Research map / Probabilistic and scientific modeling / Bayesian optimization

Co-PiLOT: Constrained Physics-Informed Latent Optimization for Target-Driven Inverse Design

NeurIPS 2026 · Probabilistic and scientific modeling › Bayesian optimization · approach: Offline black-box optimization with surrogates · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Co-PiLOT couples a ViT-based encoder and diffusion/flow-matching decoders with an orientation codec and a new latent optimizer (Meridian) to inverse-design Mg-alloy microstructures against a crystal-plasticity simulator under a 160-evaluation budget.

Problem

Inverse design of microstructures against an expensive, occasionally divergent black-box crystal-plasticity simulator, where raw design spaces are non-Euclidean and most candidates are physically invalid.

Approach

Searches a pretrained generative latent space with a GP-based optimizer that adds a failure-aware feasibility classifier, active-subspace manifold-aware trust regions with latent-shell projection, and DPP batch selection, coupled to a reversible orientation codec feeding DAMASK.

Key results

• ViT-FMDiT-768 reconstructs microstructures with FID 27.86 and MS-SSIM 0.178 on the 9,084-image test set; ViT-DiT achieves the best FID of 23.19.
• ViT-FMDiT-768 with Meridian reaches J_V2 = -0.132 ± 0.004 after 144 evaluations (sigma_y 159.1 MPa, n 0.299, sigma_u 314.0 MPa) against the AZ31 target.
• Meridian is reported best in all eight optimizer-decoder-objective studies, with a claimed 3–22% relative target-error reduction over seven baselines.

Surprising findings

• Larger 1024-D bottleneck underperforms 768-D for inverse design despite better reconstruction, because the latent becomes too sparse under the 160-evaluation budget.
• SEIKO and DDPO did not beat the warm-start incumbent within 160 evaluations.

Limitations

• Decoder can only search microstructures the learned prior can express; non-isometric orientation space with ~0.6-0.7 degree codec error but ~53 degree per-pixel disorientation in round-trip.
• Oracle is HCP phenomenological power-law plasticity without damage, isothermal loading, and only uniaxial tension; no experimental validation.
• Only one domain and a single 160-evaluation budget are evaluated; dataset expansion details deferred to future publication.

Experimental setup

Datasets: EBSD-derived Mg-alloy microstructure dataset (81,756 train / 9,084 test)

Benchmarks: Mg-alloy microstructure test set (9,084 images) AZ31 target-driven inverse design (160 simulations)

Metrics: FID MS-SSIM LPIPS grain recovery G_m orientation disorientation J_V2 target objective J_V1 toughness objective

Baselines: DANTE TuRBO BAxUS CMA-ES DDOM SEIKO DDPO

Builds on: TuRBO BAxUS DANTE DDOM DDPO SEIKO VICReg

Closest papers by meaning

• Vector Field Oriented Diffusion Model for Crystal Material Generation (AAAI 2024 · similarity 0.86)
• A Diffusion-Based Pre-training Framework for Crystal Property Prediction (AAAI 2024 · similarity 0.85)
• A Denoising Pre-training Framework for Accelerating Novel Material Discovery (AAAI 2025 · similarity 0.85)
• 3DID: Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization (NeurIPS 2025 · similarity 0.85)
• AutoMS: Multi-Agent Evolutionary Search for Cross-Physics Inverse Microstructure Design (ICML 2026 · similarity 0.85)
• Crystalformer: Infinitely Connected Attention for Periodic Structure Encoding (ICLR 2024 · similarity 0.85)