with measured log defect density, versus 0.61 for the partial-trajectory regression baseline, and the cumulative attribution curves flag long wait times at specific tools. If the framework is right, fab engineers get an interpretable, reference-free per-step blame signal and a defect-density predictor from the same model.","feed_headline":"Bellman equation blames wafer defects step by step","feed_subtitle":"PLA compares best-possible downstream routes instead of arbitrary zeroed baselines, and predicts defect density too.","key_machinery":"The central object is the Bellman equation $F^*(z_1)=\\min_{x_1,x_2,\\ldots}\\mathbb{E}\\left[\\sum_t C(z_t)\\right]$, which defines the best expected cumulative loss achievable from a wafer state by choosing optimal downstream process routes. Since the full optimization over unobserved routes is intractable, the paper replaces the Bellman optimality condition with the inequality $F_\\theta(z) \\le C(z) + F_\\theta(z')$ on observed transitions, enforced as a temporal-difference-style penalty in the objective. The difference in value between consecutive states is modeled by a positive-output network $G_\\theta(z_{t},z_{t+1}) = F_\\theta(z_{t+1}) - F_\\theta(z_t)$, and this network is the attribution engine: its output at transition $k$ is the blame score $\\alpha_k$.","core_discovery":"The central discovery is that process attribution can be reframed from comparing predictions with and without a process to comparing the best possible outcomes with and without a process. The paper formalizes the best possible outcome as $F^*(z_1) = \\min_{x_1,x_2,\\ldots} F(z_1 \\mid x_1,x_2,\\ldots)$, the Bellman optimal value for the expected cumulative defect loss from wafer state $z_1$, and defines the attribution score of process $k$ as $\\alpha_k = F_\\theta(z_k) - F_\\theta(z_{k-1})$, the fitted value increase along the observed transition. Because the value function is constrained to be non-decreasing along every observed transition, these scores are guaranteed non-negative and carry no dependence on an arbitrary reference point such as a zeroed-out process embedding or a population mean. The same fitted $F_\\theta$ also predicts the terminal defect density through the regression term of the objective, so attribution and prediction share one model rather than requiring a separate explanation step.","pith_inferences":["Beyond the paper, the same Bellman-inequality surrogate would transfer to any sequential process with embedded actions and a terminal outcome, such as other multistage manufacturing lines or clinical treatment pathways.","The non-negativity guarantee is a property of the model class, namely a positive-output difference network, not evidence about causal structure; a testable extension would perturb a single process embedding and check whether the blame shifts to the process that actually changed.","The framework's blame definition depends on the state representation carrying enough information about upstream history; if two distinct upstream states map to the same $z$, attribution between them is unidentifiable, so measuring score stability under embedding dimensionality is a natural stress test.","If a fab later changes a downstream recipe to reduce defects, PLA's upstream scores should fall even when upstream processes are untouched; this is a concrete prediction that could validate the optimal-downstream interpretation on real data."],"forward_implications":["Attribution scores are non-negative by construction, so cumulative defect-blaming curves never dip below zero and stay interpretable even for wafers with near-average defect density.","One fitted value function does double duty as a defect-density predictor and an attribution model, so fab analysis no longer needs a separate explainability step bolted onto a prediction model.","Because downstream process embeddings are not zeroed out, the scores do not depend on the coordinate origin or population mean of the process embedding, removing the arbitrary-reference-point problem faced by attribution methods that require a baseline input.","Cumulative PLA curves can point to specific events in the route, such as long wait times at a tool, giving process engineers a direct lead for root-cause investigation.","Training requires only observed route histories and terminal defect measurements, and with 787 wafers the framework is usable in real fab settings where the number of wafers is small relative to route combinatorics."],"supporting_citations":[{"why":"Supplies the partial trajectory regression baseline and the process-embedding method that PLA builds on.","marker":"(Miyaguchi, Joko, Sheraw, and Idé 2025b)"},{"why":"Provides the linear Bellman-inequality formulation on which the tractable objective is based.","marker":"(De Farias and Van Roy 2003)"},{"why":"Introduces the trajectory advantage regression idea that first discusses the Bellman inequality in this defect-attribution context.","marker":"(Miyaguchi 2025)"},{"why":"Supplies the dynamic-programming and Bellman equation formalism used to define the optimal expected cumulative loss.","marker":"(Bertsekas 2012)"},{"why":"Defines the substring kernel used to build the process token embeddings that feed the state model.","marker":"(Lodhi, Saunders, Shawe-Taylor, Cristianini, and Watkins 2002)"},{"why":"Supplies the potential-outcome framework that justifies comparing counterfactual partial trajectories.","marker":"(Rubin 2005)"}],"fun_headline_variants":["Wafer defect blame via Bellman optimal paths","Potential Loss Analysis: blame processes by best outcomes","New PLA uses Bellman to find defect culprits","Attributing wafer defects without arbitrary baselines","Bellman-based blame for wafer defects"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the training wafers span enough alternative downstream routes to pin down what the best possible continuation really is; the algorithm never generates or optimizes over unseen routes, so if the data contain only one kind of downstream continuation, the fitted best outcome is only the best outcome visible in the data.","fun_headline_variants_meta":{"raw":{"variants":["Wafer defect blame via Bellman optimal paths","Potential Loss Analysis: blame processes by best outcomes","New PLA uses Bellman to find defect culprits","Attributing wafer defects without arbitrary baselines","Bellman-based blame for wafer defects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00018,"raw_usage":{"total_tokens":1268,"prompt_tokens":875,"completion_tokens":393,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":323}},"tokens_in":491,"tokens_out":393,"duration_ms":3844,"temperature":1.0,"reasoning_tokens":323,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:44:33.947019+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled experiment with two downstream routes after a fixed upstream step: one route that repairs most defects and one that leaves them. Train PLA only on wafers that took the non-repairing route. If the blame score assigned to the upstream step does not decrease when wafers with the repairing route are then added to the training set, the framework is not actually comparing optimal downstream routes; it is only fitting the observed transitions.","supporting_citations":[],"review_version":2}