Shaoliang Yang

Inverse design of cellular metamaterials is one-to-many: many geometries realize the same target properties, but conditional generators tend to collapse onto one. Style-IH-GAN uses a style-modulated, property-conditioned generator with mode-seeking and property-consistency losses to produce genuinely different TPMS unit cells for a fixed target, and emits spatially graded fields in a single batched pass with trilinear interpolation for continuous inter-cell geometry.

• 2.2× higher shape-parameter diversity and 19% lower surrogate Young’s-modulus error than a vanilla conditional GAN
• Fidelity checked by numerical homogenization over 750 generated samples across 15 targets
• A 300-cell graded field is generated in sub-millisecond time
• Stated limitation: accuracy is strongest inside the training manifold; boundary and extrapolated-density targets need screening

@article{yang2026styleihgan,
  title={Style-IH-GAN: diverse inverse homogenization and continuous graded synthesis of cellular metamaterials},
  volume={42},
  ISSN={1435-5663},
  url={http://dx.doi.org/10.1007/s00366-026-02393-3},
  DOI={10.1007/s00366-026-02393-3},
  number={5},
  journal={Engineering with Computers},
  publisher={Springer Science and Business Media LLC},
  author={Yang, Shaoliang and Wang, Jun},
  year={2026}, month=Aug
}

Aperiodic architected cellular materials trade directional performance for isotropy and robustness, but introduce geometric uncertainty that complicates validation and manufacturing. CH-WGAN learns how aperiodic variants depart from a periodic parent: a parameter generator maps nominal controls, an aperiodicity code and noise into a convex mixture of analytic signed distance functions whose scale and periodicity distributions model the uncertainty explicitly.

• Conditional hierarchical Wasserstein GAN with an InfoGAN-style Q-head for invertibility and multimodal coverage
• A periodicalization module warps the 3D grid according to learned periodicity distributions
• Generates multiple structurally diverse aperiodic variants for one periodic parent; learned uncertainty matches real aperiodicity statistics

@article{yang2026chwgan,
  title={CH-WGAN: Uncertainty-Aware Generative Design of Three-Dimensional Aperiodic Architected Cellular Materials Via Signed Distance Fields},
  volume={148},
  ISSN={1528-9001},
  url={http://dx.doi.org/10.1115/1.4071593},
  DOI={10.1115/1.4071593},
  number={12},
  journal={Journal of Mechanical Design},
  publisher={ASME International},
  author={Yang, Shaoliang and Wang, Jun},
  year={2026}, month=May
}

A large language model acts as a state-conditioned control layer for SIMP topology optimization: every k-th iteration it reads a structured solver-state observation and sets the penalization exponent, projection sharpness, filter radius and move limit, instead of following a fixed open-loop continuation schedule. A hard grayness gate reduces premature sharpening.

• In single-run 2-D comparisons, 5.9–7.4% lower final compliance than the no-tail fixed baseline, and better than the expert, three-field, schedule-only and grayness-rule controllers
• A strong fixed-plus-tail control is slightly better (0.4–1.1%), so the paper frames the contribution as a control layer that can execute and help discover competitive strategies, not superior numerical performance
• Ablation, gate, budget, replay and model-sensitivity checks for attribution and reproducibility

@article{yang2026llmcontrollers,
  title={Large language models as optimization controllers: Adaptive continuation for SIMP topology optimization},
  volume={223},
  ISSN={0965-9978},
  url={http://dx.doi.org/10.1016/j.advengsoft.2026.104304},
  DOI={10.1016/j.advengsoft.2026.104304},
  journal={Advances in Engineering Software},
  publisher={Elsevier BV},
  author={Yang, Shaoliang and Wang, Jun and Wang, Yunsheng},
  year={2026}, month=Dec,
  pages={104304}
}

Just as a Fourier transform decomposes a signal into sine waves, this work decomposes complex periodic cellular surfaces into weighted mixtures of analytic primitives. A 3D convolutional encoder maps a signed-distance volume to physically interpretable parameters: fill fractions, periodicity exponents and soft mixture weights over a library of candidate functions.

• About 636× faster than direct parameter optimization, enabling real-time design exploration
• High-fidelity reconstruction on four datasets of synthetic cellular surfaces and TPMS

@article{yang2026neuralbasis,
  title={Interpretable neural basis approximation for periodic cellular surfaces through analytic primitive mixtures},
  volume={196},
  ISSN={0010-4485},
  url={http://dx.doi.org/10.1016/j.cad.2026.104064},
  DOI={10.1016/j.cad.2026.104064},
  journal={Computer-Aided Design},
  publisher={Elsevier BV},
  author={Yang, Shaoliang and Wang, Jun},
  year={2026}, month=July,
  pages={104064}
}

One encoder compresses a 2D curve into a latent space; three decoders infer its Bézier, B-spline and NURBS parameters, and basis-function layers reconstruct smooth closed curves such as airfoils and Superformula shapes. The inferred control points and weights are interpretable and can be edited directly.

• Three classical parameterizations unified in one network, without adversarial training
• An interactive GUI lets a designer adjust inferred control points and weights in real time

@article{yang2025tpdlf,
  title={Triple-parametric autoencoder for 2D reparameterization via Bézier, B-spline, and NURBS representations},
  volume={189},
  ISSN={0010-4485},
  url={http://dx.doi.org/10.1016/j.cad.2025.103936},
  DOI={10.1016/j.cad.2025.103936},
  journal={Computer-Aided Design},
  publisher={Elsevier BV},
  author={Yang, Shaoliang and Wang, Jun},
  year={2025}, month=Dec,
  pages={103936}
}

NURBS-OT combines optimal transport with NURBS to generate free-form curves, such as aerodynamic and hydrodynamic profiles, directly in a CAD-native parametric form rather than through a large GAN.

• 0.37M parameters versus 5.05M for BézierGAN
• Evaluated on the UIUC airfoil and Superformula datasets

@article{yang2025nurbsot,
  title={NURBS-OT: An Advanced Model for Generative Curve Modeling},
  volume={147},
  ISSN={1528-9001},
  url={http://dx.doi.org/10.1115/1.4066549},
  DOI={10.1115/1.4066549},
  number={3},
  journal={Journal of Mechanical Design},
  publisher={ASME International},
  author={Yang, Shaoliang and Wang, Jun and Wang, Kang},
  year={2024}, month=Oct
}

A lightweight low-light enhancement network (ULENet) with about 0.02M parameters, from M.Eng. thesis work at Yunnan University.

@article{yang2024ulenet,
  title={Efficient low-light image enhancement with model parameters scaled down to 0.02M},
  volume={15},
  ISSN={1868-808X},
  url={http://dx.doi.org/10.1007/s13042-023-01983-7},
  DOI={10.1007/s13042-023-01983-7},
  number={4},
  journal={International Journal of Machine Learning and Cybernetics},
  publisher={Springer Science and Business Media LLC},
  author={Yang, Shaoliang and Zhou, Dongming},
  year={2023}, month=Nov,
  pages={1575–1589}
}

LightingNet addresses noise, chromatic aberration and detail distortion in low-light images by combining a vision-transformer enhancement sub-network, which supplies local high-level features, with a complementary learning sub-network that supplies global fine-tuned features through transfer learning.

@article{yang2023lightingnet,
  title={LightingNet: An Integrated Learning Method for Low-Light Image Enhancement},
  volume={9},
  ISSN={2573-0436},
  url={http://dx.doi.org/10.1109/tci.2023.3240087},
  DOI={10.1109/tci.2023.3240087},
  journal={IEEE Transactions on Computational Imaging},
  publisher={Institute of Electrical and Electronics Engineers (IEEE)},
  author={Yang, Shaoliang and Zhou, Dongming and Cao, Jinde and Guo, Yanbu},
  year={2023},
  pages={29–42}
}

A dual-path generative adversarial network (DPGAN) for single-image low-light enhancement.

@article{yang2023dpgan,
  title={Single Image Low-Light Enhancement via a Dual-Path Generative Adversarial Network},
  volume={42},
  ISSN={1531-5878},
  url={http://dx.doi.org/10.1007/s00034-023-02311-8},
  DOI={10.1007/s00034-023-02311-8},
  number={7},
  journal={Circuits, Systems, and Signal Processing},
  publisher={Springer Science and Business Media LLC},
  author={Yang, Shaoliang and Zhou, Dongming},
  year={2023}, month=Feb,
  pages={4221–4237}
}

A vision-transformer-based GAN for low-light enhancement that controls noise during enhancement: an iterative multi-branch feature-extraction subnet followed by an image-reconstruction subnet, built on multi-head multi-covariance self-attention and a light feature-forward module.

@article{yang2022transformergan,
  title={Rethinking Low-Light Enhancement via Transformer-GAN},
  volume={29},
  ISSN={1558-2361},
  url={http://dx.doi.org/10.1109/lsp.2022.3167331},
  DOI={10.1109/lsp.2022.3167331},
  journal={IEEE Signal Processing Letters},
  publisher={Institute of Electrical and Electronics Engineers (IEEE)},
  author={Yang, Shaoliang and Zhou, Dongming and Cao, Jinde and Guo, Yanbu},
  year={2022},
  pages={1082–1086}
}

Places classical SIMP, DMF-TONN and an improved physics-informed variant, StablePINN, on identical 2D and 3D cantilever problems with matched materials, boundary conditions, loads and volume targets. StablePINN differs from DMF-TONN in one change: collocation points are sampled from a pre-allocated, GPU-resident boundary point pool.

• StablePINN reaches 10%–59% lower structural compliance than DMF-TONN across the 2D and 3D problems.
• Classical SIMP keeps exact volume-constraint satisfaction and a faster 2D wall-clock time.
• Direct SIMP-vs-neural compliance comparison needs matched energy-scaling conventions, which the paper leaves open.

@inproceedings{yang2026neuraltobench,
  title={NeuralTO-Bench: A Reproducibility Framework and Benchmark Foundation for Classical and Neural Mesh-Free Topology Optimization},
  author={S. Yang and J. Wang and Y. Wang},
  booktitle={ASME IDETC/CIE},
  year={2026},
  note={Accepted}
}

Maps qualitative engineering descriptors (e.g. "low stiffness, medium density") to quantitative property targets, then retrieves candidate unit cells from a library of implicit-function geometries and ranks them with a surrogate trained on voxel-FEM homogenization labels, instead of generating geometry directly.

• Joint Success@5 of 92.5% across the evaluation prompt suite (stiffness within tolerance and all hard constraints met).
• Stiffness mean relative error of 3.4% for the top-ranked result.
• Retrieved candidates stay consistent with Gibson–Ashby scaling over the explored density range.

@inproceedings{wang2026semantic,
  title={Semantic Inverse Design of Cellular Structures in Implicit Field Space},
  author={Y. Wang and Z. Yashengjiang and B. Yu and S. Yang and J. Wang},
  booktitle={ASME IDETC/CIE},
  year={2026},
  note={Accepted}
}

A learning-based framework that decodes the signed distance field of a complex 3D aperiodic structure into interpretable design parameters, notably a periodicity factor and a fill-fraction scale, with an automatic dataset-generation procedure that supervises those parameters directly.

• Recovers the governing parameters and reconstructs the SDF from them

@inproceedings{yang2025inverseinference,
  series={IDETC-CIE2025},
  title={Inverse Inference of Aperiodic Structures: Decoding Signed Distance Fields Into Learnable Parameters for 3D Reconstruction},
  url={http://dx.doi.org/10.1115/detc2025-168575},
  DOI={10.1115/detc2025-168575},
  booktitle={Volume 2A: 45th Computers and Information in Engineering Conference (CIE)},
  publisher={American Society of Mechanical Engineers},
  author={Yang, Shaoliang and Wang, Jun},
  year={2025}, month=Aug,
  collection={IDETC-CIE2025}
}

Connects machine-learning curve generation to isogeometric analysis: generated curves keep their NURBS parameterization, and a companion analytical framework constructs the coordinates and computes physical displacements, so generated geometry can be analysed directly on UIUC airfoil and Superformula shapes.

@inproceedings{yang2024igasynthesis,
  series={IDETC-CIE2024},
  title={Enhancing Isogeometric Analysis With NURBS-Based Synthesis},
  url={http://dx.doi.org/10.1115/detc2024-142195},
  DOI={10.1115/detc2024-142195},
  booktitle={Volume 2B: 44th Computers and Information in Engineering Conference (CIE)},
  publisher={American Society of Mechanical Engineers},
  author={Yang, Shaoliang and Wang, Kang and Wang, Jun},
  year={2024}, month=Aug,
  collection={IDETC-CIE2024}
}

In a matrix-free multigrid FGMRES solver for 3D SIMP, a reported converged solve is not always converged: the projected residual used for stopping can fall below tolerance while the true residual is far above it. The paper treats density-floor selection as a verified control problem and accepts no solution until the true relative residual is recomputed and meets 1e-6.

• On 4 of 102 held-out states the stopping residual passes while the true residual is 1.35–49.5× the tolerance
• A two-feature rule matches 98 of 102 reference classifications; the residual guard catches the other four, so all 102 selected solves meet the tolerance
• Preserves the original operator on 24 admissible states, at 2.5× the mean wall time of always using a 1e-3 floor

@misc{yang2026densityfloor,
  title={When a Positive SIMP Density Floor Is Not Enough: Solver Admissibility and Guarded Floor Selection in Matrix-Free 3D Topology Optimization},
  author={S. Yang and J. Wang and Y. Wang},
  year={2026},
  eprint={2607.26382},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2607.26382}
}

A single fused CUDA kernel performs gather, per-element stiffness multiplication and scatter accumulation in one pass, removing the DRAM traffic between stages of the conventional three-stage matrix-free implementation.

• 4.6–7.3× end-to-end SIMP speedup on one RTX 4090 across 216k–4.9M cantilever elements; 4.4× on a 499,125-element torsion case
• 2.3–4.6× against the same-precision FP32 three-stage baseline
• 3.2–4.9× lower energy than matched FP64 runs in instrumented board-power traces
• BF16 is shown to be ill-conditioned for this problem class rather than adopted

@misc{yang2026fusedggs,
  title={Matrix-Free 3D SIMP Topology Optimization with Fused Gather-GEMM-Scatter Kernels},
  author={S. Yang and J. Wang and Y. Wang},
  year={2026},
  eprint={2604.18020},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2604.18020}
}

A matrix-free Galerkin geometric multigrid hierarchy built around a fused fine-level operator, studied as a linear-solver result rather than an end-to-end speed claim, with an explicit failure-mode screen for high-contrast 3D SIMP systems.

• FP32-GMG: 1.62×, 1.75× and 3.12× wall-time ratios against a capped Jacobi-PCG baseline at 64k, 216k and 512k elements
• A 1M-element uniform system solved in 1.50 ± 0.58 s
• Reports its limits: heterogeneous pass rates fall to 1/9 at 512k, and BF16 is not faster than FP32

@misc{yang2026gmg,
  title={A Matrix-Free Galerkin Multigrid Solver and Failure-Mode Screen for Single-GPU 3D SIMP Linear Systems},
  author={S. Yang and J. Wang and Y. Wang},
  year={2026},
  eprint={2604.26441},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2604.26441}
}

A pipeline from a plain-English structural problem to a validated binary topology: an LLM configurator, a boundary-condition generator, a three-field SIMP solver, an eight-check structural evaluator and a closed-loop retry.

• Valid specifications on all 10 test problems, with a median +0.3% compliance penalty against expert ground truth
• Across 17 benchmarks the LLM controller reaches the lowest median compliance but a 76.5% pass rate; the deterministic schedule passes 100% at +1.5% compliance

@misc{yang2026autosimp,
  title={AutoSiMP: Autonomous Topology Optimization from Natural Language via LLM-Driven Problem Configuration and Adaptive Solver Control},
  author={S. Yang and J. Wang and Y. Wang},
  year={2026},
  eprint={2603.27000},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2603.27000}
}

Tests whether multimodal LLMs can act as inspectable spatial proposal modules around a SIMP solver: after each solve, density and stress fields are rendered, the model proposes ranked interventions, and deterministic safeguards accept, reject or stop each one.

• 1.9% lower geometric-mean compliance on a 16-problem 2D benchmark, reported as not statistically significant (p = 0.382)
• Deterministic hotspot seeding remains competitive; the paper supports an inspectable framework, not a claim that LLM visual reasoning improves the optimum

@misc{yang2026itersimp,
  title={IterSIMP-σ: Evaluating LLM-Assisted Spatial Interventions in Stress-Aware Topology Optimization},
  author={S. Yang and J. Wang and Y. Wang},
  year={2026},
  eprint={2605.19110},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2605.19110}
}

A language model parses an engineer's thermal-mechanical request into a typed query; deterministic search over 1,397 homogenised unit cells and 19 base materials (26,543 combinations) then either returns a catalogue row that a second solver can rebuild, or refuses an unsatisfiable request with the minimal conflicting constraints and the slack of a repair, instead of returning the nearest neighbour. One registry declaration generates both the prompt and the evaluator, so undeclared parsed keys cannot reach search.

• On a frozen suite of 64 typed queries, matches min-repair on every feasible query (48/48) and refuses every empty one (16/16) with minimal unsatisfiable subsets and slack.
• Constraint-ignoring nearest-neighbour and penalty search violate stated cost or density limits.
• Repairs printed at three significant figures and rounded outward stay feasible on 216/216 empty queries.
• The 308-request parse benchmark is template text, not unconstrained engineer prose.

@misc{yang2026registry,
  title={A property-registry contract for retrieve-or-refuse thermal-mechanical lattice search},
  author={S. Yang and H. Chu and Z. Yashengjiang and J. Wang},
  year={2026},
  eprint={2609.14741},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2609.14741}
}

Super-resolution can make inspection images look sharper without preserving the evidence needed to detect a defect. This benchmark separates reconstruction from detection and scores both at a predeclared low false-positive rate.

• The two learned reconstruction models reach the highest structural similarity yet detect fewer defect pixels than bicubic interpolation in every paired repetition
• Passing calibration does not guarantee transfer: a joint model passes clean calibration but exceeds the held-out limit in all ten repetitions

@misc{yang2026waferinspectsr,
  title={Does Super-Resolution Preserve Defect Evidence? A Low-False-Call Benchmark for Semiconductor Inspection},
  author={S. Yang and J. Wang},
  year={2026},
  eprint={2607.17401},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2607.17401}
}

Treats image restoration before an inspection detector as a selective decision: return the measured image, one of several restorations, or send it to review. A fixed gate is certified on held-out data with exact binomial bounds on evidence-loss and excess-activation incidents.

• Evaluated on 4,591 public Carinthia-S images
• An auditable framework for when a transformed image may be returned automatically, explicitly not a claim that adaptive routing beats simpler policies

@misc{yang2026saferestore,
  title={SafeRestore: Detector-Relative Risk Certificates for Selective Industrial Image Restoration},
  author={S. Yang and J. Wang},
  year={2026},
  eprint={2609.03475},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2609.03475}
}

Remaining-useful-life estimates help maintenance decisions only if both accuracy and prediction intervals hold when operating conditions change. On a 10-bearing PHME subset with held-out load–speed regimes, a calibrated model fusing vibration windows, engineered descriptors and operating context is evaluated under strict data separation.

• Normalized MAE 0.1477 and empirical 90% coverage 0.900, against 0.1538 and 0.871 for a 400-tree random forest
• Conditional diagnostics expose non-uniform reliability, e.g. 0.666 coverage in a low-load, high-speed cell

@misc{yang2026bearingrul,
  title={Empirical Calibration and Conditional-Reliability Diagnostics for Bearing RUL Prediction under Operating-Regime Shift},
  author={S. Yang and J. Wang and Y. Wang},
  year={2026},
  eprint={2607.08273},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2607.08273}
}