Co-PiLOT: Constrained Physics-Informed Latent Optim... | AI Research

Inverse design asks a difficult question: if a material must exhibit a particular behaviour, what structure should be built to produce it? In the paper Co-PiLOT: Constrained Physics-Informed Latent Optimization for Target-Driven Inverse Design, the authors address this problem for magnesium-alloy microstructures. Their system searches a compact latent representation learned by a generative model, converts proposed designs into crystallographic orientation fields, and evaluates them with an expensive crystal-plasticity simulator. The goal is to find structures whose simulated stress–strain properties match a specified target while avoiding designs that are invalid or cause the simulator to fail.

What Co-PiLOT is designed to solve

Directly optimizing a microstructure is challenging because the design space is structured rather than an ordinary vector space. It contains grains, grain boundaries, and crystallographic orientations, and many possible arrangements are physically invalid. Gradients are generally unavailable through a black-box simulator, while each 300 × 300 simulation takes about six minutes and can diverge.
Co-PiLOT moves the search into a learned latent space. An encoder maps microstructure images into a latent vector, while a decoder maps latent vectors back into candidate microstructures. The optimizer then searches within a bounded latent region instead of modifying every pixel, grain, or orientation independently. This gives the search a learned prior: decoded candidates remain near the distribution of microstructures represented in the training data.
The framework is written as a general combination of a decoder, a physical oracle, an objective, constraints, and an optimizer. Although the authors describe possible applications to molecules, photonics, and topology optimization, the demonstrated application is magnesium-alloy microstructure and texture design.

How the method connects images to physics

The generative models are trained on 81,756 electron backscatter diffraction (EBSD)-derived microstructures from 17 magnesium-alloy classes. A Vision Transformer encoder compresses each 512 × 512 orientation image into a bottleneck vector. The paper evaluates bottleneck sizes of 512, 768, and 1,024 dimensions, and pairs the encoder with several diffusion-based decoders.
The strongest reported combination is ViT-FMDiT, which uses a rectified-flow or flow-matching diffusion transformer decoder. With a 768-dimensional latent representation, it reconstructs high-fidelity microstructure images, achieving a Fréchet Inception Distance of 27.86 and an MS-SSIM of 0.178. These image metrics are used to assess reconstruction quality; they do not by themselves show that a decoded structure will achieve a desired mechanical response.
A central component is the orientation codec. A generic image decoder produces RGB values, but the crystal-plasticity solver requires a grid of valid hexagonal close-packed crystal orientations. The codec represents orientations as quaternions, unfolds neighbouring grains to reduce artificial discontinuities, applies a class-level alignment, and maps the result to RGB through stereographic projection. Its closed-form inverse converts generated images back into orientation fields.
The codec also segments generated images into grains without relying on ground-truth labels. The reported round-trip orientation error is approximately 0.6–0.7 degrees, below the 5-degree grain-boundary threshold cited by the authors. This bridge allows an image-domain generator to supply inputs to a crystal-plasticity solver rather than treating the generated image as the final design.
Latent diffusion is also used in other settings to iteratively refine compact representations, as in continuous latent diffusion for reasoning. Co-PiLOT uses a related latent-space idea for a different purpose: not generating a reasoning sequence, but proposing physical structures that must survive conversion into a simulator-ready representation.

How MERIDIAN chooses the next simulation

The optimization loop is called MERIDIAN. At each step, a latent vector is decoded, converted through the orientation codec, and evaluated by the crystal-plasticity oracle. The simulator returns properties such as yield strength, hardening behaviour, and ultimate strength. These are combined into a scalar objective measuring distance from the target, while also incorporating a ductility floor and penalties related to unphysical grain counts.
MERIDIAN uses a deep-kernel Gaussian-process surrogate to estimate likely objective values and uncertainty. It also trains a feasibility classifier on failed decoder, codec, or solver evaluations. This matters because a failed simulation is not simply a poor score: it may provide no usable objective value at all. By learning where failures occur, the optimizer can avoid repeatedly querying regions likely to diverge.
The method further uses manifold-aware trust regions. Rather than making unrestricted moves throughout the latent box, it performs local search shaped by the surrogate and projects candidates onto a latent shell intended to remain within the decoder’s valid region. A determinantal point process selects diverse batches of candidates. Together, these mechanisms are intended to make better use of a very small number of expensive simulations.

What the reported results show—and what remains open

Under a budget of 160 simulations, the ViT-FMDiT-768 decoder combined with MERIDIAN achieves the best target-driven objective among the tested combinations. The authors report a 3–22% reduction in relative target error compared with seven baselines: DANTE, TuRBO, BAxUS, CMA-ES, DDOM, SEIKO, and DDPO, all evaluated with the same decoder.
The comparison supports the paper’s central claim that handling simulator failures and respecting the learned design manifold can improve black-box search efficiency. It does not establish that Co-PiLOT is universally superior across all inverse-design problems. The experiments cover one application domain—magnesium-alloy microstructures—and the paper’s broader framework is presented as a possible template for other domains rather than as a demonstrated result there.
The approach also depends on the quality and coverage of its learned microstructure prior, the orientation codec, and the simulator. A decoder can make candidates look plausible without guaranteeing the desired physical response, while the optimizer remains limited by the small simulation budget. The authors release the code, but further applications would be needed to determine how well the framework transfers beyond the magnesium-alloy setting.