GUIDe: Generative & Uncertainty-Based Inverse Design

• GUIDe is an inverse design framework that combines generative models with explicit uncertainty quantification to address ill-posed one-to-many problems.
• It leverages latent variable formulations such as GANs, VAEs, and normalizing flows to map target properties to multiple feasible designs.
• By integrating physics-informed training and budget-aware optimization, GUIDe enhances reliability and decision-making in materials and aerodynamic designs.

Generative and Uncertainty-informed Inverse Design (GUIDe) is an Editor’s term for inverse-design workflows that combine a generative model of feasible designs with an explicit treatment of the one-to-many uncertainty inherent in inverse problems. A concrete early instance is a framework for microstructural materials design in which a generative adversarial network (GAN) learns a low-dimensional latent design space and a Mixture Density Network (MDN) learns a conditional distribution $p(z\mid y)$, so that a desired optical absorption yields multiple candidate microstructures rather than a single deterministic output (Yang et al., 2021). In broader formulations, inverse design is cast as constrained, budget-aware optimization over a representable design space $\mathcal{X}$ and a feasible manifold $\mathcal{C}$, with generative models supplying priors over admissible designs and uncertainty-aware surrogates or acquisition rules steering search under expensive evaluations such as DFT, CFD, experiment, or PDE solves (Babu et al., 1 Jun 2026).

1. Problem structure: ill-posed, one-to-many, and budget-constrained

Inverse design in GUIDe-style systems begins from the asymmetry between forward and inverse maps. In microstructural materials design, the forward relation maps a microstructure image $x$ to a scalar property $y$, while inverse design seeks $x$ such that $f_{\text{phys}}(x)\approx y^\ast$. Because different microstructures can yield the same or very similar properties, the inverse is naturally a multivalued relation; in latent form this is expressed as a conditional distribution $p(z\mid y)$, not a single-valued function (Yang et al., 2021). The same structural issue appears across scientific inverse problems: in imaging, $y=Ax_\ast+e$; in PDE-governed media, sparse or noisy observations constrain only a subset of possible coefficient fields; in materials discovery, multiple crystal tuples $x=(\mathbf{L},\mathbf{Z},\mathbf{S})$ may satisfy the same target profile (Jin et al., 1 Jun 2026).

A general formulation makes this explicit as constrained optimization over representable and feasible sets,

$\mathcal{X}$0

and, when high-fidelity evaluation is costly,

$\mathcal{X}$1

Here $\mathcal{X}$2 is the representable design space, $\mathcal{X}$3 is the feasible manifold, and $\mathcal{X}$4 folds together target properties, cost, and robustness (Babu et al., 1 Jun 2026). This suggests that GUIDe is not merely conditional generation. It is a search problem under feasibility, ambiguity, and evaluation budget.

A persistent technical obstacle is dimensionality mismatch. In the microstructure setting, the inverse input may be a scalar property $\mathcal{X}$5, while the naïve output is an image $\mathcal{X}$6. Direct $\mathcal{X}$7 learning must hallucinate missing information, and the cited GAN–MDN study reports that direct MDN inversion tends to collapse toward lower-absorption microstructures and fails for high targets, especially on larger images (Yang et al., 2021). This motivates latent-variable formulations in which inverse inference is performed in a compact space and decoded into full designs only afterward.

2. Generative parameterizations of feasible design manifolds

The generative component of GUIDe acts as a decoder of feasible designs. In the GAN–MDN microstructure framework, the generator $\mathcal{X}$8 learns a nonlinear manifold of physically realistic two-phase microstructures, and inverse design becomes $\mathcal{X}$9 (Yang et al., 2021). This separates representation learning of feasible designs from probabilistic inverse inference on that representation.

Across materials and scientific inverse problems, several model families supply this generative layer. Variational autoencoders use latent variables $\mathcal{C}$0 and decoders $\mathcal{C}$1, trained by the ELBO; normalizing flows learn invertible maps with exact likelihoods; autoregressive transformers model $\mathcal{C}$2; and diffusion or score-based models learn denoising or reverse-time dynamics, often with classifier-free guidance for property conditioning (Babu et al., 1 Jun 2026). In crystalline solids, these generators operate on crystal-native representations such as lattice matrices, species, and fractional coordinates, with equivariant architectures and symmetry-aware encodings used to reduce spurious degrees of freedom (Babu et al., 1 Jun 2026).

Physics-informed operator models extend this idea when design variables are fields rather than compact vectors. IGNO encodes high-dimensional coefficient fields $\mathcal{C}$3 and boundary data into low-dimensional latents, decodes both coefficient fields and PDE solutions using MultiONet neural operators, and is trained without labeled $\mathcal{C}$4 pairs using PDE residuals and boundary conditions (Bao et al., 5 Nov 2025). GenPANIS goes further for discrete multiphase media by keeping microstructures exactly discrete while performing inference in a continuous latent space, with a learnable normalizing-flow prior and a decoder that contains a differentiable coarse-grained PDE solver (Chatzopoulos et al., 16 Feb 2026). Design-GenNO applies the same pattern to inverse microstructure design: a flow-regularized latent space is decoded both to microstructures and to full PDE solution fields, enabling multiple objectives without retraining (Zang et al., 10 Sep 2025).

This suggests a common GUIDe architecture: a learned generative prior defines where sampling is allowed, and physics-aware decoding defines what counts as a valid response. The more that feasibility is baked into representation, decoder, and training objective, the less the inverse stage must rely on rejection after generation.

3. Uncertainty representation, posterior inference, and trustworthiness

The uncertainty-aware part of GUIDe replaces point estimates with conditional distributions over feasible designs. In the microstructure GAN–MDN framework, the inverse model is

$\mathcal{C}$5

with $\mathcal{C}$6 diagonal Gaussian components, so that different mixture components represent different latent design modes and the component variances represent dispersion around each mode (Yang et al., 2021). Sampling from this mixture and decoding through the GAN yields multiple candidate microstructures for the same target property.

A more explicitly Bayesian variant performs posterior analysis directly in latent space. In the VAE-based framework for Bayesian inverse problems, the generative prior is learned once on uncorrupted data, and latent posterior inference is then performed for arbitrary corruption operators by approximating $\mathcal{C}$7 with Gaussian or Gaussian-mixture surrogates using EL$\mathcal{C}$8O and local Hessian information (Böhm et al., 2019). GenPANIS uses a learned normalizing-flow prior and Hamiltonian Monte Carlo in latent space to sample from

$\mathcal{C}$9

then decodes posterior samples into discrete microstructures, posterior means, and pixel-wise variances (Chatzopoulos et al., 16 Feb 2026). GenAI4UQ takes a related conditional-generative route for inverse uncertainty quantification by learning a direct mapping from observations and Gaussian noise to samples from parameter posteriors, thereby avoiding per-query MCMC (Fan et al., 2024).

Trustworthiness enters when plausible samples are not necessarily measurement-supported. In generative imaging inverse problems, a local manifold model yields a tangent space $x$0 and a worst-direction compatibility constant

$x$1

together with a log-determinant score

$x$2

which quantify how well the measurement operator observes prior-relevant directions (Jin et al., 1 Jun 2026). The corresponding theory shows that small $x$3 permits large tangent error even when the prior is strong, explaining measurement-induced hallucinations. Posterior clouds—sets of conditional reconstructions sampled from the current posterior—then become uncertainty objects in their own right: their covariance identifies ambiguous directions that can guide additional measurements (Jin et al., 1 Jun 2026).

GUIDe-style uncertainty can also be architecture-intrinsic. Diagonal Flow Matching introduces Zero-Deviation and Self-Consistency as internal reliability scores; these enable selecting the best candidate among multiple generations, abstaining from unreliable predictions, and detecting out-of-distribution targets, and are reported to consistently outperform ensemble and general-purpose alternatives across the evaluated tasks (Campos et al., 16 Mar 2026). This suggests that uncertainty in GUIDe need not be limited to posterior variance; it can also be operationalized as abstention, self-consistency, or measurement support.

4. Physics-informed training and closed-loop optimization

Many GUIDe implementations are distinguished less by the generator than by how tightly the generator is coupled to physics and to decision-making under evaluation budgets. In the physics-informed operator setting, PDE residuals become virtual observables. IGNO trains on a loss combining PDE residual, boundary-condition loss, and coefficient reconstruction,

$x$4

and performs inverse inference by optimizing a low-dimensional latent variable rather than a full coefficient field (Bao et al., 5 Nov 2025). GenPANIS uses an ELBO with three terms corresponding to unlabeled microstructures, labeled $x$5 pairs, and virtual residual observations, thereby integrating simulation-free structural data and physics residuals in one probabilistic objective (Chatzopoulos et al., 16 Feb 2026). Design-GenNO uses a similar ELBO-style construction with virtual PDE residual likelihoods and reports 95% trust intervals for effective properties (Zang et al., 10 Sep 2025).

Closed-loop GUIDe systems additionally require search policies. The 2026 inverse-materials-design review places Bayesian optimization, reinforcement learning, and active learning into a common budget-aware loop. It gives upper confidence bound,

$x$6

expected improvement, constrained acquisition $x$7, and target-oriented acquisition $x$8 as canonical choices for expensive evaluations (Babu et al., 1 Jun 2026). The same review also advocates a four-stage constraint pipeline: representation-level priors, training-time constraints, sampling-time guidance, and post-generation screening plus relaxation (Babu et al., 1 Jun 2026).

The measurement-design literature shows that the “design variable” in GUIDe can itself be the measurement operator. Posterior-cloud measurement design starts from an initial acquisition $x$9, samples a posterior cloud conditioned on $y$0, estimates a local subspace $y$1, and greedily chooses additional measurements to maximize a log-determinant criterion on $y$2 (Jin et al., 1 Jun 2026). This is inverse design over acquisition geometry rather than over object space, but it follows the same generative-and-uncertainty-informed logic.

A broader system view is explicit in the 2026 review: multimodal encoder, generator, verifier, optional robotic lab, characterization stack, and scheduler exchanging both data and uncertainty. This suggests that GUIDe is best understood as the algorithmic core of an orchestrated closed loop, not as a standalone generator (Babu et al., 1 Jun 2026).

5. Representative implementations and empirical behavior

In microstructural materials design, the GAN–MDN framework provides one of the clearest early demonstrations of the GUIDe pattern. It uses a GAN trained on Gaussian-random-field microstructures, an MDN with $y$3 Gaussians in latent space, and RCWA to evaluate optical absorption. On two datasets—Data-I with $y$4 images and latent dimension $y$5, and Data-II with $y$6 images and latent dimension $y$7—the proposed GAN-MDN reports minimum residual error percentage values that are consistently very low, often below $y$8, with running time around $y$9 seconds for 30 designs, whereas optimization-based inverse modeling requires $x$0 hours per target and produces only one solution (Yang et al., 2021).

In inverse microstructure design with operator decoders, Design-GenNO reports $x$1 success for effective-property matching in a target region within the training distribution and $x$2 for a target region outside the training distribution, compared with $x$3 and $x$4 for PoreFlow; for sparse-field matching it reports a cross-correlation index $x$5 (Zang et al., 10 Sep 2025). GenPANIS, on Darcy and Helmholtz problems, is reported to outperform state-of-the-art baselines while using $x$6 times fewer parameters and to maintain accuracy under unseen boundary conditions, volume fractions, and morphologies with sparse, noisy observations (Chatzopoulos et al., 16 Feb 2026).

In aerodynamic design, Dflow-SUR combines a flow-matching generator with surrogate-based physical loss, but separates physical optimization from flow inference by differentiating through the full flow map. Relative to the strongest energy-based baseline on the airfoil case, it reduces physical loss by four orders of magnitude and cuts wall-clock time by $x$7; on wing design, it increases the mean lift-to-drag ratio by $x$8 over traditional Latin-hypercube sampling (Yang et al., 9 Dec 2025). Diagonal Flow Matching, evaluated on airfoil, gas-turbine combustor, and analytical tasks up to design dimension $x$9, reports order-of-magnitude improvements in round-trip accuracy over standard conditional flow matching and invertible neural network baselines, while its Zero-Deviation and Self-Consistency scores support candidate selection, abstention, and OOD detection (Campos et al., 16 Mar 2026).

In diffusion-based airfoil inverse design, a classifier-free guided denoising diffusion probabilistic model generates pressure-coefficient distributions conditioned on six pressure features and then maps them to airfoil geometry. On classical transonic airfoils, the method reports a $f_{\text{phys}}(x)\approx y^\ast$0 precision improvement over Wasserstein GAN methods in airfoil generating tasks, while the classifier-free guidance coefficient provides an explicit control over fidelity to the designated pressure features (Deng et al., 10 Mar 2025). In metamaterials, RIGID uses a random-forest forward classifier, target-conditioned likelihood extraction, and MCMC sampling to operate in a small-data regime with fewer than 250 training samples, generating satisfactory solutions that cover a broader range of the design space than the genetic-algorithm baseline (Chen et al., 2023).

These case studies indicate that GUIDe is not tied to a single generative family. GANs, diffusion models, flow matching, random-forest likelihoods, neural operators, and latent-variable Bayesian models all instantiate the same pattern when generation is coupled to uncertainty-aware inverse inference.

6. Evaluation standards, misconceptions, and limitations

A recurring misconception is that a plausible generative sample is therefore trustworthy. Measurement-geometry analysis shows that plausible reconstructions may be filled in by the prior along unobserved tangent directions of the data manifold; low $f_{\text{phys}}(x)\approx y^\ast$1 or low $f_{\text{phys}}(x)\approx y^\ast$2 signals elevated hallucination risk even when reconstructions appear realistic (Jin et al., 1 Jun 2026). A related misconception in materials discovery is that validity of representation implies utility: the PGCGM evaluation study reports that while generated crystals can satisfy syntactic validity checks, most generated structures are predicted to be thermodynamically unstable by a separate property-prediction model, and the default PGCGM input space is not smooth with respect to parameter variation, making material optimization difficult and limited (New et al., 2023).

The 2026 review formalizes broader failure modes: surrogate exploitation, diversity collapse, distribution shift, and the stability-synthesizability gap (Babu et al., 1 Jun 2026). These motivate staged evaluation rather than single metrics. The proposed ledger comprises validity, uniqueness, novelty, stability, and cost/efficiency, each of which can function either as a constraint or as a component of utility. This reporting protocol is intended to prevent metric gaming and trivial rediscovery (Babu et al., 1 Jun 2026).

Several current GUIDe instances also remain only partially probabilistic. In the GAN–MDN microstructure framework, GAN and MDN are trained sequentially rather than jointly, the MDN uses diagonal Gaussian components, and the number of components $f_{\text{phys}}(x)\approx y^\ast$3 is fixed by hand (Yang et al., 2021). In IGNO, inversion is performed by deterministic latent optimization and thus yields a point estimate unless additional posterior machinery is added (Bao et al., 5 Nov 2025). These facts suggest that a mature GUIDe system will often need further layers: richer density estimators, full posterior sampling, calibration under shift, and explicit decision rules for abstention or further measurement.

What emerges is less a single algorithm than a design doctrine. A GUIDe system learns a generative prior over feasible designs, constrains that prior with physics and feasibility, performs inverse inference in a space where multimodality can be represented rather than collapsed, and uses uncertainty not as an afterthought but as a control signal for candidate ranking, acquisition, abstention, and closed-loop refinement.