{"id":"51a60081-f05e-44be-9e2b-2e84c85df068","arxiv_id":"2502.06838","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A differentiable, physics-based photoresist simulator with about twenty parameters beats threshold-based baselines on LithoBench while enabling gradient-based co-optimization.","lead":"TorchResist is a new open-source photoresist simulator that models resist development with physics-based equations, all differentiable in PyTorch. It reports lower pixel and edge errors than threshold baselines on the LithoBench dataset, and runs fast on a GPU.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Superiority claim rests on comparison with only two threshold baselines; efficiency is never compared to any baseline.","rationale":"I read the paper's central claim as a comparative one: TorchResist is better than existing resist solutions. The paper's own framing in the introduction contrasts threshold-based and network-based methods, so 'existing solutions' must include the network-based class. The experimental section only benchmarks against two threshold baselines, leaving the most relevant alternative untested. This is not an attack on the internal consistency of the calibrated model; rather, it targets whether the reported accuracy numbers demonstrate superiority. The efficiency claim is even weaker because Table 2 contains no comparison at all. The reader's weakest assumption was about generalization across process conditions; that is a valid external-validity concern, and the paper indeed does not test transfer. However, I find the more directly load-bearing gap to be the missing comparison against the alternatives named in the same paper. If a compact CNN or a calibrated commercial model performs comparably, the headline claim loses its force; if baseline runtimes were measured and TorchResist is slower, the efficiency component fails outright. This concern does not overturn the paper's usefulness as an open-source differentiable simulator, and the public code plus the clear analytical formulation count in its favor, but the comparative claim needs stronger support. The reader's CONDITIONAL verdict remains appropriate; the conditions should include adding relevant baselines and a timing comparison.","tokens_in":7666,"tokens_out":6983,"duration_ms":72439,"concrete_test":"On the same LithoBench MetalSet calibration/test split, calibrate a compact CNN resist model (e.g., the neural-network-based approach of [24]) and run the existing Fixed and Variable Threshold baselines on the same GPU. Measure test Pixel Difference, EPE-mean, EPE-max, and per-tile wall-clock time for all methods. If the CNN matches or beats TorchResist on accuracy, or if the threshold baselines are substantially faster, then the 'superior accuracy and efficiency' claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the abstract is that TorchResist 'achieves superior accuracy and efficiency compared to existing solutions.' The accuracy half of this claim is tested in §3.2, Table 1, but only against Fixed Threshold [7,30] and Variable Threshold [23], both simple threshold-based methods. The introduction explicitly identifies network-based resist models [24] as the relevant alternative class, yet no neural, commercial, or otherwise state-of-the-art resist model is included in the comparison. The efficiency half is even less supported: Table 2 reports only TorchResist's absolute runtime at 7 nm/pixel and 1 nm/pixel; no baseline runtime is measured or reported. Thus the comparative claim 'superior to existing solutions' is not established by the experiments as presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"TorchResist is an open-source, differentiable photoresist simulator built on analytical Dill exposure and Mack development models with a small set of interpretable parameters. The parameters are calibrated on a LithoBench MetalSet split using gradient descent, and the simulator is evaluated against two threshold-based baselines in terms of pixel difference and edge-placement error. The paper also reports runtime at two resolutions and introduces two variants for open-source lithography models. The claimed contributions are white-box interpretability, differentiability, depth output, and superior accuracy and efficiency relative to existing solutions.","tokens_in":7816,"tokens_out":6530,"duration_ms":56284,"significance":"If the central claims hold, TorchResist would be a valuable tool for computational lithography because it combines a physical analytical model with automatic differentiation, enabling joint optimization with mask and source optimization. The open-source release is a concrete strength, and the experimental results on LithoBench show edge-placement errors well below the two threshold baselines. However, the current evidence does not establish the 'superior accuracy and efficiency compared to existing solutions' claim as stated: the comparison omits network-based and commercial resist models, and no efficiency baselines are measured. The paper's value as an open, differentiable, interpretable alternative to black-box models is clear, but the experimental support needs strengthening.","major_comments":[{"comment":"Equation (7) lists ∂I(h,t)/∂t = −I(h,t)M(h,t)C, but the Dill model derived in Eq. (3) gives ∂m1/∂t = −m1 I C, which after normalization becomes ∂M/∂t = −I M C. The equation as written is inconsistent with the boundary condition M(0,t) = exp(−RCt) in Eq. (8), which follows from ∂M/∂t rather than ∂I/∂t. Please correct the equation and re-check the subsequent derivation.","section":"§2.1, Eq. (7)"},{"comment":"The abstract and conclusion claim 'superior accuracy ... compared to existing solutions,' but Table 1 compares TorchResist only with Fixed Threshold and Variable Threshold, both threshold-based baselines. The introduction explicitly identifies network-based resist models [24] as the relevant alternative class, and no such baseline—nor any commercial resist model—is included in the comparison. Either add a network-based or commercial baseline, or restrict the claim to 'compared with threshold-based baselines.'","section":"§3.2, Table 1; Abstract; §4"},{"comment":"The efficiency claim is unsupported as presented: Table 2 reports only TorchResist's absolute runtime at 7 nm/pixel and 1 nm/pixel, with no runtime measured for any baseline (Fixed Threshold, Variable Threshold, or a network-based model). Since the abstract claims 'superior efficiency compared to existing solutions,' the paper should either measure and report baseline runtimes on the same hardware or temper the claim to an absolute runtime report.","section":"§3.2, Table 2; Abstract"},{"comment":"The numerical method for computing development depth is not specified. After Eq. (17) the text says fast-marching level-set methods 'can be employed,' but it is never stated whether the implemented simulator uses the vertical integration of Eq. (17), a fast-marching method, or another level-set variant. Because the paper lists depth simulation as a contribution and uses the development model to produce the binary results in Table 1, the actual algorithm used in the experiments must be described for reproducibility.","section":"§2.1 (Eqs. (16)–(18)); §3.1"},{"comment":"The calibrated parameter values are not reported. The paper states that 'we fix the parameters' after training and emphasizes that TorchResist has 'interpretable parameters,' but the fitted values of C, m_TH, r_max, r_min, and τ are never listed, nor is the complete budget of the 'at most twenty' parameters enumerated. Without this information, the white-box/interpretability claim cannot be checked by readers.","section":"§3.1; §2.2"}],"minor_comments":[{"comment":"There are several typos: 'T orchResist' in the abstract has a stray space; 'gradient-decent' in §2.2 should be 'gradient descent'; 'the exist of L0 norm' should be 'the existence of the L0 norm'; and 'the envelop' in §2.1 should be 'the envelope.'","section":"Abstract; §2.2"},{"comment":"The 'commercial tools' used to generate the aerial and resist images are not named, and the lithography model in Table 1 is described only as 'a commercial tool.' Please identify the tool and version to support reproducibility.","section":"§3.1"},{"comment":"The reported metrics are single-point averages without standard deviations or multiple calibration seeds; given a single calibration/test split, the stability of the 0.22% pixel difference is not quantified.","section":"§3.2"},{"comment":"In the scale-robustness test, please clarify whether the 0.17% pixel difference is computed between the 7 nm/pixel and 1 nm/pixel outputs after up-sampling, and what the reference is; as written, the comparison is ambiguous.","section":"§3.2"},{"comment":"Table 3 is labeled 'for reference' and its metrics are said not to be comparable with those in Table 1; consider explaining the intended use of the TorchResist-F/I results or removing the table, since the current text leaves the takeaway unclear.","section":"§3.2, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a promising and timely tool, and the open-source release is a strength. However, the 'superior accuracy and efficiency' claim in the abstract and conclusion is not supported by the current experiments, which lack both a network-based baseline and any runtime comparison. The typographical error in Eq. (7) should be corrected promptly, and the development solver needs to be specified. With these fixes and a modest strengthening of the evaluation, the paper could become a solid contribution to the computational lithography community."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: TorchResist is a real, shippable artifact—an open-source differentiable resist simulator with at most twenty interpretable parameters, calibrated by gradient descent on a held-out split of LithoBench, with code publicly available. That is the contribution: not new physics (the formulas are Dill 1975 and Mack 1987), but a clean differentiable implementation that also outputs development depth. The reported numbers on the MetalSet test set (0.22% pixel difference, 0.73 nm EPE-mean) beat the two threshold baselines by a comfortable margin. Credit where due: the empirical setup is standard, not circular, and the open-source code makes the work reproducible. That is real evidence.\n\nThe soft spots are in the claims, not the artifact. The abstract says 'superior accuracy and efficiency compared to existing solutions,' but the accuracy comparison is only against Fixed Threshold and Variable Threshold—both simple threshold baselines. The introduction explicitly names network-based resist models as the relevant alternative class, yet no such model is in Table 1. The efficiency claim has no baseline at all: Table 2 reports TorchResist's absolute runtime at two resolutions, with no runtime for any existing simulator. So the comparative superiority claim is not established by these experiments. It may still be true, but the paper doesn't show it.\n\nThere is also a load-bearing typo in the derivation: Equation (7) writes dI/dt = -I M C, but the correct Dill equation is dM/dt = -I M C, as the authors themselves wrote in Equation (3) for m1. That is not a cosmetic slip; it is the central kinetic equation. A referee would need that fixed. Other minor issues: no error bars or variance across tiles, no reported fitted parameter values (so no physical plausibility check), and the development solver is ambiguous—the text mentions fast-marching level-set methods but the experiments may use the simplified vertical path; the paper should state which.\n\nThe citation pattern is fine; the authors cite the classic models and relevant prior work. Self-citations point to related lithography papers and are not a problem here.\n\nWho is this for: people building differentiable lithography pipelines and wanting a lightweight resist model to co-optimize with mask or source. They will get value from the code. It deserves a serious referee: the artifact is reproducible, the approach is sound, and the claimed superiority is a fixable overreach rather than a fatal flaw. I would send it to review with a request to tighten the claims and fix Equation (7).","headline":"A useful, open-source differentiable resist simulator built on classic Dill/Mack physics, but the 'superior to existing solutions' claim overshoots the experiments, which pit it only against two threshold baselines and no efficiency baseline at all.","tokens_in":8344,"tokens_out":1924,"would_cite":true,"duration_ms":17485,"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":"TorchResist claims that a white-box analytical resist simulator with fewer than twenty interpretable parameters can beat threshold-based baselines on LithoBench MetalSet while staying fully differentiable.","keywords":["photoresist simulation","differentiable programming","analytical resist model","white-box model","lithography simulation","edge placement error","development depth","LithoBench"],"falsifier":"Compute the same predictions on aerial images generated at a different exposure dose or focus, compare against measured wafer images; if pixel difference or EPE-mean jumps well above the reported 0.22% and 0.73 nm, the fixed analytical parameters are fitted to the calibration condition rather than capturing the resist process.","tokens_in":7475,"feed_emoji":"🔬","tokens_out":7635,"duration_ms":69216,"temperature":0.7,"pith_summary":"The paper aims to show that resist simulation in computational lithography does not need a large black-box neural network: a white-box analytical model with at most twenty physically interpretable parameters can reproduce the resist outcome of a commercial tool more accurately than threshold-based baselines. This matters because lithography simulators are a bottleneck in semiconductor manufacturing cost, and existing open-source simulators either lack resist modeling or use non-interpretable, data-hungry networks. TorchResist is built so that every step from aerial image to developed resist is differentiable, which lets it be tuned end-to-end alongside mask optimization and to emit development depth for 3D downstream tasks. The reported results on LithoBench MetalSet support the claim that this combination of accuracy, efficiency, interpretability, and differentiability is simultaneously achievable.","feed_headline":"TorchResist cuts resist-image pixel error to 0.22 percent","feed_subtitle":"Under 20 interpretable parameters, it beats threshold baselines and slots into mask-optimization loops.","key_machinery":"The central object is a parameterized function that maps an aerial image to a continuous development-depth image. It chains three differentiable stages: an absorption stage following the exponential light-absorption law with three absorbing species, parameterized by the customary A, B, C constants plus inhibitor decay; a development-rate stage following the positive-resist development-rate law $r = r_{\\max}(a+1)(1-M)^n/(a+(1-M)^n) + r_{\\min}$, where $M$ is the fractional inhibitor concentration and $a$ encodes the inflection point; and a fast-marching solution of $|\\nabla T| = 1/r$ for the development time field, binarized by a learned sigmoid threshold. This machinery keeps every parameter physically interpretable while making the whole map trainable by gradient descent on aerial-image/wafer-image pairs.","core_discovery":"On its own terms, the paper's central claim is that resist modeling can be kept classical and transparent without sacrificing accuracy. TorchResist instantiates the standard exposure kinetics (light absorption through three absorbing species, first-order inhibitor destruction) and the standard development-rate curve, solves the development front by fast-marching, and fits the few remaining constants plus a binarization threshold by gradient descent. With at most twenty interpretable parameters it reaches 0.22% pixel difference and 0.73 nm EPE-mean on the LithoBench MetalSet test set, improving on both fixed- and variable-threshold baselines, while remaining differentiable end-to-end and able to emit development depth.","pith_inferences":["Editorial inference: the same calibration pipeline should transfer to other lithography simulators by refitting only the twenty parameters, which the paper's two open-source variants already hint at; a direct test would be to fit on one simulator's aerial images and evaluate on another without retraining.","Editorial inference: because the model outputs continuous development depth, it could be plugged into etch-bias or optical proximity correction loops that currently consume only binary resist contours; the paper does not demonstrate such a co-optimization.","Editorial inference: a natural stress test is to vary dose, focus, or resist thickness across the calibration set and see whether the fitted parameters become dose/focus dependent; if they do, the fixed-parameter analytical form is a benchmark-fit rather than a process model."],"forward_implications":["Resist simulation can be calibrated from a small labeled set: the paper fits all parameters in roughly one hour on a single GPU, where network-based resist models would need far more data and compute.","Because every stage is differentiable, TorchResist can be embedded in larger optimization loops such as mask optimization, with resist accuracy no longer a weak link.","The same analytical core can serve different lithography simulators by refitting parameters; the paper demonstrates this with two open-source lithography models, reporting usable accuracy for each.","Outputting development depth in addition to the binary resist image enables 3D-aware downstream tasks like etch simulation without a separate resist model.","Running at 7 nm/pixel instead of 1 nm/pixel changes predictions by only 0.17% pixel difference, so low-resolution inference can be used where speed matters."],"supporting_citations":[{"why":"Supplies the exposure kinetics equations with three absorbing species and inhibitor decay that form the exposure stage.","marker":"[16]"},{"why":"Supplies the development-rate law with rmax, rmin, and an inflection-point constant that forms the development stage.","marker":"[17]"},{"why":"Supplies the MetalSet aerial-image/wafer-image dataset and the calibration/test split used for fitting and evaluation.","marker":"[30]"},{"why":"Provides the variable-threshold baseline method against which TorchResist is compared.","marker":"[23]"},{"why":"Supplies the fixed-threshold baseline and the benchmark layout patterns underlying the MetalSet subset.","marker":"[7]"},{"why":"Supplies the fast-marching level-set method used to solve the development front equation.","marker":"[18]"},{"why":"Provides the automatic-differentiation engine that makes the whole pipeline trainable by gradient descent.","marker":"[26]"}],"fun_headline_variants":["TorchResist: 20 parameters beat threshold baselines","TorchResist: white-box resist sim, 20 params, 0.22% error","Differentiable litho resist model hits 0.22% pixel error","Open-source TorchResist: 20 params, 0.73 nm EPE","TorchResist: interpretable, differentiable, 0.73 nm EPE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the classical analytical exposure-and-development equations, with one fitted parameter set, capture the actual resist behavior of the evaluated process well enough that accuracy on the held-out test set reflects real predictive power, not just fit to the calibration split.","fun_headline_variants_meta":{"raw":{"variants":["TorchResist: 20 parameters beat threshold baselines","TorchResist: white-box resist sim, 20 params, 0.22% error","Differentiable litho resist model hits 0.22% pixel error","Open-source TorchResist: 20 params, 0.73 nm EPE","TorchResist: interpretable, differentiable, 0.73 nm EPE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001282,"raw_usage":{"total_tokens":5199,"prompt_tokens":863,"completion_tokens":4336,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":4229}},"tokens_in":479,"tokens_out":4336,"duration_ms":30722,"temperature":1.0,"reasoning_tokens":4229,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T00:42:31.779594+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the same predictions on aerial images generated at a different exposure dose or focus, compare against measured wafer images; if pixel difference or EPE-mean jumps well above the reported 0.22% and 0.73 nm, the fixed analytical parameters are fitted to the calibration condition rather than capturing the resist process.","supporting_citations":[{"cited_title":"Characterization of positive photoresist,","cited_arxiv_id":null,"evidence_quote":"Supplies the exposure kinetics equations with three absorbing species and inhibitor decay that form the exposure stage."},{"cited_title":"Development of positive photoresists,","cited_arxiv_id":null,"evidence_quote":"Supplies the development-rate law with rmax, rmin, and an inflection-point constant that forms the development stage."},{"cited_title":"Lithobench: Benchmarking ai computational lithog- raphy for semiconductor manufacturing,","cited_arxiv_id":null,"evidence_quote":"Supplies the MetalSet aerial-image/wafer-image dataset and the calibration/test split used for fitting and evaluation."},{"cited_title":"Variable-threshold resist models for lithography simulation,","cited_arxiv_id":null,"evidence_quote":"Provides the variable-threshold baseline method against which TorchResist is compared."},{"cited_title":"Iccad-2013 cad contest in mask optimization and benchmark suite,","cited_arxiv_id":null,"evidence_quote":"Supplies the fixed-threshold baseline and the benchmark layout patterns underlying the MetalSet subset."},{"cited_title":"Fast-marching level-set methods for three-dimensional photolithography development,","cited_arxiv_id":null,"evidence_quote":"Supplies the fast-marching level-set method used to solve the development front equation."},{"cited_title":"Pytorch: An imperative style, high-performance deep learning library,","cited_arxiv_id":null,"evidence_quote":"Provides the automatic-differentiation engine that makes the whole pipeline trainable by gradient descent."}],"review_version":1}