Physics-Aware Uncertainty Pruning (PAUP)

• Physics-Aware Uncertainty Pruning (PAUP) is a pruning paradigm that uses physical inconsistency signals as a proxy for predictive uncertainty.
• It applies thresholding and adaptive uncertainty scoring to selectively suppress or defer outputs in both surrogate-assisted inverse design and underwater 3D Gaussian Splatting.
• By substituting ensemble methods with residual-based checks against known physics, PAUP efficiently allocates resources and ensures artifact-free reconstructions.

Physics-Aware Uncertainty Pruning (PAUP) denotes a class of procedures in which physics-derived inconsistency signals are used to suppress, defer, or selectively validate model outputs that are likely unreliable. In the current literature represented by two distinct formulations, PAUP appears explicitly in underwater 3D reconstruction as a branch that prunes noisy floating Gaussians in 3D Gaussian Splatting, and implicitly in surrogate-assisted inverse design as physics-informed uncertainty embedded in a multi-fidelity, uncertainty-aware optimization loop. In both cases, the central idea is the same: predictions that violate governing physical structure, or that are unstable under physically meaningful checks, should not be trusted on equal footing with physically consistent ones (Xing et al., 8 Aug 2025, Xue et al., 26 Jan 2026).

1. Definition and scope

PAUP is not presented as a single standardized algorithm across the literature. Rather, the available work shows two operationally different but conceptually aligned instantiations.

Setting	Object pruned or de-prioritized	Physics signal
FSS inverse design	Candidate designs or surrogate suggestions	Continuity residuals of tangential electromagnetic fields
Underwater 3DGS	Gaussian primitives	Multi-view instability plus Beer–Lambert-aware inconsistency

In the frequency-selective surface (FSS) study, the paper does not use the name “PAUP”; instead, its “physics-informed uncertainty” inside a “multi-fidelity uncertainty-aware BPSO” plays the PAUP role. The stated role is to compute a physics-informed uncertainty score for each candidate, prune or de-prioritize high-uncertainty candidates in exploitation steps, select some high-uncertainty candidates for high-fidelity (HF) exploration, and use HF-validated beacons to guide subsequent search (Xue et al., 26 Jan 2026).

In UW-3DGS, PAUP is a named branch. It assigns each Gaussian a Physics-aware Uncertainty Score (PUS), converts that score into a pruning probability, and masks Gaussians whose scores indicate that they are likely to represent medium-induced artifacts rather than true scene structure. The paper describes this as adaptive removal of noisy floating Gaussians to ensure artifact-free geometry (Xing et al., 8 Aug 2025).

This suggests that PAUP is best understood as a physics-constrained pruning paradigm rather than a domain-specific heuristic. Its core function is resource allocation under uncertainty: it decides which predictions should propagate, which should be discounted, and which require further evaluation.

2. Core principle: physics as a proxy for uncertainty

The common principle across formulations is that violation of known physics is treated as a computationally cheap proxy for predictive uncertainty. In the FSS inverse-design setting, this is called physics-informed uncertainty (PIU). The surrogate’s credibility is assessed by the degree to which its predicted electromagnetic responses violate the continuity of tangential electromagnetic field across the metasurface. For reflection $R$ and transmission $T$ under $xx$ polarization, the boundary condition yields

$$1 + \mathrm{Re}(R) = \mathrm{Re}(T), \qquad \mathrm{Im}(R) = \mathrm{Im}(T).$$

Residuals of these equalities across the frequency band are aggregated into an uncertainty score. The rationale given is that surrogates extrapolate in data-sparse regions and may confidently predict unrealistic responses; physics residuals detect such inconsistencies without requiring ensembles or Bayesian surrogates (Xue et al., 26 Jan 2026).

In UW-3DGS, the physics-awareness comes from the underwater image formation model. The observed intensity is decomposed into direct and backscatter components,

$I(x) = D(x) + B(x),$

and the adopted participating-media formulation is

$$I = J \cdot e^{-\beta^D \cdot z} + B^\infty \cdot \left(1 - e^{-\beta^B \cdot z}\right).$$

UW-3DGS further uses the channel-wise Beer–Lambert form

$$I_c(x) = J_c(x)\, t_c(x) + B_c(x)\,(1 - t_c(x)), \qquad t_c(x) = \exp(-\beta_c(x)\, d(x)).$$

Underwater conditions can bias gradients in standard 3DGS, causing the optimizer to “explain” haze by spawning or inflating Gaussians in empty water. PAUP targets these spurious primitives by combining multi-view inconsistency with inconsistency relative to the learned attenuation model (Xing et al., 8 Aug 2025).

A plausible unifying interpretation is that PAUP replaces generic epistemic uncertainty estimators with residuals against structure that the model should satisfy anyway. The practical advantage, made explicit in the FSS study, is that this can avoid the cost of evaluating multiple models while retaining useful correlation with actual error.

3. PAUP in surrogate-assisted inverse design for frequency-selective surfaces

The FSS formulation is built around a multi-fidelity optimization problem over a very large binary design space. The meta-atom is $5.4 \times 5.4\ \mathrm{mm}^2$, represented by $18 \times 18$ binary pixels with eight-fold symmetry, which reduces the search to 45 design bits, for a total space of approximately $2^{45}$ designs. The operational band is 20–30 GHz with 100 uniformly sampled frequency points. The low-fidelity (LF) surrogate is a WideResNet with 3 groups, repeated $T$0 times, each group following the sequence Conv2D $T$1 Dropout $T$2 Conv2D $T$3 BatchNorm $T$4 ReLU, and it predicts $T$5, $T$6, $T$7, and $T$8 for $T$9 polarization; $xx$0 by symmetry. HFSS was used to generate 10,000 random eight-fold symmetric designs, with 1–10 minutes per design to convergence $xx$1, corresponding to approximately 2 months of generation time. The optimization loop itself uses an in-house PEEC solver validated against HFSS (Xue et al., 26 Jan 2026).

The paper defines residual-based physics-informed uncertainty as

$xx$2

with

$xx$3

The model outputs are defined over the 100 sampled frequencies. The paper describes applying the residual per frequency and aggregating across the band; a frequency-averaged form consistent with the description is reconstructed as

$xx$4

where $xx$5. Ensemble-based uncertainty is also defined for comparison: $xx$6 with $xx$7.

PAUP-like pruning is realized through explicit decision rules. The basic rule is thresholding: $xx$8 The paper synthesis also describes adaptive $xx$9 strategies: percentile-based pruning of the top 20–40% highest-$$1 + \mathrm{Re}(R) = \mathrm{Re}(T), \qquad \mathrm{Im}(R) = \mathrm{Im}(T).$$0 candidates for exploitation, or a calibrated mapping from $$1 + \mathrm{Re}(R) = \mathrm{Re}(T), \qquad \mathrm{Im}(R) = \mathrm{Im}(T).$$1 to expected error. Exploitation selects among low-$$1 + \mathrm{Re}(R) = \mathrm{Re}(T), \qquad \mathrm{Im}(R) = \mathrm{Im}(T).$$2, low-$$1 + \mathrm{Re}(R) = \mathrm{Re}(T), \qquad \mathrm{Im}(R) = \mathrm{Im}(T).$$3 candidates, whereas exploration periodically sends promising high-$$1 + \mathrm{Re}(R) = \mathrm{Re}(T), \qquad \mathrm{Im}(R) = \mathrm{Im}(T).$$4 candidates for HF evaluation.

Optimization is conducted with Binary PSO (BPSO) using the S-shaped transfer

$$1 + \mathrm{Re}(R) = \mathrm{Re}(T), \qquad \mathrm{Im}(R) = \mathrm{Im}(T).$$5

over 30 iterations. The staged multi-fidelity workflow has two phases. First is a constant-attraction stage of 10 iterations, where candidates are evaluated by the surrogate, pruned by uncertainty, and one selected design is sent to HF evaluation; the resulting HF-evaluated design is added to a beacon set $$1 + \mathrm{Re}(R) = \mathrm{Re}(T), \qquad \mathrm{Im}(R) = \mathrm{Im}(T).$$6 with weight

$$1 + \mathrm{Re}(R) = \mathrm{Re}(T), \qquad \mathrm{Im}(R) = \mathrm{Im}(T).$$7

Second is an alternating-strategy stage of 20 iterations with velocity update

$$1 + \mathrm{Re}(R) = \mathrm{Re}(T), \qquad \mathrm{Im}(R) = \mathrm{Im}(T).$$8

where

$$1 + \mathrm{Re}(R) = \mathrm{Re}(T), \qquad \mathrm{Im}(R) = \mathrm{Im}(T).$$9

If there is no improvement in minimum $I(x) = D(x) + B(x),$0 for 2 iterations, the scheme switches to exploration; otherwise it uses exploitation. Only one HF evaluation is performed per iteration.

The study also reports optional compatible extensions not used in the paper’s metric: energy conservation $I(x) = D(x) + B(x),$1, reciprocity $I(x) = D(x) + B(x),$2, and passivity via maximum singular value $I(x) = D(x) + B(x),$3. These can be incorporated into a generic residual template

$I(x) = D(x) + B(x),$4

This suggests that, in inverse design, PAUP functions as a residual-gated search mechanism that intervenes before the optimizer can overcommit to false minima.

4. PAUP in underwater 3D Gaussian Splatting

UW-3DGS represents the scene with anisotropic Gaussians $I(x) = D(x) + B(x),$5 having mean $I(x) = D(x) + B(x),$6, covariance $I(x) = D(x) + B(x),$7 decomposed as

$I(x) = D(x) + B(x),$8

opacity $I(x) = D(x) + B(x),$9, and view-dependent color $$I = J \cdot e^{-\beta^D \cdot z} + B^\infty \cdot \left(1 - e^{-\beta^B \cdot z}\right).$$0 parameterized with spherical harmonics. The base renderer uses front-to-back alpha compositing,

$$I = J \cdot e^{-\beta^D \cdot z} + B^\infty \cdot \left(1 - e^{-\beta^B \cdot z}\right).$$1

with cumulative transmittance

$$I = J \cdot e^{-\beta^D \cdot z} + B^\infty \cdot \left(1 - e^{-\beta^B \cdot z}\right).$$2

Depth is also composited, but uncertain Gaussians are down-weighted: $$I = J \cdot e^{-\beta^D \cdot z} + B^\infty \cdot \left(1 - e^{-\beta^B \cdot z}\right).$$3 The physics module uses voxel grids $$I = J \cdot e^{-\beta^D \cdot z} + B^\infty \cdot \left(1 - e^{-\beta^B \cdot z}\right).$$4 with $$I = J \cdot e^{-\beta^D \cdot z} + B^\infty \cdot \left(1 - e^{-\beta^B \cdot z}\right).$$5 and vector–matrix low-rank decomposition of rank $$I = J \cdot e^{-\beta^D \cdot z} + B^\infty \cdot \left(1 - e^{-\beta^B \cdot z}\right).$$6, queried via trilinear interpolation to obtain $$I = J \cdot e^{-\beta^D \cdot z} + B^\infty \cdot \left(1 - e^{-\beta^B \cdot z}\right).$$7 and $$I = J \cdot e^{-\beta^D \cdot z} + B^\infty \cdot \left(1 - e^{-\beta^B \cdot z}\right).$$8. The resulting underwater rendering is

$$I = J \cdot e^{-\beta^D \cdot z} + B^\infty \cdot \left(1 - e^{-\beta^B \cdot z}\right).$$9

where $$I_c(x) = J_c(x)\, t_c(x) + B_c(x)\,(1 - t_c(x)), \qquad t_c(x) = \exp(-\beta_c(x)\, d(x)).$$0 is the URI after PAUP pruning (Xing et al., 8 Aug 2025).

The PAUP branch computes a per-Gaussian Physics-aware Uncertainty Score

$$I_c(x) = J_c(x)\, t_c(x) + B_c(x)\,(1 - t_c(x)), \qquad t_c(x) = \exp(-\beta_c(x)\, d(x)).$$1

with learnable weights $$I_c(x) = J_c(x)\, t_c(x) + B_c(x)\,(1 - t_c(x)), \qquad t_c(x) = \exp(-\beta_c(x)\, d(x)).$$2 initialized to $$I_c(x) = J_c(x)\, t_c(x) + B_c(x)\,(1 - t_c(x)), \qquad t_c(x) = \exp(-\beta_c(x)\, d(x)).$$3. The rendering-instability term is

$$I_c(x) = J_c(x)\, t_c(x) + B_c(x)\,(1 - t_c(x)), \qquad t_c(x) = \exp(-\beta_c(x)\, d(x)).$$4

where $$I_c(x) = J_c(x)\, t_c(x) + B_c(x)\,(1 - t_c(x)), \qquad t_c(x) = \exp(-\beta_c(x)\, d(x)).$$5, $$I_c(x) = J_c(x)\, t_c(x) + B_c(x)\,(1 - t_c(x)), \qquad t_c(x) = \exp(-\beta_c(x)\, d(x)).$$6, and $$I_c(x) = J_c(x)\, t_c(x) + B_c(x)\,(1 - t_c(x)), \qquad t_c(x) = \exp(-\beta_c(x)\, d(x)).$$7 neighboring views are used. Effective opacity is

$$I_c(x) = J_c(x)\, t_c(x) + B_c(x)\,(1 - t_c(x)), \qquad t_c(x) = \exp(-\beta_c(x)\, d(x)).$$8

The physics inconsistency term is

$$I_c(x) = J_c(x)\, t_c(x) + B_c(x)\,(1 - t_c(x)), \qquad t_c(x) = \exp(-\beta_c(x)\, d(x)).$$9

The second term encodes the stated intuition that a highly opaque Gaussian at substantial depth, where strong attenuation should apply, is physically inconsistent and indicative of a floating splat.

PUS is converted to a pruning probability via a 2-layer MLP $5.4 \times 5.4\ \mathrm{mm}^2$0 with 32 hidden units: $5.4 \times 5.4\ \mathrm{mm}^2$1 Pruning uses an adaptive threshold $5.4 \times 5.4\ \mathrm{mm}^2$2: $5.4 \times 5.4\ \mathrm{mm}^2$3 The paper notes percentile-based adaptation per iteration; the PAUP description mentions the 95% percentile, whereas the implementation states the median of $5.4 \times 5.4\ \mathrm{mm}^2$4. In practice, pruned Gaussians are masked from rendering, equivalently setting $5.4 \times 5.4\ \mathrm{mm}^2$5 for those Gaussians.

Training minimizes

$5.4 \times 5.4\ \mathrm{mm}^2$6

with $5.4 \times 5.4\ \mathrm{mm}^2$7, $5.4 \times 5.4\ \mathrm{mm}^2$8, and $5.4 \times 5.4\ \mathrm{mm}^2$9. The underwater image loss is

$18 \times 18$0

with $18 \times 18$1. The Physics-Aware Pruning Supervision Loss is

$18 \times 18$2

with $18 \times 18$3 and $18 \times 18$4. Attenuation regression uses PUS-weighted supervision and low-r