PODiff: Latent Diffusion in Proper Orthogonal Decomposition Space for Scientific Super-Resolution

Abstract: Probabilistic super-resolution of high-dimensional spatial fields using diffusion models is often computationally prohibitive due to the cost of operating directly in pixel space. We propose PODiff, a structured conditional generative framework that performs diffusion in a fixed, variance-ordered Proper Orthogonal Decomposition (POD) coefficient space, exploiting the orthogonality of POD modes to impose an interpretable, variance-ordered latent geometry. This design enables efficient ensemble generation, preserves dominant spatial structure, and yields spatially interpretable, well-calibrated uncertainty at substantially lower computational cost. We evaluate PODiff on sea surface temperature downscaling over the West Australian coast and on a controlled advection-diffusion benchmark. PODiff achieves reconstruction accuracy comparable to pixel-space diffusion while requiring significantly less memory and producing more reliable uncertainty estimates than deterministic and Monte Carlo Dropout baselines.

• The paper introduces a novel diffusion framework in a variance-ordered POD latent space to achieve super-resolution with interpretable uncertainty quantification.
• It leverages linear POD decomposition to reduce computational cost and ensure accurate spatial uncertainty propagation over high-dimensional scientific fields.
• Experimental results on SST downscaling and PDE benchmarks demonstrate significant improvements over pixel-based approaches while drastically lowering computational requirements.

PODiff: Latent Diffusion in Proper Orthogonal Decomposition Space for Scientific Super-Resolution

Introduction and Motivation

High-resolution spatial field estimation is computationally intensive in scientific applications such as climate analysis and oceanography. Super-resolution approaches address this by attempting to recover fine-scale structure from low-resolution measurements. While deep conditional generative models based on diffusion processes have proven capable in probabilistic super-resolution tasks, their application in scientific contexts is challenged by high computational cost, especially when operating directly in pixel space. Typical latent diffusion methods address some efficiency bottlenecks by encoding data into a learned nonlinear latent space, but these have limited interpretability and difficult uncertainty propagation—contradicting core needs in scientific modeling.

In "PODiff: Latent Diffusion in Proper Orthogonal Decomposition Space for Scientific Super-Resolution" (2605.03399), the authors introduce a conditional generative diffusion framework operating in a reduced, interpretable, variance-ordered latent POD space rather than pixel or encoded latent space. This approach leverages the orthogonality, linearity, and variance-ordering of POD, matching the low-rank structure common to scientific fields and enabling efficient, interpretable uncertainty quantification.

Methodology

POD as an Interpretable Latent Space

Proper Orthogonal Decomposition (POD) offers a linear basis for scientific fields, with modes ordered by explained variance. Each spatial field is projected onto the leading K POD modes (usually small, K ≪ d), capturing dominant structures efficiently and supporting hierarchical representation by spatial scale. This provides several intrinsic advantages for diffusion-based generation: (1) reduced dimensionality for computational efficiency; (2) orthogonality and linear mapping for stability; (3) variance-ordering for modal interpretability and structured uncertainty propagation.

The mean-removed, high-dimensional fields $\mathbf{u}_i \in \mathbb{R}^d$ are decomposed using SVD, and the leading $K$ orthonormal eigenvectors comprise the basis matrix $\Phi \in \mathbb{R}^{d \times K}$. Each field is then expressed as $$\mathbf{u} \approx \overline{\mathbf{u}} + \Phi \mathbf{a}$$, where $\mathbf{a} \in \mathbb{R}^K$ are the POD coefficients.

Latent Diffusion Process

Diffusion operates on standardized POD coefficients. Conditioning is accomplished by projecting bicubically upsampled low-res input onto the same truncated POD basis, providing a consistent K-dimensional context vector. Forward diffusion perturbs the target latent vector with an additive Gaussian process over multiple steps; a neural network denoiser (a residual MLP) is trained to estimate the injected noise, conditioned on the noisy coefficients, the context vector, and the diffusion timestep.

Sampling is performed by initializing at pure noise in the latent space and applying reverse diffusion. Each denoised latent sample is mapped back to the physical space via linear reconstruction.

Baselines and Ablations

Rigorous baselines include:

• Deterministic POD-projection (no learning or stochasticity; direct projection from conditioning)
• RandOrthDiff (diffusion in a random orthonormal latent basis, controlling for the necessity of a variance-ordered basis)
• Deterministic/MC Dropout U-Nets (pixel-space baselines for point prediction and uncertainty via dropout)
• Pixel-space diffusion (PixelDiff) (direct application of diffusion models in pixel space)
• Radial basis function interpolation (classical non-learning approach)

These ablations precisely isolate the impact of the structured latent basis and diffusion mechanism.

Uncertainty Quantification

A primary advantage of PODiff is analytic spatial uncertainty propagation. The covariance in the POD coefficients propagates linearly to the reconstructed field space, enabling direct computation of spatial uncertainty maps. This allows both well-calibrated predictive intervals and interpretable spatial localization of uncertainty. Calibration is evaluated with empirical coverage, reliability curves, mean absolute calibration error (MACE), and proper scoring metrics (CRPS).

Computational Advantages

The move from pixel space ($d$ dimensions) to POD latent space ($K \ll d$) reduces both memory and parameter count by orders of magnitude. PODiff's per-sample inference is competitive with deterministic models. Since reconstruction is linear, generating sample ensembles and propagating uncertainty is significantly less expensive than pixel-based ensemble generation.

Experimental Validation

SST Downscaling over the Western Australian Coast

Using a challenging test case (SST during a 2011 marine heatwave) with strong gradients and spatial heterogeneity, PODiff demonstrates superior performance on both reconstruction and uncertainty metrics.

Model	RMSE (°C)	MAE (°C)	Ext. RMSE	Ext. MAE
PODiff-K40	0.3923	0.2976	0.4836	0.3537
PixelDiff	0.4118	0.3158	0.4899	0.3600
U-Net	0.6788	0.5141	0.8366	0.6109
MC Dropout U-Net	—	—	—	—
RBF	0.7784	0.5804	0.7899	0.5936

PODiff outperforms pixel-based U-Nets and interpolation, and even achieves slightly lower error than pixel-space diffusion (at a fraction of the memory/cost). The ablation using a random basis (RandOrthDiff) performs much worse, underscoring the criticality of variance-ordering and data-adaptivity in the POD latent space.

Uncertainty Calibration

Empirical coverages closely track nominal levels (e.g., 90% interval coverage: 0.9009 for PODiff-K40; superior to MC Dropout). The mean absolute calibration error is 0.0128, and CRPS is lower than baselines. The spatial structure of predictive uncertainty is concentrated in regions with high physical gradients (e.g., near coasts and heatwave regions), not noise-dominated or uniform, further highlighting the interpretability and selectivity enabled by the POD basis.

Efficiency

PODiff drastically reduces computational burden:

• Parameters: 0.2M (PODiff, K=40) vs. 33M (U-Net/PixelDiff)
• Peak GPU memory: 1.4GB (PODiff) vs. 12.5GB (PixelDiff)
• Training time: 3.8h (PODiff) vs. 48h (PixelDiff)
• Inference time/sample: 0.08s (PODiff) vs. 1.24s (PixelDiff)

Despite these reductions, accuracy and uncertainty calibration remain comparable or better than the best pixel-based methods.

Synthetic Advection-Diffusion PDE Benchmark

In a PDE regime with known dynamics, PODiff achieves ensemble mean RMSE of 0.018 and MAE of 0.0098. Uncertainty is sharply concentrated at regions of localized peaks, matching epistemic uncertainty inherent to the physical dynamics. Reliability curves show near-perfect alignment between empirical and nominal coverage, demonstrating the robustness of both the generative and uncertainty quantification capabilities.

Implications and Future Directions

PODiff's demonstration of diffusion in a variance-ordered, fixed, interpretable linear subspace addresses several key gaps in scientific generative modeling:

• Interpretability: Enables inspection and understanding of uncertainty by spatial scale.
• Computational efficiency: Makes diffusion-based probabilistic super-resolution practical for large climate or geophysical data.
• Uncertainty propagation: Ensures analytic, physically principled transition of uncertainty from latent to real space.

This framework fits scenarios where the signal of interest is well described by low-rank structure, as in climate, ocean, and other PDE-based scientific systems.

Future Research Trajectories

Potential directions include:

• Extension to non-stationary or adaptive subspaces: For phenomena with strong non-linearities or nonstationarity, methods to adapt the POD basis online, or combine with nonlinear manifold approaches, could further improve robustness.
• Integration with physics-informed constraints: Embed PDE knowledge or physical loss priors within the diffusion process.
• Hybridization with learned latents: Where POD truncation is limiting, combine variance-ordered and learned representations.

Conclusion

PODiff offers a principled, efficient, and interpretable approach to probabilistic super-resolution for high-dimensional scientific fields. By conducting conditional diffusion in a variance-ordered POD latent space, the framework achieves competitive or superior accuracy to pixel-based approaches, drastically lowers computational cost, and yields well-calibrated, spatially structured uncertainty estimates. This method is particularly well-aligned with the low-rank, spatially coherent nature of many scientific datasets and stands as a practical alternative to latent and pixel-based diffusion for uncertainty-aware scientific modeling.