REVIEW 5 major objections 5 minor 31 references
TorchResist: Open-Source Differentiable Resist Simulator
T0 review · 5 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash
Pith's one-line read 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.
desk verdict 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. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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.
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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 (5)
- [§2.1, Eq. (7)] 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.
- [§3.2, Table 1; Abstract; §4] 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.'
- [§3.2, Table 2; Abstract] 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.
- [§2.1 (Eqs. (16)–(18)); §3.1] 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.
- [§3.1; §2.2] 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.
minor comments (5)
- [Abstract; §2.2] 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.'
- [§3.1] 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.
- [§3.2] 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.
- [§3.2] 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.
- [§3.2, Table 3] 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.
Circularity Check
No significant circularity: the resist model follows classical Dill/Mack equations, parameters are calibrated on a training split and evaluated on a held-out test split, and no fitted quantity is presented as an independent prediction.
full rationale
The paper's derivation chain is self-contained in the sense required here. The exposure and development models are taken from external classical references (Dill et al., 1975; Mack, 1987), not from the authors' own prior work. The unknown parameters theta and the threshold tau are fitted by minimizing a differentiable BCE objective on a calibration split of LithoBench, and the reported Pixel Difference, EPE-mean, and EPE-max are computed on the held-out test split using those fixed fitted parameters. This is standard empirical validation, not circular reasoning: no test-set quantity is equal to a training-set fit by construction, and the threshold tau is not re-fit on the test set. The self-citations that appear (LithoBench [30], FUILT [9], the differentiable lithography framework [8], and other prior group papers) are used as dataset/tool references or as context, not as the load-bearing justification for the resist model equations or for the accuracy claim. The cited LithoBench benchmark was generated with commercial tools, so the evaluation data are external to the model being proposed. The main weakness is that the abstract's 'superior accuracy and efficiency compared to existing solutions' claim is only partially supported by the experiments: Table 1 compares against two threshold-based baselines only, and Table 2 reports no baseline runtime. That is an evidence/comparison limitation, not a circularity. Overall, there is no step in which a prediction reduces to a fitted parameter, nor a load-bearing self-citation chain that forces the paper's conclusions.
Assumptions & free parameters
free parameters (5)
- Dill C (exposure rate)
- m_TH (inflection point in Mack model)
- r_max (maximum development rate)
- r_min (finite dissolution rate of unexposed resist)
- threshold tau (binarization threshold)
assumptions (5)
- domain assumption Lambert-Beer absorption law and the Dill exposure model describe the resist exposure step
- domain assumption The Mack development model with integer n and the stated inflection-point relation is accurate for the evaluated resist
- domain assumption The final resist image can be represented by thresholding the development depth image
- domain assumption Automatic differentiation through the resist simulation provides correct gradients for parameter optimization
- domain assumption The LithoBench MetalSet aerial/wafer pairs, generated by commercial tools, are a valid proxy for real resist behavior
Cite this review
Pith. "Pith review of TorchResist: Open-Source Differentiable Resist Simulator." pith.science (2026). https://pith.science/paper/OZC5HBQM
@misc{pith2026250206838,
author = {Pith},
title = {Pith review of: TorchResist: Open-Source Differentiable Resist Simulator},
year = {2026},
howpublished = {\url{https://pith.science/paper/OZC5HBQM}},
note = {Machine review of arXiv:2502.06838}
}
read the original abstract
Recent decades have witnessed remarkable advancements in artificial intelligence (AI), including large language models (LLMs), image and video generative models, and embodied AI systems. These advancements have led to an explosive increase in the demand for computational power, challenging the limits of Moore's Law. Optical lithography, a critical technology in semiconductor manufacturing, faces significant challenges due to its high costs. To address this, various lithography simulators have been developed. However, many of these simulators are limited by their inadequate photoresist modeling capabilities. This paper presents TorchResist, an open-source, differentiable photoresist simulator.TorchResist employs an analytical approach to model the photoresist process, functioning as a white-box system with at most twenty interpretable parameters. Leveraging modern differentiable programming techniques and parallel computing on GPUs, TorchResist enables seamless co-optimization with other tools across multiple related tasks. Our experimental results demonstrate that TorchResist achieves superior accuracy and efficiency compared to existing solutions. The source code is publicly available.
Figures
Reference graph
Works this paper leans on
-
[24]
Neural-network-based approach to resist modeling and opc,
Zach, F. X., “Neural-network-based approach to resist modeling and opc,” in [ Optical Microlithography XVII ], 5377, 670–679, SPIE (2004)
work page 2004
-
[1]
Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al., “Deepseek-v3 technical report,” arXiv preprint arXiv:2412.19437 (2024)
arXiv 2024
-
[2]
Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al., “Gpt-4 technical report,” arXiv preprint arXiv:2303.08774 (2023)
arXiv 2023
-
[3]
MoreauPruner: Robust Pruning of Large Language Models against Weight Perturbations
Wang, Z., Zhang, J., Zhao, W., Farnia, F., and Yu, B., “Moreaupruner: Robust pruning of large language models against weight perturbations,” arXiv preprint arXiv:2406.07017 (2024)
work page Pith review arXiv 2024
-
[4]
High-resolution image synthesis with latent diffusion models,
Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B., “High-resolution image synthesis with latent diffusion models,” in [ Proceedings of the IEEE/CVF conference on computer vision and pattern recognition], 10684–10695 (2022)
2022
-
[5]
Consistent Distributed Ranking of Generative Models via Kernel Distances
Wang, Z., Farnia, F., Lin, Z., Shen, Y., and Yu, B., “On the evaluation of generative models in distributed learning tasks,” arXiv preprint arXiv:2310.11714 (2023)
work page Pith review arXiv 2023
-
[6]
Mack, C., [ Fundamental principles of optical lithography: the science of microfabrication ], John Wiley & Sons (2007)
work page 2007
-
[7]
Iccad-2013 cad contest in mask optimization and benchmark suite,
Banerjee, S., Li, Z., and Nassif, S. R., “Iccad-2013 cad contest in mask optimization and benchmark suite,” in [2013 IEEE/ACM International Conference on Computer-Aided Design (ICCAD) ], 271–274, IEEE (2013)
work page 2013
Show all 31 references
-
[8]
Open-source differentiable lithography imaging framework,
Chen, G., Geng, H., Yu, B., and Pan, D. Z., “Open-source differentiable lithography imaging framework,” in [DTCO and Computational Patterning III ], 12954, 118–127, SPIE (2024)
2024
-
[9]
Fuilt: Full chip ilt system with boundary healing,
Yin, S., Zhao, W., Xie, L., Chen, H., Ma, Y., Ho, T.-Y., and Yu, B., “Fuilt: Full chip ilt system with boundary healing,” in [ Proceedings of the 2024 International Symposium on Physical Design ], ISPD ’24 , 13–20, Association for Computing Machinery, New York, NY, USA (2024)
2024
-
[10]
Accurate lithography simulation model based on convolutional neural networks,
Watanabe, Y., Kimura, T., Matsunawa, T., and Nojima, S., “Accurate lithography simulation model based on convolutional neural networks,” in [ Optical Microlithography XXX ], 10147, 137–145, SPIE (2017)
2017
-
[11]
F¨ uhner, T.,Artificial evolution for the optimization of lithographic process conditions, PhD thesis, Friedrich- Alexander-Universit¨ at Erlangen-N¨ urnberg (F AU) (2014)
2014
-
[12]
True process variation aware optical proximity correction with variational lithography modeling and model calibration,
Yu, P., Shi, S. X., and Pan, D. Z., “True process variation aware optical proximity correction with variational lithography modeling and model calibration,” Journal of Micro/Nanolithography, MEMS and MOEMS 6(3), 031004–031004 (2007)
2007
-
[13]
Diffpattern: Layout pattern generation via discrete diffusion,
Wang, Z., Shen, Y., Zhao, W., Bai, Y., Chen, G., Farnia, F., and Yu, B., “Diffpattern: Layout pattern generation via discrete diffusion,” in [ 2023 60th ACM/IEEE Design Automation Conference (DAC) ], 1–6, IEEE (2023)
2023
-
[14]
Chatpattern: Layout pattern cus- tomization via natural language,
Wang, Z., Shen, Y., Yao, X., Zhao, W., Bai, Y., Farnia, F., and Yu, B., “Chatpattern: Layout pattern cus- tomization via natural language,” in [ Proceedings of the 61st ACM/IEEE Design Automation Conference ], 1–6 (2024)
2024
-
[15]
Ultra-fast source mask optimization via conditional discrete diffusion,
Chen, G., Wang, Z., Yu, B., Pan, D. Z., and Wong, M. D., “Ultra-fast source mask optimization via conditional discrete diffusion,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (2024)
2024
-
[16]
Characterization of positive photoresist,
Dill, F. H., Hornberger, W. P., Hauge, P. S., and Shaw, J. M., “Characterization of positive photoresist,” IEEE Transactions on electron devices 22(7), 445–452 (1975)
1975
-
[17]
Development of positive photoresists,
Mack, C. A., “Development of positive photoresists,” Journal of the Electrochemical Society 134(1), 148 (1987)
1987
-
[18]
Fast-marching level-set methods for three-dimensional photolithography development,
Sethian, J. A., “Fast-marching level-set methods for three-dimensional photolithography development,” in [Optical Microlithography IX ], 2726, 262–272, SPIE (1996)
1996
-
[19]
Pursuit of lower critical dimensional uniformity in euv resists,
Thackeray, J., Cameron, J., Jain, V., LaBeaume, P., Coley, S., Ongayi, O., Wagner, M., Rachford, A., and Biafore, J., “Pursuit of lower critical dimensional uniformity in euv resists,” Journal of Photopolymer Science and Technology 26(5), 605–610 (2013)
2013
-
[20]
The lithographic impact of resist model parameters,
Smith, M. D., Byers, J. D., and Mack, C. A., “The lithographic impact of resist model parameters,” in [Advances in Resist Technology and Processing XXI ], 5376, 322–332, SPIE (2004)
2004
-
[21]
Chemically amplified resists resolving 25 nm 1: 1 line: space features with euv lithography,
Thackeray, J. W., Nassar, R. A., Brainard, R., Goldfarb, D., Wallow, T., Wei, Y., Mackey, J., Naulleau, P., Pierson, B., and Solak, H. H., “Chemically amplified resists resolving 25 nm 1: 1 line: space features with euv lithography,” in [Emerging Lithographic Technologies XI],...
2007
-
[22]
Materials challenges for sub-20-nm lithography,
Thackeray, J. W., “Materials challenges for sub-20-nm lithography,” Journal of Micro/Nanolithography, MEMS, and MOEMS 10(3), 033009–033009 (2011)
2011
-
[23]
Variable-threshold resist models for lithography simulation,
Randall, J., Ronse, K. G., Marschner, T., Goethals, A.-M., and Ercken, M., “Variable-threshold resist models for lithography simulation,” in [ Optical Microlithography XII ], 3679, 176–182, SPIE (1999)
1999
-
[25]
TensorFlow: Large-scale machine learning on heterogeneous systems,
Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, G. S., Davis, A., Dean, J., Devin, M., Ghemawat, S., Goodfellow, I., Harp, A., Irving, G., Isard, M., Jia, Y., Jozefowicz, R., Kaiser, L., Kudlur, M., Levenberg, J., Man´ e, D., Monga, R., Moore, S.,...
2015
-
[26]
Pytorch: An imperative style, high-performance deep learning library,
Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al., “Pytorch: An imperative style, high-performance deep learning library,” Advances in neural information processing systems 32 (2019)
2019
-
[27]
JAX: composable transformations of Python+NumPy programs,
Bradbury, J., Frostig, R., Hawkins, P., Johnson, M. J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., Van- derPlas, J., Wanderman-Milne, S., and Zhang, Q., “JAX: composable transformations of Python+NumPy programs,” (2018)
2018
-
[28]
Level set methods: an overview and some recent results,
Osher, S. and Fedkiw, R. P., “Level set methods: an overview and some recent results,” Journal of Com- putational physics 169(2), 463–502 (2001)
2001
-
[29]
A fast path-based method for 3-d resist development simulation,
Dai, Q., Guo, R., Lee, S.-Y., Choi, J., Lee, S.-H., Shin, I.-K., Jeon, C.-U., Kim, B.-G., and Cho, H.-K., “A fast path-based method for 3-d resist development simulation,” Microelectronic engineering 127, 86–96 (2014)
2014
-
[30]
Lithobench: Benchmarking ai computational lithog- raphy for semiconductor manufacturing,
Zheng, S., Yang, H., Zhu, B., Yu, B., and Wong, M., “Lithobench: Benchmarking ai computational lithog- raphy for semiconductor manufacturing,” Advances in Neural Information Processing Systems 36 (2024)
2024
-
[31]
Adam: A method for stochastic optimization,
Kingma, D. P., “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 (2014)
2014 arXiv
Reviewed August 9, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.