Reinforcement learning for inverse silhouette design and rapid laser cutting of compact parallelogram quad kirigami prototypes

The gist The RL-Kirigami framework combines an optimal-transport conditional flow matching prior with Group Relative Policy Optimization to perform inverse design for compact reconfigurable parallelogram quad kirigami structures under hard geometric feasibility constraints, leading to rapid laser cutting prototypes.

Inverse Design Challenge

Kirigami inverse design remains a bottleneck because feasible cut layouts must satisfy discrete compatibility rules, avoid overlap, and map one target shape to valid designs <ref:2605.08098#pg2> The problem is complicated by the fact that the deployed response of the full structure arises from many coupled, almost rigid panels connected by thin ligaments, which can exhibit elastic instabilities as well as strong geometric nonlinearity <ref:2605.08098#pg3> Classical inverse design approaches typically rely on search strategies such as topology optimization or evolutionary algorithms and use repeated forward simulations to evaluate candidate patterns, which leads to a high compute cost per target <ref:2605.08098#pg4>. Data-driven generative models offer an alternative where models can sample cut patterns conditioned on a prescribed target without iterative search at inference time, but learning kirigami constraints remains challenging because cut rules are discrete and relational <ref:2605.08098#pg5>.

RL-Kirigami Framework

RL-Kirigami combines an optimal-transport conditional flow matching (OT-CFM) generator with reinforcement learning fine-tuning under rewards from the forward simulator and evaluator <ref:2605.08098#pg6>. The OT-CFM generator is treated as a stochastic policy over kirigami cut patterns conditioned on the target silhouette, which is then refined in a closed loop using reinforcement learning <ref:2605.08098#pg7>. The rewards combine kirigami validity rules, overlap penalties, and evaluator terms such as shape error and ratio-field regularity computed after forward simulation <ref:2605.08098#pg8>. Group Relative Policy Optimization (GRPO) is used to optimize this nondifferentiable black-box reward, aligning the generator with nondifferentiable task rewards for silhouette matching, feasibility, and ratio-field regularity <ref:2605.08098#pg8>.

Model Components and Constraints

The kirigami pattern is modeled as an m × n array of voids where each void is parameterized by an interior deployment angle phi ij and a ratio of side lengths x ij <ref:2605.08098#pg2>. Global compatibility is enforced by decoding in a marching order from left to right and top to bottom, reusing vertices that were already created by earlier neighbors <ref:2605.08098#pg5>. The pipeline includes nondifferentiable geometric checks that act as hard feasibility filters, such as an invalid void count Ninv for collapsed or self intersecting voids, and an overlap ratio rov <ref:2605.08098#pg5>.

Experimental Results and Performance

The OT-CFM prior reached 94.2% sIoU and outperformed solver baselines, reducing forward simulator evaluations from hundreds to 1 <ref:2605.08098#pg8>. GRPO improved accuracy to 94.91% sIoU and reduced the total variation TV(x) from 0.95 to 0.81 while maintaining a high sIoU <ref:2605.08098#pg8>. The resulting layouts were exported to DXF and laser cut in 50 μm polymeric sheets, producing deployable prototypes in 8.0 ± 1.0 min per part <ref:2605.08098#pg8>.

Comparative Analysis

The OT-CFM prior achieved a sIoU of 94.2% and a success rate of 98.7% with only one forward evaluation per target, significantly outperforming solver baselines which required hundreds to 1000 evaluations <ref:2605.08098#pg8>. The OT-CFM prior was shown to be competitive even at very low sampling-step counts, unlike diffusion which required a much longer sampling time for comparable performance <ref:2605.08098#pg8>. Furthermore, the accuracy-only RL reward increased sIoU while keeping TV(x) near the OT-CFM prior, while the combined reward preserved high sIoU at 94.83% and lowered TV(x) to 0.81 <ref:2605.08098#pg8>.

Fabrication Demonstration

Representative targets like a heart, a star, and a hexagon were fabricated on polyamide (PA, 50 μm thick) sheets using laser cutters, with the full end-to-end workflow taking 8.0 ± 1.0 min per prototype <ref:2605.08098#pg8>. The laser-cut polyamide patterns successfully transformed into the target shapes, demonstrating that successful deployment depends jointly on cut topology, connector geometry, sheet thickness, and material compliance <ref:2605.08098#pg8>.

Limitations

The study focused on compact reconfigurable parallelogram quad kirigami with fixed rectangular boundary anchors and single-state silhouette matching at one fixed deployment parameter phi <ref:2605.08098#pg8>. The geometry-based forward simulator does not model forces, material nonlinearity, contact along the full deployment path, or fabrication tolerances <ref:2605.08098#pg8>. Therefore, the reported results should be read as strong evidence for fast inverse design within this constrained geometry class and simulator <ref:2605.08098#pg8>.

Conclusion

RL-Kirigami successfully combines an OT-CFM prior with GRPO fine-tuning to achieve a manufacturing-aware inverse design workflow for deployable kirigami metamaterials under hard geometric feasibility constraints <ref:2605.08098#pg8>. The pretrained prior provided strong single-sample solutions, and GRPO effectively aligned the generator to nondifferentiable task rewards and design preferences <ref:2605.08098#pg8>. Ultimately, the framework demonstrates a practical route for inverse kirigami design where fast conditional sampling handles the one-to-many inverse map, while RL supports preference alignment under discrete feasibility rules <ref:2605.08098#pg8>.

References

Albergo, M., Vanden-Eijnden, E., 2023. Building normalizing flows with stochastic interpolants. International Conference on Learning Representations <ref:2605.08098#pg2>

Alderete, N.A., Pathak, N., Espinosa, H.D., 2022. Machine learning assisted design of shape-programmable 3d kirigami metamaterials npj Computational Materials 8, 191 <ref:2605.08098#pg2>

Babaee, S., Shi, Y., Abbasalizadeh, S., Tamang, S., Hess, K., Collins, J.E., Ishida, K., Lopes, A., Williams, M., Albaghdadi, M., et al. 2021 Kirigami-inspired stents for sustained local delivery of therapeutics Nature Materials 20, 1085–1092 <ref:2605.08098#pg2>

Bastek, J.H., Kochmann, D.M., 23 Inverse design of nonlinear mechanical metamaterials via video denoising diffusion models Nature Machine Intelligence 5, 1466–1475 <ref:2605.08098#pg2>

Bertoldi, K., Vitelli, V., Christensen, J., Van Hecke, M., 23 Flexible mechanical metamaterials Nature Reviews Materials 2, 1–11 <ref:2605.08098#pg2>

Blees, M.K., Barnard, A.W., Rose, P.A., Roberts, S.P., McGill, K.L., Huang, P.Y., Ruyack, A.R., Kevek, J.W., Kobrin, B., Muller, D.A., et al. 2015 Graphene kirigami Nature 524, 204–207 <ref:2605.08098#pg2>

Bliah, O., Hegde, C., Tan, J.M.R., Magdassi, S., 25 Fabrication of soft robotics by additive manufacturing: from materials to applications Chemical Reviews 125, 7275–7320 <ref:2605.08098#pg2>

Boender, C.G.E., Rinnooy Kan, A., Timmer, G., Stougie, L., 1982 A stochastic method for global optimization Mathematical programming 22, 125–140 <ref:2605.08098#pg2>

Branyan, C., Rafsanjani, A., Bertoldi, K., Hatton, R.L., Mengüç, Y., 2022 Curvilinear kirigami skins let soft bending actuators slither faster Frontiers in Robotics and AI 9, 872007 <ref:2605.

1. Improved generative prior learning using Optimal Transport Conditional Flow Matching (OT-CFM) allows for fast conditional sampling, which is critical because decoding and feasibility checks are required for each candidate. This enables a rapid inverse design pipeline by generating feasible designs without requiring iterative search at inference time.

2. Integration of Group Relative Policy Optimization (GRPO) fine-tuning with rewards that include both silhouette matching and ratio-field regularity aligns the generator with nondifferentiable rewards for silhouette matching, feasibility, and ratio-field regularity. This allows the RL system to steer the model toward specific design preferences, such as achieving lower TV(x) or a high sIoU.

3. The RL framework enables preference alignment by optimizing augmented rewards where accuracy-only GRPO gives the highest sIoU at 94.91%, while its TV(x) is 0.981. This demonstrates that RL can be used to shift the sampling distribution toward better task scores or higher ratio-field regularity, depending on the reward, without requiring a differentiable surrogate.

4. The system supports an end-to-end manufacturing workflow where generated patterns are exported to DXF and laser cut in situ, with reported results showing that the generated layouts can be translated deterministically and directly to a DXF file and immediately passed to the laser cutting workflow rather than being redrawn by hand. This proves the utility of the design for rapid laser cutting of kirigami prototypes.

Kirigami is an increasingly useful fabrication method to produce shape-programmable metamaterial structures. However, inverse design remains difficult because deployment is nonlinear, and feasible cut layouts must satisfy discrete compatibility rules, avoid overlap, and map one target shape to valid designs. We present RL-Kirigami, an inverse design framework that combines optimal-transport conditional flow matching (OT-CFM) with reinforcement learning to generate compatible ratio fields for compact reconfigurable parallelogram quad kirigami. A marching decoder enforces global geometric compatibility, and Group Relative Policy Optimization (GRPO) aligns the generator with nondifferentiable rewards for silhouette matching, feasibility, and ratio-field regularity. Across procedurally generated target shape instances, a single sample from the pretrained OT-CFM prior reached 94.2% sIoU and outperformed solver baselines while reducing forward simulator evaluations from hundreds to 1. GRPO improved accuracy to 94.91% sIoU and, with regularity included, reduced TV(x) from 0.95 to 0.81 while maintaining 94.83% sIoU. Generated layouts were exported to DXF and laser cut in 50 μ m polymeric sheets to produce deployable prototypes in 8.0 plus or minus 1.0 minutes per part. These results support a manufacturing-aware inverse design workflow for deployable kirigami metamaterials under geometric feasibility constraints. Code is available at https://github.com/milad1378yz/RL-Kirigami.