MP-AG: Physics-Inspired 3D Defect Generator

• MP-AG is a physics-inspired module that converts normal 3D point clouds into synthetic defects using parameterized deformation operators.
• It employs five defect primitives—bulge, concavity, hole, crack, and bend—to create realistic, morphologically plausible anomalies.
• Integration with prototype guidance and local-global attention in PA3AD yields superior offset prediction and defect detection performance.

Searching arXiv for the cited works and related MP-AG context. Multi-Physics Anomaly Generator (MP-AG) denotes the physics-inspired pseudo-anomaly generation module introduced within PA3AD, a 3D point-cloud anomaly-detection framework for industrial manufacturing. In that setting, MP-AG transforms an anomaly-free point cloud into a pseudo-anomalous but physically plausible defective version by applying one of several parameterized deformation operators derived from qualitative physical behavior, while also providing a per-point offset field used as supervision for regression-based anomaly scoring. The paper explicitly uses the label “MP‑AG” for this module, although it does not give a formal acronym expansion; operationally, it is the mechanism that supplies synthetic defect geometry when real anomalous samples are scarce (Ning et al., 12 Jul 2026).

1. Definition and role within PA3AD

In PA3AD, training begins from a normal point cloud

$P_{\text{normal}} \in \mathbb{R}^{N \times 3}.$

MP-AG receives this normal cloud, samples one of five physics-inspired deformation models, and produces a pseudo-anomalous point cloud

$P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$

together with implicit per-point offsets

$o_i = p_i' - p_i.$

These synthetic offsets become the target for an offset-prediction head, so the generator is not merely augmentation; it is the direct source of supervision that teaches the detector how normal geometry transitions into defective geometry (Ning et al., 12 Jul 2026).

The full PA3AD training pipeline couples MP-AG to a shared-weight backbone, prototype guidance, and a difference-aware fusion block. A sparse 3D encoder–decoder with local-global attention extracts features from both $P_{\text{normal}}$ and $P_{\text{anomaly}}$ using the same weights. A momentum-updated prototype stores a stable normal reference. A fusion-and-regression stage then predicts per-point offsets $\hat{o}_i \in \mathbb{R}^3$ from feature pairs and their differences. Point-level anomaly scores are defined as

$$s(p_i) = |\hat{o}_{i,x}| + |\hat{o}_{i,y}| + |\hat{o}_{i,z}|,$$

and the object-level score is

$s^{(o)}(P) = \frac{1}{N}\sum_i s(p_i).$

A crucial operational detail is that MP-AG is used only during training. At inference, only a test cloud is available. The saved normal prototype replaces the normal branch, and predicted offsets against that reference yield anomaly heatmaps and object scores. This directly counters a common misunderstanding: the deployed detector does not require synthetic anomaly generation at test time.

2. Defect primitives and deformation laws

MP-AG takes as input a normal point cloud

$P = \{p_i \in \mathbb{R}^3\}_{i=1}^N$

with estimated normals $n_i$ and local structure, and outputs a transformed cloud

$P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$0

plus offsets $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$1. It samples one of five local defect primitives: bulge, concavity, hole, crack, and bend. During training, the paper uses weighted random sampling over these modes: bulge $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$2, concavity $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$3, hole $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$4, crack $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$5, and bend $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$6 (Ning et al., 12 Jul 2026).

Defect primitive	Deformation principle	Sampling weight
Bulge	Gaussian RBF displacement along surface normals	0.35
Concavity	Gaussian RBF displacement along surface normals with negative magnitude	0.35
Hole	Material removal plus smooth inward sinking	0.10
Crack	PCA-aligned linear separation	0.10
Bend	Euler–Bernoulli beam-inspired regional deflection	0.10

Bulges and concavities are modeled by a Gaussian radial basis function aligned with surface normals: $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$7 where $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$8 is the deformation center, $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$9 is the unit surface normal, $o_i = p_i' - p_i.$0 is the perturbation magnitude, and $o_i = p_i' - p_i.$1 controls the spatial extent. Positive $o_i = p_i' - p_i.$2 yields a bulge and negative $o_i = p_i' - p_i.$3 yields a concavity. The paper interprets this as an elastic displacement field under localized load: maximal displacement near the impact point with smooth radial decay.

The hole primitive combines a smooth sink with topology change: $o_i = p_i' - p_i.$4 followed by removal of points satisfying

$o_i = p_i' - p_i.$5

Here $o_i = p_i' - p_i.$6 is the sinking depth, $o_i = p_i' - p_i.$7 is a smoothing coefficient, and $o_i = p_i' - p_i.$8 is the hole radius. The use of $o_i = p_i' - p_i.$9 is intended to model a compact central sink with smooth transition, approximating a crater-like profile from material removal.

The crack primitive approximates fracture opening along a locally dominant direction. A crack center $P_{\text{normal}}$0 is chosen, a $P_{\text{normal}}$1-NN neighborhood is extracted, PCA is computed, and the eigenvector $P_{\text{normal}}$2 with largest eigenvalue is used as an approximate principal stress direction. The cloud is then split by

$P_{\text{normal}}$3

where $P_{\text{normal}}$4 is the crack width. This is explicitly presented as a simple kinematic fracture model rather than a full stress simulation.

The bend primitive is inspired by Euler–Bernoulli beam bending: $P_{\text{normal}}$5 with

$P_{\text{normal}}$6

Here $P_{\text{normal}}$7 is the bending strength, $P_{\text{normal}}$8 is the bending axis direction, $P_{\text{normal}}$9 is the vector from $P_{\text{anomaly}}$0 to the axis orthogonal component, $P_{\text{anomaly}}$1 is a smoothing weight, and $P_{\text{anomaly}}$2 provides numerical stability. The intent is to reproduce smooth curvature change over a region without solving PDEs.

3. Physical plausibility and diversity control

The defining claim of MP-AG is that it is physics-inspired and morphology-aware rather than random. The paper emphasizes several mechanisms for physical plausibility. Bulge and concavity deformations are continuous and normal-direction aligned, so the geometry remains smooth and manifold-like. Hole generation combines point removal with a surrounding elastic-like relaxation. Crack direction is chosen from local PCA rather than arbitrary coordinate perturbation. Bend is smooth and distance-weighted around an axis. The generator therefore avoids random per-point noise and instead encodes defect-specific morphologies associated with stamping, casting, and injection molding (Ning et al., 12 Jul 2026).

Diversity is controlled at three levels. First, there is categorical randomness over defect types. Second, each primitive has its own parameter set: $P_{\text{anomaly}}$3, $P_{\text{anomaly}}$4, and center $P_{\text{anomaly}}$5 for bulge or concavity; $P_{\text{anomaly}}$6, $P_{\text{anomaly}}$7, $P_{\text{anomaly}}$8, and $P_{\text{anomaly}}$9 for holes; $\hat{o}_i \in \mathbb{R}^3$0 and local PCA neighborhoods for cracks; $\hat{o}_i \in \mathbb{R}^3$1, axis $\hat{o}_i \in \mathbb{R}^3$2, region definition, and $\hat{o}_i \in \mathbb{R}^3$3 for bending. Third, there is spatial randomness through the sampled deformation center, bending axis, and neighborhood geometry.

The paper reports default deformation parameters sampled from $\hat{o}_i \in \mathbb{R}^3$4 and studies a uniform scaling factor $\hat{o}_i \in \mathbb{R}^3$5 applied to all parameters. On selected categories, the reported O‑AUROC values are:

• $\hat{o}_i \in \mathbb{R}^3$6: headset0 $\hat{o}_i \in \mathbb{R}^3$7, jar0 $\hat{o}_i \in \mathbb{R}^3$8, fish $\hat{o}_i \in \mathbb{R}^3$9, shell $$s(p_i) = |\hat{o}_{i,x}| + |\hat{o}_{i,y}| + |\hat{o}_{i,z}|,$$0
• $$s(p_i) = |\hat{o}_{i,x}| + |\hat{o}_{i,y}| + |\hat{o}_{i,z}|,$$1: headset0 $$s(p_i) = |\hat{o}_{i,x}| + |\hat{o}_{i,y}| + |\hat{o}_{i,z}|,$$2, jar0 $$s(p_i) = |\hat{o}_{i,x}| + |\hat{o}_{i,y}| + |\hat{o}_{i,z}|,$$3, fish $$s(p_i) = |\hat{o}_{i,x}| + |\hat{o}_{i,y}| + |\hat{o}_{i,z}|,$$4, shell $$s(p_i) = |\hat{o}_{i,x}| + |\hat{o}_{i,y}| + |\hat{o}_{i,z}|,$$5
• $$s(p_i) = |\hat{o}_{i,x}| + |\hat{o}_{i,y}| + |\hat{o}_{i,z}|,$$6: headset0 $$s(p_i) = |\hat{o}_{i,x}| + |\hat{o}_{i,y}| + |\hat{o}_{i,z}|,$$7, jar0 $$s(p_i) = |\hat{o}_{i,x}| + |\hat{o}_{i,y}| + |\hat{o}_{i,z}|,$$8, fish $$s(p_i) = |\hat{o}_{i,x}| + |\hat{o}_{i,y}| + |\hat{o}_{i,z}|,$$9, shell $s^{(o)}(P) = \frac{1}{N}\sum_i s(p_i).$0
• $s^{(o)}(P) = \frac{1}{N}\sum_i s(p_i).$1: headset0 $s^{(o)}(P) = \frac{1}{N}\sum_i s(p_i).$2, jar0 $s^{(o)}(P) = \frac{1}{N}\sum_i s(p_i).$3, fish $s^{(o)}(P) = \frac{1}{N}\sum_i s(p_i).$4, shell $s^{(o)}(P) = \frac{1}{N}\sum_i s(p_i).$5

This establishes a central empirical property of MP-AG: too small a deformation makes pseudo anomalies nearly normal, while too large a deformation makes them unrealistic. The generator is therefore best understood as a parameterized family of qualitative physics proxies whose useful regime is bounded by both geometric plausibility and downstream detection performance.

4. Coupling to prototypes, fusion, and loss design

MP-AG in PA3AD is inseparable from prototype-guided feature learning. The framework maintains a single normal prototype $s^{(o)}(P) = \frac{1}{N}\sum_i s(p_i).$6 as an exponential moving average of batchwise normal features: $s^{(o)}(P) = \frac{1}{N}\sum_i s(p_i).$7 where $s^{(o)}(P) = \frac{1}{N}\sum_i s(p_i).$8 is the mean normal feature of the current batch and $s^{(o)}(P) = \frac{1}{N}\sum_i s(p_i).$9 is the step-dependent momentum. The schedule starts around $P = \{p_i \in \mathbb{R}^3\}_{i=1}^N$0 and increases linearly to approximately $P = \{p_i \in \mathbb{R}^3\}_{i=1}^N$1, so early training emphasizes adaptability and late training emphasizes stability. The paper explicitly interprets this EMA as a low-pass filter over minibatch noise (Ning et al., 12 Jul 2026).

The difference-aware fusion block receives either $P = \{p_i \in \mathbb{R}^3\}_{i=1}^N$2 during training or $P = \{p_i \in \mathbb{R}^3\}_{i=1}^N$3 at inference. It forms a concatenated feature

$P = \{p_i \in \mathbb{R}^3\}_{i=1}^N$4

and a raw difference

$P = \{p_i \in \mathbb{R}^3\}_{i=1}^N$5

The fused representation is

$P = \{p_i \in \mathbb{R}^3\}_{i=1}^N$6

while the weighting branch computes

$P = \{p_i \in \mathbb{R}^3\}_{i=1}^N$7

The difference-enhanced feature is then

$P = \{p_i \in \mathbb{R}^3\}_{i=1}^N$8

which feeds an MLP that predicts offsets.

The total loss is

$P = \{p_i \in \mathbb{R}^3\}_{i=1}^N$9

with $n_i$0 and $n_i$1. The offset norm loss is

$n_i$2

the offset direction loss is

$n_i$3

the feature discrepancy regularizer is

$n_i$4

and the weight regularizer is

$n_i$5

A recurrent misconception is that MP-AG alone explains PA3AD’s performance. The reported ablations indicate otherwise. Pseudo anomalies without prototype guidance already provide a useful baseline, but the largest additional gains arise when synthetic deformation is combined with stable normal-reference estimation and difference-aware feature separation.

5. Backbone, training protocol, and empirical performance

The feature extractor used with MP-AG is a sparse 3D encoder–decoder with a local-global attention bottleneck. Point clouds are voxelized as

$n_i$6

with voxel size $n_i$7, non-empty voxel coordinates $n_i$8, and aggregated voxel features $n_i$9. At the bottleneck, two transformer layers are inserted: $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$00

$P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$01

Local attention operates on overlapping cubic windows using relative positional embeddings, while the global branch applies a linear-attention formulation

$P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$02

reducing complexity from $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$03 to $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$04. The intended division of labor is explicit: local attention is sensitive to small high-curvature defects, whereas global attention stabilizes detection under viewpoint and shape variation (Ning et al., 12 Jul 2026).

The implementation details reported for PA3AD are specific. The datasets are Anomaly-ShapeNet, Real3D-AD, and MVTec 3D-AD. For each normal training sample, one pseudo anomaly is generated via MP-AG. Training uses voxel size $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$05, AdamW with learning rate $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$06, batch size $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$07, and $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$08 epochs, with no pretraining.

Extensive experiments are summarized in three ablation families. First, MP-AG is compared to three simpler synthetic strategies under the same network: random point deletion, random displacement, and Gaussian noise. On selected classes, the reported AUROC values are:

• Random deletion: bucket0 $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$09, vase0 $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$10, car $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$11, seahorse $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$12
• Random displacement: bucket0 $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$13, vase0 $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$14, car $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$15, seahorse $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$16
• Gaussian noise: bucket0 $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$17, vase0 $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$18, car $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$19, seahorse $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$20
• MP‑AG: bucket0 $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$21, vase0 $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$22, car $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$23, seahorse $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$24

Second, the prototype-guidance ablation reports:

• $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$25: pseudo anomalies only — AnomalyShapeNet O‑AUROC $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$26, P‑AUROC $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$27; Real3D O‑AUROC $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$28, P‑AUROC $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$29
• $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$30: $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$31 prototype guidance — AnomalyShapeNet O‑AUROC $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$32, P‑AUROC $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$33; Real3D O‑AUROC $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$34, P‑AUROC $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$35
• $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$36: $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$37 prototype guidance $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$38 momentum schedule — AnomalyShapeNet O‑AUROC $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$39, P‑AUROC $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$40; Real3D O‑AUROC $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$41, P‑AUROC $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$42

Third, local-global attention yields further improvements. On selected categories, the full local+global model reaches cap3 $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$43, vase3 $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$44, fish $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$45, and seahorse $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$46, compared with a no-attention baseline of $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$47, $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$48, $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$49, and $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$50, respectively.

6. Broader interpretations and related formulations

The strict sense of MP-AG in the literature provided here is the PA3AD module for 3D point clouds. However, adjacent work suggests broader interpretations of the term as a design pattern for anomaly generation or anomaly-grounded reasoning across multiple physical regimes.

In physics-grounded video anomaly detection, a multi-turn VLM framework represents object physics by the tuple

$P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$51

where $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$52 denotes component recognition, $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$53 denotes dynamic subject focusing, and $P_{\text{anomaly}} = \{p_i'\}_{i=1}^{N},$54 denotes motion specification. That work defines anomalies as violations of physical laws and constraints, such as irregular rotations or violated mechanical motions, rather than appearance defects. This suggests a temporal generalization of MP-AG in which anomaly generation would target violations of expected dynamics over time rather than purely geometric surface defects (Gu et al., 16 Mar 2026).

In calorimeter data-quality monitoring, Lorenzetti-based synthetic data generation injects two defect families during digitization: dead modules, implemented by setting cell signals to zero over selected modules and time windows, and increased noise, implemented by multiplying the noise standard deviation by a configurable factor in selected cells. The resulting multivariate time series is then used to benchmark TSAD methods. This suggests another plausible extension of MP-AG: a modular anomaly generator acting at the sensor or electronics layer, with downstream effects observed only after reconstruction and feature aggregation (Boggia et al., 9 Sep 2025).

In collider anomaly detection, conditional generative models are trained on sidebands to synthesize background in a signal region, first in full phase-space resonant anomaly detection and then in the more general Generator Based Inference framework. There the generator is learned from data and becomes part of a likelihood or likelihood-ratio construction. These works do not define MP-AG in the PA3AD sense, but they suggest a broader, generator-centric view in which an anomaly generator need not only create synthetic defects; it can also emulate normal background distributions against which anomalies are scored and interpreted statistically (Buhmann et al., 2023, Cheng et al., 30 May 2025).

Under that broader reading, MP-AG can be understood as an architectural principle with three recurring elements: a physically motivated parameterization of perturbations or processes, a controllable sampling scheme over anomaly or background modes, and a downstream detector that uses the generated data either as direct supervision, as in PA3AD, or as a reference distribution for likelihood-ratio-style inference. The PA3AD instantiation remains the most explicit and fully specified realization of that principle in the present corpus.