Making chip features smaller than the wavelength of the light that prints them requires a deliberately distorted photomask, a stencil whose odd shapes cancel out the blur that diffraction adds. Inverse lithography computes that mask by running a physics model of the optics backwards. It meets two goals that pull against each other. The printed pattern should match the intended layout closely, measured as edge placement error (how far printed edges land from intended edges) and process variation band (how much the print moves under focus and dose drift). The mask should also be cheap to write. Masks are drawn by an electron beam that exposes one rectangle at a time, so cost scales with the shot count, the number of rectangles needed. Highly faithful masks are curvy and need many shots. Simple masks print poorly.
Current learned mask generators commit to one point on this trade-off before training. Moving to another point means retraining, and how much write cost a given layer can tolerate changes with the mask shop, the layer's criticality, and wafer volume. We propose to train a single generator that takes a trade-off preference as an input and returns, in about a tenth of a second, a mask near the best achievable compromise for that preference. The model is trained with a non-linear scalarization that can reach compromise points a weighted sum cannot reach. We also treat as an open empirical question whether such points exist in lithography at all.
Core Claim
The closest existing system is LithoGRPO. It pretrains a flow-matching mask generator, fine-tunes it through a differentiable optics simulator, and then applies Group Relative Policy Optimization (GRPO). GRPO is a policy-gradient method that scores a group of sampled outputs against each other. LithoGRPO uses it to optimize a fixed linear weighting of fidelity and shot count, computed with a fast integer-programming shot counter that runs in 0.2 s and preserves the ranking of candidate masks. Its authors state that navigating the fidelity–shot-count Pareto frontier "will continue to be a central challenge." Changing their weights means another RL run.
Our claim has two parts. First, a generator conditioned on a preference vector and trained with randomly drawn preferences can cover the whole frontier. Its hypervolume, the area of objective space it dominates, should match or exceed that of a set of separately trained fixed-weight models at equal total compute. Second, the scalarization used inside the group-relative reward matters exactly in proportion to how non-convex the lithography frontier is. We expect it to be non-convex for a structural reason: shot count is an integer that jumps when the fracturing has to add a rectangle for a new jog or serif, while fidelity improves smoothly. This produces "dents" in the frontier. Optimal Scalarizations for Sublinear Hypervolume Regret shows that a weighted sum cannot reach points inside such dents. The same paper shows that hypervolume scalarization with uniformly random weights achieves hypervolume regret that matches the lower bound for any frontier shape.
What would be new is a preference-conditioned, frontier-covering mask generator, together with the first measurement of how much of the lithography frontier is unreachable by weighted sums. If the frontier turns out to be essentially convex, the second part of the claim fails cleanly. The conditioning contribution survives without it.
Research Program
We start from a LithoGRPO-style pipeline built on LithoBench data and the TorchLitho differentiable simulator. TorchLitho supplies Abbe/Hopkins imaging with adjoint gradients and reports agreement with commercial ground truth of 99.45% mean pixel accuracy. We add a preference input $\lambda$, a positive unit vector over the objectives. It enters the flow-matching network through feature-wise modulation in every block. We also reuse LithoGRPO's low-frequency colored exploration noise, which keeps masks from fragmenting.
For each training layout, we draw $\lambda$ uniformly from the positive unit sphere and sample a group of masks. We map each mask's costs (edge placement error, process variation band, shot count) to a gain vector $y \in \mathbb{R}^k_{+}$ by subtracting them from a fixed reference point and normalizing by per-objective ranges estimated on the pretrained model's outputs. Each mask is then scored by
$$s_\lambda(y) = \min_{i=1,\dots,k} \left( y_i / \lambda_i \right)^k,$$
where $k$ is the number of objectives. This is the hypervolume scalarization from Zhang's paper. Group-relative advantages are computed from $s_\lambda$ within the group, which shares a single $\lambda$.
The min operator rewards a mask for improving its worst objective relative to the requested direction, so it cannot trade an unreachable dent for a supported vertex. Normalizing gains also addresses the failure that PD-MORL identifies, where objectives with large numeric ranges swamp small ones. Here the large-range objective is shot count, in the thousands, and the small one is nanometer edge error. PD-MORL also supplies the precedent that one preference-conditioned network can replace a family of policies with an order of magnitude fewer parameters (0.34M vs 2.8M on Walker).
We will also support a budget mode, in which the user specifies a shot-count cap rather than a direction. Following the zero-duality-gap result of Paternain et al., we train the cap as a conditioning input with a learned Lagrange multiplier per budget bin. Mask shops naturally state requirements this way, so we report both modes.
The analysis component is a frontier atlas. For a few hundred held-out tiles, we pool every mask produced by every method and every sampled preference. We compute the empirical non-dominated set with exact (slow) fracturing rather than the fast counter, and we measure the fraction of non-dominated points that lie strictly inside the convex hull's lower envelope, meaning points no weighted sum can select. We report this fraction separately for MetalSet, ViaSet, and the out-of-distribution standard-cell sets, because dense metal and isolated vias may differ.
Baselines
The naive reference is the pretrained-plus-simulator-fine-tuned generator with no RL stage. It shows how far the frontier moves from its default output. A second reference is the classic numerical inverse lithography solver. ILILT, an implicit fixed-point learner, reports average edge placement error of 0.08 against that solver's 0.21, with a 10× speedup. We include ILILT as a fidelity-only point, since it optimizes no write-cost objective.
The strongest competitor is LithoGRPO itself. The file states it achieved the lowest average rank across metrics against traditional and learned baselines, with roughly 0.1 s inference. No reported result was found on Lacuna for LithoGRPO's absolute edge placement error, band, or shot count values, so we will rerun it. To make the comparison fair, we train $K$ fixed-weight LithoGRPO models with weights evenly spaced in angle. Their combined RL fine-tuning GPU-hours equal our single conditioned model's budget, and we also run $K$ at full per-model budget as an upper reference. Every method uses the same pretrained checkpoint, simulator, shot counter, and data split.
Beating LithoGRPO means two things. Our model must achieve higher hypervolume on held-out tiles at matched compute, and it must not lose more than the run-to-run variance at any of the $K$ preferences LithoGRPO was trained on.
TokMan generates Manhattan polygons token by token with a diffusion transformer. It reports over 20% edge placement improvement over prior methods and 4–7× shot reduction on ICCAD 2013. It is the strongest low-shot-count anchor, and we compare at its single operating point. Beating it means producing a mask that dominates or matches its point at our corresponding preference.
The key ablation keeps everything in our model but replaces $s_\lambda$ with the linear score $\sum_i \lambda_i y_i$. This isolates the scalarization. A second ablation removes conditioning and uses random $\lambda$ only during training, to separate diversity effects from controllability.
Why not just…?
Why not just sweep the weight in LithoGRPO's linear reward? This is a real threat. If the frontier is convex, a sweep of $K$ models recovers it, apart from the $K$-fold cost and the inability to serve unseen preferences. Zhang's result shows a weighted sum misses dent points, but only if dents exist. The frontier atlas measures exactly this, and the linear-conditioned ablation shows whether the scalarization or the conditioning carries the gain. If the non-supported fraction is near zero, we will report the contribution as a compute and controllability gain, not a frontier-quality gain.
Why not just sample many masks from one model and keep the non-dominated ones? LithoGRPO's exploration noise already produces diverse candidates, and filtering is cheap with a 0.2 s counter. This is the simplest frontier method and could plausibly come close. We run it explicitly at matched wall-clock, giving it the inference time our model saves. Our expectation is that samples cluster near the trained weighting and leave the frontier extremes empty. Zhang's BBOB experiments make a related observation, finding that random-weight hypervolume scalarization reached the frontier's extreme ends better than an expected-hypervolume-improvement method. If best-of-N matches our hypervolume, the conditioned model's value reduces to latency.
Why not just state a shot budget and solve the constrained problem? Paternain et al. show such constrained problems have no duality gap, so a Lagrangian method is sound. It still fixes one budget per run, and picking the budget requires knowing where the trade-off bends. We adopt this formulation as our budget mode rather than arguing against it. The comparison is one budget-conditioned model against separately trained single-budget Lagrangian runs, scored on edge placement error at each budget.
Why not just restrict masks to Manhattan geometry, as TokMan does? That removes shot explosion by construction, but TokMan itself notes it cannot freely place sub-resolution assist features, the small non-printing shapes that widen the process window around isolated features. Our frontier measures what that restriction gives up: the gap in process variation band between TokMan's point and our high-fidelity end, evaluated on isolated-via ViaSet tiles where assist features matter most.
Proposed Experiments
Training uses LithoBench MetalSet and ViaSet tiles of 2 µm × 2 µm at 1 nm per pixel. Testing covers held-out tiles from both sets, StdMetal and StdContact for out-of-distribution generalization, and ICCAD 2013 for comparability with TokMan and TorchLitho. Metrics are edge placement error, process variation band, exact shot count on the final masks, inference time, and hypervolume over normalized objectives from a common reference point. We also measure controllability as the Spearman correlation between the requested preference angle and achieved shot count, and report budget attainment, the edge placement error at a requested cap together with the cap-violation rate.
The decisive checks come in order. The frontier atlas is run first, because its non-supported fraction predicts whether the hypervolume scalarization should beat its linear ablation. A non-supported fraction of a few percent with no ablation gap would falsify the second claim. The matched-compute comparison against the $K$ linear models and against best-of-N filtering decides the first claim. Out-of-distribution hypervolume on the standard-cell sets checks that conditioning does not trade generalization for coverage.
One known risk is that the fast shot counter's ranking fidelity may degrade near low counts. We test this by checking rank agreement against exact fracturing on 500 masks per frontier region before trusting any reward-driven result.
References
- LithoGRPO: Fast Inverse Lithography via GRPO Reinforced Flow Matching
- Optimal Scalarizations for Sublinear Hypervolume Regret
- LithoBench: Benchmarking AI Computational Lithography for Semiconductor Manufacturing
- Open-Source Differentiable Lithography Imaging Framework
- PD-MORL: Preference-Driven Multi-Objective Reinforcement Learning Algorithm
- Constrained Reinforcement Learning Has Zero Duality Gap
- ILILT: Implicit Learning of Inverse Lithography Technologies
- TokMan:Tokenize Manhattan Mask Optimization for Inverse Lithography