---
title: 'MP-AG: Physics-Inspired 3D Defect Generator'
url: https://www.emergentmind.com/topics/multi-physics-anomaly-generator-mp-ag
type: topic
---

# MP-AG: Physics-Inspired 3D Defect Generator

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 [2607.10544].

## 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 [2607.10544].

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
\[
P' = \{p_i'\}_{i=1}^N
\]
plus offsets \(o_i = p_i' - p_i\). 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 \(0.35\), concavity \(0.35\), hole \(0.10\), crack \(0.10\), and bend \(0.10\) [2607.10544].

| 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:
\[
p_i' = p_i + \alpha \cdot \exp\left(-\frac{\|p_i-c\|^2}{2\sigma_{bh}^2}\right)\cdot n_i,
\]
where \(c\) is the deformation center, \(n_i\) is the unit surface normal, \(\alpha\) is the perturbation magnitude, and \(\sigma_{bh}\) controls the spatial extent. Positive \(\alpha\) yields a bulge and negative \(\alpha\) 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:
\[
p_i' = p_i - \beta \cdot \tanh\left(\frac{\|p_i-c\|}{\sigma_h}\right)\cdot n_i,
\]
followed by removal of points satisfying
\[
\|p_i-c\| < r_h.
\]
Here \(\beta\) is the sinking depth, \(\sigma_h\) is a smoothing coefficient, and \(r_h\) is the hole radius. The use of \(\tanh\) 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 \(c\) is chosen, a \(k\)-NN neighborhood is extracted, PCA is computed, and the eigenvector \(v\) with largest eigenvalue is used as an approximate principal stress direction. The cloud is then split by
\[
p_i'=
\begin{cases}
p_i + \gamma \cdot v, & \text{if } (p_i-c)^T v > 0,\\
p_i - \gamma \cdot v, & \text{otherwise},
\end{cases}
\]
where \(\gamma\) 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:
\[
p_i' = p_i + \eta \cdot f_{\text{bend}}(p_i,c,a),
\]
with
\[
f_{\text{bend}}(p_i,a) =
w_i \,\frac{a \times d_{i,\perp}}{\|d_{i,\perp}\|+\epsilon}.
\]
Here \(\eta\) is the bending strength, \(a\) is the bending axis direction, \(d_{i,\perp}\) is the vector from \(p_i\) to the axis orthogonal component, \(w_i\) is a smoothing weight, and \(\epsilon\) 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 [2607.10544].

Diversity is controlled at three levels. First, there is categorical randomness over defect types. Second, each primitive has its own parameter set: \(\alpha\), \(\sigma_{bh}\), and center \(c\) for bulge or concavity; \(\beta\), \(\sigma_h\), \(r_h\), and \(c\) for holes; \(\gamma\) and local PCA neighborhoods for cracks; \(\eta\), axis \(a\), region definition, and \(w_i\) 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 \([0.08, 0.24]\) and studies a uniform scaling factor \(k\) applied to all parameters. On selected categories, the reported O‑AUROC values are:

- \(0.5\times\): headset0 \(0.946\), jar0 \(0.903\), fish \(0.912\), shell \(0.753\)
- \(1.0\times\): headset0 \(1.000\), jar0 \(0.952\), fish \(0.954\), shell \(0.808\)
- \(2.0\times\): headset0 \(0.982\), jar0 \(0.928\), fish \(0.937\), shell \(0.784\)
- \(5.0\times\): headset0 \(0.843\), jar0 \(0.781\), fish \(0.773\), shell \(0.641\)

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 \(z^{(t)} \in \mathbb{R}^d\) as an exponential moving average of batchwise normal features:
\[
z^{(t+1)} = \mu(t)\,z^{(t)} + (1-\mu(t))\,f^{(t)},
\]
where \(f^{(t)}\) is the mean normal feature of the current batch and \(\mu(t)\in[0,1)\) is the step-dependent momentum. The schedule starts around \(0.1\) and increases linearly to approximately \(0.999\), so early training emphasizes adaptability and late training emphasizes stability. The paper explicitly interprets this EMA as a low-pass filter over minibatch noise [2607.10544].

The difference-aware fusion block receives either \((F_{\text{normal}}, F_{\text{anomaly}})\) during training or \((z^{(T)}, F_{\text{test}})\) at inference. It forms a concatenated feature
\[
F_{\text{concat}} = [F_{\text{normal}}, F_{\text{anomaly}}] \in \mathbb{R}^{N\times 2d}
\]
and a raw difference
\[
\Delta F = F_{\text{anomaly}} - F_{\text{normal}} \in \mathbb{R}^{N\times d}.
\]
The fused representation is
\[
F_{\text{fuse}} = \text{Linear}\big(\text{ReLU}(\text{Linear}(F_{\text{concat}}))\big),
\]
while the weighting branch computes
\[
w_{\text{diff}} = \text{Sigmoid}\big(\text{Linear}(\text{ReLU}(\text{Linear}(\Delta F)))\big).
\]
The difference-enhanced feature is then
\[
F_{\text{diff}}^{\text{enh}} = F_{\text{fuse}} \otimes w_{\text{diff}},
\]
which feeds an MLP that predicts offsets.

The total loss is
\[
L = L_{\text{norm}} + L_{\text{dir}} + \lambda_1 L_{\text{feat}} + \lambda_2 L_{\text{weight}},
\]
with \(\lambda_1 = 0.01\) and \(\lambda_2 = 0.001\). The offset norm loss is
\[
L_{\text{norm}} = \frac{1}{N}\sum_{i=1}^{N}\left(\|\hat{o}_i-o_i\|_1 + \|\hat{o}_i-o_i\|_2\right),
\]
the offset direction loss is
\[
L_{\text{dir}} = \frac{1}{N}\sum_{i=1}^{N}\left(1-\frac{\hat{o}_i\cdot o_i}{\|\hat{o}_i\|_2\|o_i\|_2+\varepsilon}\right),
\quad \varepsilon=10^{-8},
\]
the feature discrepancy regularizer is
\[
L_{\text{feat}} = \exp\left(-\frac{1}{N}\sum_{i=1}^{N}\|F_n^{(i)}-F_a^{(i)}\|_2^2\right),
\]
and the weight regularizer is
\[
L_{\text{weight}} = \frac{1}{N\cdot d}\sum_{i=1}^{N}\sum_{j=1}^{d}|w_{\text{diff}}^{(i,j)}-\alpha|,
\quad \alpha=0.5.
\]

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
\[
X = V(P,s) = (C,G),
\]
with voxel size \(s\), non-empty voxel coordinates \(C\), and aggregated voxel features \(G\). At the bottleneck, two transformer layers are inserted:
\[
Z_i^{\text{att}} = Z_0 + \text{Attn}(\text{LN}(Z_0)),
\]
\[
Z_{i+1} = Z_i^{\text{att}} + \text{MLP}(\text{LN}(Z_i^{\text{att}})).
\]
Local attention operates on overlapping cubic windows using relative positional embeddings, while the global branch applies a linear-attention formulation
\[
Z_{\text{global}} = D^{-1}(\phi(Q)\phi(K)^T V),
\]
reducing complexity from \(O(M^2)\) to \(O(M)\). 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 [2607.10544].

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 \(0.05\), AdamW with learning rate \(1\times 10^{-3}\), batch size \(32\), and \(2000\) 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 \(0.657\), vase0 \(0.689\), car \(0.468\), seahorse \(0.527\)
- Random displacement: bucket0 \(0.711\), vase0 \(0.718\), car \(0.402\), seahorse \(0.473\)
- Gaussian noise: bucket0 \(0.783\), vase0 \(0.802\), car \(0.531\), seahorse \(0.614\)
- MP‑AG: bucket0 \(0.968\), vase0 \(0.975\), car \(0.769\), seahorse \(0.875\)

Second, the prototype-guidance ablation reports:

- \(V_1\): pseudo anomalies only — AnomalyShapeNet O‑AUROC \(0.874\), P‑AUROC \(0.865\); Real3D O‑AUROC \(0.724\), P‑AUROC \(0.772\)
- \(V_2\): \(+\) prototype guidance — AnomalyShapeNet O‑AUROC \(0.921\), P‑AUROC \(0.915\); Real3D O‑AUROC \(0.768\), P‑AUROC \(0.815\)
- \(V_3\): \(+\) prototype guidance \(+\) momentum schedule — AnomalyShapeNet O‑AUROC \(0.936\), P‑AUROC \(0.924\); Real3D O‑AUROC \(0.789\), P‑AUROC \(0.837\)

Third, local-global attention yields further improvements. On selected categories, the full local+global model reaches cap3 \(0.933\), vase3 \(0.918\), fish \(0.954\), and seahorse \(0.875\), compared with a no-attention baseline of \(0.906\), \(0.889\), \(0.921\), and \(0.832\), 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
\[
\mathcal{P}_c = (S_{com}, S_{dyn}, S_{mot}),
\]
where \(S_{com}\) denotes component recognition, \(S_{dyn}\) denotes dynamic subject focusing, and \(S_{mot}\) 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 [2603.15237].

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 [2509.07451].

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 [2310.06897] [2506.00119].

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.

Source: https://www.emergentmind.com/topics/multi-physics-anomaly-generator-mp-ag