Source-linked AI summary
SHRAV: State-Hypothesis-Reason-Action-Verify Framework for Physical Modeling and Inverse Design
Ziheng Guo, Yang Bu
TL;DR
Physical modeling and inverse design require computation that can continue from reusable state. SHRAV organizes this continuation around explicit roles for learned evolution and numerical quantities, with inverse configurations incorporating target-directed changes and evaluator feedback. In computational lithography, four fixed-weight updates improve thresholded aerial-image IoU under independent replay.
Problem
Physical modeling and inverse design require computation that can continue from reusable state while relating internal computation, physical changes, and evaluated consequences.
Source: §1 · p. 1Method
SHRAV defines a continuation core with declared reuse boundaries and learned/numerical roles; inverse configurations additionally use retained state, target-directed mask updates, and evaluator feedback.
Source: §1 · p. 1, §2.3 · p. 3, §2.2 · p. 3Results
Four fixed-weight design updates improve independent-replay thresholded-image IoU from 0.5313 to 0.8153, with a maximum absolute prediction-replay difference of 0.000824.
Source: §2.3 · p. 3, §4 · p. 5Takeaways & Limitations
The shared interface connects reusable learned computation with physical assessment without binding the framework to one network or domain.
Source: §6 · p. 7Independent scalar-pupil replay shares nominal optical assumptions and is not a wafer measurement.
Source: §4 · p. 5
Abstract
from arXiv · showhide
Physical modeling and inverse design require computation that can continue from reusable state. We introduce SHRAV, an architecture-independent computational framework organized around State, Hypothesis, Reason, Action, and Verify. Its central mechanism is a state-continuation core with declared reuse boundaries and explicit roles for learned evolution and numerical quantities. Forward configurations evolve predictive state and read out physical responses; inverse-design configurations additionally generate target-directed modifications and consume evaluator feedback. Electromagnetic world-model studies are mapped to forward configurations, with selected readout and reuse diagnostics reported here. Computational lithography demonstrates an inverse-design configuration: four fixed-weight design updates improve thresholded aerial-image intersection-over-union from 0.5313 to 0.8153 under independent scalar-pupil replay, with a maximum absolute IoU difference of approximately 0.000824 between predictor estimates and independent replay.
1. From physical prediction to design evolution
SHRAV frames physical design as continuing computation: reusable state, learned evolution, numerical quantities, design changes, and evaluated consequences meet through explicit interfaces. It is intended to support inverse design while also accommodating forward configurations.
1. From physical prediction to design evolution
- SHRAV treats physical computation, including design, as an evolving process in which reusable state and evaluated consequences inform subsequent computation.The framework asks what persists, what is computed internally, what changes physically, and how evaluated consequences enter the next step.Source: §1 · p. 1
- The framework separates computational roles: learned latent evolution need not reproduce a complete numerical solution, and numerical observations can enter continuation through an explicit interface.This separation does not assert statistical independence or guarantee transfer across conditions.Source: §1 · p. 1
- FieldSeer I [3] maps to a forward configuration with externally supplied geometry interventions, reporting physical-domain field MSE of 0.01122 after 180 observed and 100 predicted frames.SHRAV is intended to support inverse design, while the framework itself is not specific to lithography.Source: §1 · p. 1
- Reason and Verify assessment enter the state-continuation update U, and the updated design and internal state initialize the next cycle.Feedback from one update is an input to the next; an explicit acceptance gate is not required.Source: §1 · p. 2
2. Shared core and configuration contracts
SHRAV defines a shared continuation core for forward and inverse configurations, then adds requirements for target-directed design and evaluator feedback. Its lithography implementation reuses retained latent state across executed mask updates under fixed target and optical conditions.
2.1 Shared continuation core
- The shared core requires persistent state with a declared reuse boundary, distinct learned-evolution and numerical roles, and a learned component that evolves or recontextualizes retained state.A purely numerical solver alone is not a SHRAV configuration; the state need not be a complete or proven sufficient Markov state.Source: §2.1 · p. 2
2.2 Inverse-design roles and criterion
- Inverse-design configurations select and execute target-directed modifications and consume evaluated consequences in subsequent computation; feedback alone does not make a predictor an inverse designer.The configuration also declares internal hypothesis development, evaluator criteria, and feedback consumed by U; reporting-only scores do not qualify.Source: §2.1 · p. 2, §2.2 · p. 3
2.3 Implemented inverse feedback
- The lithography exemplar reuses retained latent state across mask updates at fixed target and optical conditions, with four internal Reason iterations per outer update and no intervening optical evaluation.All candidates use fixed weights without an acceptance gate; independent scalar-pupil replay audits saved masks separately.Source: §2.3 · p. 3
3. Forward configurations of the interface
The forward configurations illustrate numerical readout refinement that leaves latent evolution unchanged, alongside distinct latent-numerical organizations. The reported diagnostics are selected examples, and state reuse is limited to declared conditions.
3. Forward configurations of the interface
- Changed histories or unvalidated conditions require reinitialization, so the examples do not establish unrestricted state transfer.The forward examples continue under fixed geometry and source conditions, or fixed mask and optics.Source: §3 · p. 4
- 100/100 saved trajectories show decreased whole-rollout coupled-field MSE, with a median relative reduction of 30.9%, while corrected fields do not feed the recurrent transition.The diagnostic assimilates 32 frames and predicts 80; the representative trajectory is the median-ranked case among 100.Source: Figure 2 · p. 4, §3 · p. 4
- Figure 3 contrasts latent and hybrid lithographic propagation; the hybrid model carries an explicit numerical propagation state, illustrating distinct latent-numerical organizations rather than isolating a single cause.The saved outputs come from separately trained implementations under identical display scales.Source: Figure 3 · p. 4, §3 · pp. 4–5
4. Numerical exemplar: computational lithography
Figure 4 presents a selected lithography validation trajectory with a fixed 0.225 threshold and independent numerical replay. IoU improves across four updates, while contour accuracy favors an earlier update.
4. Numerical exemplar: computational lithography
- The fixed 0.225 intensity threshold is used consistently in feedback construction and numerical assessment; thresholded-image IoU measures foreground-pixel intersection over union.Figure 4 shows target, masks, predicted focal intensities, thresholded images, and feedback discrepancies; columns 0–3 feed the next update, while column 4 shows terminal discrepancies.Source: §4 · p. 5
- 0.5313 to 0.8153: independent IoU gains 0.2840 across four updates, with a maximum absolute prediction-replay difference of 0.000824.Overlap favors update 4, whereas contour accuracy favors update 2; these preferences do not imply an executed rollback.Source: Figure 4 · p. 5, §4 · p. 5
- 21.34, 13.93, 9.53, 11.96, 15.00 nm: independent full-contour EPE P95 across updates, favoring update 2 over update 4.Independent replay shares nominal optical assumptions and is not a wafer measurement.Source: §4 · p. 5
5. Positioning, domains, and progression
SHRAV positions itself as a configuration contract that makes reuse boundaries and computational roles explicit, and maps physical-modeling examples to forward and inverse configurations. The paper distinguishes capability levels and reserves detailed self-adaptation mechanisms and evaluations for companion studies.
5. Positioning, domains, and progression
- SHRAV makes reuse boundaries and latent/numerical roles explicit, while learned optimizers, latent world models, and planners supply related ingredients rather than being excluded by the contract.Adjoint methods can provide sensitivities or local refinement.Source: §5 · p. 6
5.1 Domain mappings and evidence
- The evidence index maps FieldSeer I to a forward configuration and reports cached-versus-uncached output equivalence in 30 edited cases, with maximum absolute difference 0.This establishes execution equivalence, not an independent physical-accuracy test; Table 1 distinguishes configuration mappings from evidence status and says illustrative inverse mappings are not validated implementations.Source: §5.1 · p. 6, Table 1 · p. 6
- Two companion directions are role-attribution studies across physical domains and experimental-feedback adaptation under shifted process conditions; their mechanisms, protocols, and limitations are reserved for later publications.Source: §5.1 · p. 6
5.2 Capability levels and scope
- Capability levels are cumulative: Level-0 establishes fixed-weight continuation, Level-1 adds learned-operator reuse across a task family, and Level-2 adds parameter adaptation verified against declared criteria before adoption.AI-native inverse design means target-directed processes jointly conditioned by reusable learned state and evaluator feedback; detailed Level-2 mechanisms and evaluations are reserved for companion studies.Source: §5.2 · pp. 6–7
6. Conclusion
SHRAV specifies a continuation core with explicit reuse boundaries and distinct roles for learned evolution and numerical quantities. Its shared interface supports forward prediction and readout, as well as inverse-design actions informed by evaluator feedback.
6. Conclusion
- SHRAV specifies a continuation core with explicit reuse boundaries and distinct roles for learned evolution and numerical quantities.Source: §6 · p. 7
- Forward configurations expose predictive evolution and physical-response readout through the shared interface.Source: §6 · p. 7
- Inverse-design configurations additionally execute target-directed changes and consume evaluation feedback through that interface.Source: §6 · p. 7