Physics-Informed Detection Framework

• Physics-informed detection frameworks are systems that integrate physical laws and structured priors with learned models to improve anomaly detection.
• They employ techniques like residual encoding, latent partitioning, and multimodal fusion to capture domain-specific dynamics and enhance interpretability.
• Empirical results demonstrate significant performance gains in applications such as video anomaly detection, smart-grid security, and SAR-based diagnostics.

Searching arXiv for papers on physics-informed detection frameworks across anomaly detection and related inverse problems. Physics-informed detection framework denotes a class of inference systems that couple learned detectors with explicit physical structure—governing equations, conservation laws, kinematic or thermodynamic constraints, mechanistic simulators, or structured priors—to decide whether an observation is normal, anomalous, attacked, occupied, phase-transitioning, or boundary-containing. Recent literature instantiates this idea in video anomaly detection, multivariate time-series monitoring, traffic phase inference, smart-grid security, radar perception, SAR detection, battery diagnostics, and inverse flow or thermal reconstruction (Gu et al., 16 Mar 2026, Soni et al., 15 Aug 2025, Deng et al., 22 May 2026, Elnour et al., 16 Feb 2026, Zhu et al., 31 Mar 2025).

1. Definition, scope, and representative forms

The term does not refer to a single architecture. It instead covers a family of methods in which the detection objective is constrained or guided by domain physics. In some systems, physics enters as an explicit residual of an ODE or PDE. In others, it enters through structured prompts, latent priors, evolutionary rule operators, geometry-aware simulation, or signal-level invariants. The resulting detectors may be supervised, weakly supervised, or unsupervised, and they may score anomalies by classification confidence, reconstruction error, ELBO, residual energy, or equilibrium breakdown (Gu et al., 16 Mar 2026, Lin et al., 20 Oct 2025, Soni et al., 15 Aug 2025, Deng et al., 22 May 2026, Li et al., 2020).

Framework	Physical prior	Detection target
Multi-turn VLM (Gu et al., 16 Mar 2026)	Object properties, motion paradigms, dynamic constraints	Physics-grounded video anomaly
PILLM (Lin et al., 20 Oct 2025)	Thermodynamic and control-theoretic constraints	HVAC anomaly rules
PIGTN (Elnour et al., 16 Feb 2026)	AC power flow constraints	Smart-grid attack detection
SpinFlow (Deng et al., 22 May 2026)	Mass conservation and spatial smoothness	Traffic phase transition
PICAE (Li et al., 2020)	Elliptical voltage-current trajectory	High-impedance fault
Pi-Transformer (Maleki et al., 24 Sep 2025)	Prior attention for self-similarity and phase synchrony	Time-series anomaly

This breadth is important because it prevents an overly narrow identification of physics-informed detection with PINNs alone. The literature includes PINN-based inverse detection, but it also includes graph transformers, multimodal fusion systems, diffusion models, instruction-tuned VLMs, and sim-to-real digital twins (Rodriguez et al., 2024, Pessoa et al., 18 May 2026, Sharma et al., 8 Jun 2025, Trinh et al., 25 Jan 2026, Huang et al., 2024).

2. Sources of physical prior

A first major design axis is how physical knowledge is encoded. In the Multi-Turn Physics-Informed Vision-LLM, each object category is assigned a physics-prior tuple $$\mathcal{P}_c=(S_{\mathrm{com}},S_{\mathrm{dyn}},S_{\mathrm{mot}})$$, where the prompt strings describe components, governing laws such as $F=m_c\cdot a$ and $\tau=I_c\cdot\alpha$, and expected motion relations such as $\omega(t)=\omega_0+\alpha t$ and $v(t)=v_0+at$ (Gu et al., 16 Mar 2026). In PILLM, the prior is encoded as thermodynamic and control-theoretic constraints inside reflection and crossover prompts, including conditions such as $$\dot m\,c_p\,(T_{\mathrm{supply}-T_{\mathrm{return}})\ge 0}$$ and $0\le \text{damper\_position}\le 1$ (Lin et al., 20 Oct 2025). In Pi-Transformer, the prior is an attention distribution parameterized by a local Hurst exponent $H_i$ and timing-stiffness $\tau_i$, with affinity

$A^{(\ell,h)}_{i,j}=-\frac{|i-j|^{2H_i}}{\tau_i},$

so that nominal self-similarity and phase synchrony become part of the detector itself (Maleki et al., 24 Sep 2025).

A second axis is residual-based encoding. UTOPYA adds soft physics constraints for temporal smoothness and thermodynamic monotonicity through a combined loss

$F=m_c\cdot a$0

where monotonicity enforces a temperature ordering along the distillation column (Pessoa et al., 18 May 2026). PIGTN regularizes predicted phasors by AC power flow identities,

$F=m_c\cdot a$1

and penalizes deviations in regenerated active and reactive power (Elnour et al., 16 Feb 2026). Pi-DPM imposes the 3-DoF kinematic bicycle model through penalties on $F=m_c\cdot a$2, $F=m_c\cdot a$3, $F=m_c\cdot a$4, and $F=m_c\cdot a$5 (Sharma et al., 8 Jun 2025). PICAE uses the elliptical voltage-current trajectory of steady sinusoidal circuits as its physical regularizer, biasing reconstructed voltages to obey the same ellipse as the currents (Li et al., 2020).

A third axis is structural or geometric prior. PGD for SAR airplanes constructs scattering-structure heatmaps from Harris–Laplace corner detection and a Gaussian Mixture Model, then transfers those priors into detector features and instance-level supervision (Huang et al., 2024). FDTRImageEnhancer uses a Gaussian pump–probe point-spread abstraction and k-means clustering of structural images to reduce conductivity inversion to a small set of region-wise parameters (Odufisan, 8 Aug 2025). The FMCW radar digital twin uses a physics-informed geometric simulator and calibrated domain randomization to match the simulated RD-map noise floor to unlabeled real empty-room statistics while preserving micro-Doppler peaks (Trinh et al., 25 Jan 2026). This suggests that “physics-informed” may refer to physically plausible observation formation, not only to downstream decision rules.

3. Architectural patterns

Several recurrent architectural patterns appear across the literature. One is selective adaptation on top of a frozen perceptual backbone. The multi-turn VLM detector builds on Video-LLaVA with a frozen ViT and frozen multimodal projector, and fine-tunes only the Vicuna-7B LLM’s QKV and FFN weights while injecting physics priors through four dialogue turns (Gu et al., 16 Mar 2026). PGD similarly augments standard backbone–neck–head detectors with three modules—PGSSL, PGFE, and PGIP—rather than replacing the detector wholesale (Huang et al., 2024).

A second pattern is multimodal fusion conditioned by static context. UTOPYA encodes eight modalities separately, aggregates tabular, text, and molecular information into a context vector $F=m_c\cdot a$6, uses FiLM conditioning to modulate each dynamic embedding, applies a single layer of bidirectional cross-modal attention, and performs gated fusion into a shared latent $F=m_c\cdot a$7 (Pessoa et al., 18 May 2026). The explicit role of context is central: the reported multimodal ablation identifies static context via FiLM conditioning as the key enabler (Pessoa et al., 18 May 2026).

A third pattern is expert decomposition or latent partitioning. POU-PINN introduces partition-of-unity gating functions $F=m_c\cdot a$8 satisfying $F=m_c\cdot a$9, learns region-specific conductivity parameters $\tau=I_c\cdot\alpha$0, and feeds the composite field

$\tau=I_c\cdot\alpha$1

into a global PINN (Rodriguez et al., 2024). SpinFlow takes an analogous decomposition route for traffic states, but replaces spatial partitions by phase-mixture weights $\tau=I_c\cdot\alpha$2 derived from a latent spin vector $\tau=I_c\cdot\alpha$3 and a competitive-equilibrium softmax (Deng et al., 22 May 2026). In both cases, the detector is not merely classifying observations; it is first recovering a latent physical structure that organizes the decision problem.

A fourth pattern is search or optimization in the loop. PILLM embeds an LLM inside a “generate–evaluate–reflect–crossover–mutate” loop that returns the best rules by F1 after repeated physics-informed reflection and crossover (Lin et al., 20 Oct 2025). PIGTN couples the detector to NSGA-II so that sensor placement and detection performance are co-optimized in a closed loop (Elnour et al., 16 Feb 2026). Residual-loss anomaly analysis of PINNs uses a two-stage pipeline: overlapping subinterval decomposition for coarse localization and unified joint optimization for precise change-point and parameter identification (Bai et al., 28 Apr 2026). These methods treat detection as a coupled design problem rather than a fixed post hoc classifier.

4. Training objectives, inference, and scoring

The objective functions are typically composite. In the multi-turn VLM framework, training is a four-turn autoregressive instruction-tuning problem with loss

$\tau=I_c\cdot\alpha$4

where the first three turns inject physical priors and the fourth produces a verdict and causal explanation (Gu et al., 16 Mar 2026). In UTOPYA, the total loss is a weighted sum of prediction, classification, reconstruction, and physics terms, with the production model using $\tau=I_c\cdot\alpha$5, $\tau=I_c\cdot\alpha$6, $\tau=I_c\cdot\alpha$7, and $\tau=I_c\cdot\alpha$8 (Pessoa et al., 18 May 2026). PIGTN combines cross-entropy with AC-residual penalties as

$\tau=I_c\cdot\alpha$9

with $\omega(t)=\omega_0+\alpha t$0 and $\omega(t)=\omega_0+\alpha t$1 (Elnour et al., 16 Feb 2026).

Diffusion-based frameworks use a similar additive principle but attach the physics term to generative training. TPIDM adds a weighted physics residual to the diffusion loss, with a static schedule $\omega(t)=\omega_0+\alpha t$2 that emphasizes earlier, less noisy diffusion steps (Soni et al., 15 Aug 2025). Pi-DPM combines the diffusion VLB, a spatial reconstruction term, and the kinematic penalty in

$\omega(t)=\omega_0+\alpha t$3

using $\omega(t)=\omega_0+\alpha t$4 in the reported experiments (Sharma et al., 8 Jun 2025). Pi-Transformer alternates series and prior updates via two coupled losses, one subtracting and one adding the symmetric KL divergence between data-driven and prior attention while both retain reconstruction and prior regularization terms (Maleki et al., 24 Sep 2025).

Scoring mechanisms differ by task. The multi-turn VLM defines the anomaly score as the model confidence in “Verdict: Anomalous” or as a likelihood ratio between anomalous and normal verdicts (Gu et al., 16 Mar 2026). TPIDM scores windows by ELBO and thresholds with a trimmed mean plus $\omega(t)=\omega_0+\alpha t$5 rule (Soni et al., 15 Aug 2025). Pi-DPM uses reconstruction error $\omega(t)=\omega_0+\alpha t$6 against a user-specified threshold $\omega(t)=\omega_0+\alpha t$7 (Sharma et al., 8 Jun 2025). Pi-Transformer fuses an alignment-weighted reconstruction signal and a mismatch signal through a soft-OR, $\omega(t)=\omega_0+\alpha t$8 (Maleki et al., 24 Sep 2025). SpinFlow detects transitions by the minimum of the Phase Equilibrium Degree inside the phase-coexistence set, $\omega(t)=\omega_0+\alpha t$9 (Deng et al., 22 May 2026). PICAE performs local thresholding on the relative reconstruction error $v(t)=v_0+at$0 and then fuses node-level scores at system level (Li et al., 2020).

A common misconception is that physics-informed detection always requires labeled anomalies. Several frameworks explicitly avoid that regime. POU-PINN learns subdomains without labeled indicators (Rodriguez et al., 2024). TPIDM is trained on normal data and then flags low-support windows (Soni et al., 15 Aug 2025). Pi-DPM trains only on normal trajectories (Sharma et al., 8 Jun 2025). PICAE uses no labeled HIF examples in training (Li et al., 2020). The FMCW radar sim2real pipeline uses no real labeled data in training (Trinh et al., 25 Jan 2026).

5. Empirical performance and application domains

Reported results indicate that physically grounded inductive bias can materially change detector behavior. On Phys-AD, the multi-turn VLM achieves 96.7% AUROC in video-level detection versus prior SOTA 66.9%, reaches 0.777 LLM score for explanations, and ablation shows that single-turn concatenation fails on 8/21 categories while the multi-turn scheme recovers 0.967 average AUROC (Gu et al., 16 Mar 2026). On the public Building Fault Detection dataset, PILLM reaches Precision 0.968, Recall 0.859, and F1 0.926, while removing PIR or PIC drops F1 to 0.869 and 0.868 respectively (Lin et al., 20 Oct 2025). On batch distillation, UTOPYA reports a window-level test AUROC of 0.832 and 0.874 under multi-signal experiment-level scoring, with a +0.147 window-level AUROC gain over the best external baseline and a 0.717 to 0.663 drop when physics loss is removed in the ablation study (Pessoa et al., 18 May 2026).

Time-series generative detectors show similar patterns. TPIDM reports on Predator–Prey an F1 of 0.983 versus 0.971 for DM and 0.959 for PIDM, and on Lenze an F1 of 0.999 versus 0.999 for DM and 0.962 for PIDM (Soni et al., 15 Aug 2025). Pi-Transformer reports point-adjusted F1 scores of 91.23% on SMD, 95.96% on MSL, 97.02% on SMAP, 96.82% on SWaT, and 98.08% on PSM, with ablations indicating that removing the prior can collapse F1 by up to 60 points on SWaT (Maleki et al., 24 Sep 2025). For anomalous trajectories, Pi-DPM reports, at 5% anomalies, Geolife Accuracy 0.98 and F1 0.98, MarineCadastre Accuracy 0.97 and F1 0.97, and Danish Maritime Accuracy 0.98 and F1 0.98, compared with the best baseline at F1 approximately 0.91–0.92 (Sharma et al., 8 Jun 2025).

Beyond anomaly detection in the narrow sense, physics-informed detection is also used for structural or event inference. SpinFlow attains $v(t)=v_0+at$1 up to 0.940, PED drops of 94.9–100%, bottleneck localization T.MAE of 0–34 m, and runtime of 0.7–3.9 s on a single CPU, reported as 9× faster than a physics-informed DeepONet (Deng et al., 22 May 2026). In smart grids, PIGTN with optimized placement improves accuracy by up to 37%, detection rate by up to 73%, keeps mean false alarms around 0.3%, and yields up to 98% reduction in voltage-angle error and 61–94% reduction in magnitude error for state estimation (Elnour et al., 16 Feb 2026). PICAE, evaluated on the IEEE 34-node feeder, reports F1 values of 92.9%, 97.1%, 97.6%, and 100.0% at 30, 50, 70, and 90 dB SNR, outperforming AE98, PCA, and ellipse-only baselines (Li et al., 2020).

Image and field reconstruction tasks show the same principle in inverse settings. PGD for SAR airplanes reaches 90.7% mAP on SAR-AIRcraft-1.0 and improves existing detectors by as much as 3.1% mAP in fine-grained detection (Huang et al., 2024). FDTRImageEnhancer recovers bulk conductivity values within less than 0.5% error and reports learned bulk-region errors below 0.5% in the heterointerface case, although the grain-boundary conductivity estimate remains substantially biased because of limited resolution (Odufisan, 8 Aug 2025). The FMCW radar digital twin attains 97 percent accuracy for occupancy detection and 72 percent accuracy for people counting using models trained purely on CDR-adjusted simulation (Trinh et al., 25 Jan 2026). The transferable battery framework reports source-scenario phase-detection accuracy of 96%, fine-tuned target-scenario phase accuracy improving from 67.2% to 88.2%, and a knee-onset to knee-point regression $v(t)=v_0+at$2 with $v(t)=v_0+at$3 (Zhang et al., 24 Jan 2025). For hidden-boundary detection in fluid mechanics, the PINN framework reconstructs fixed-cylinder flow at sub-1% field errors and moving-cylinder or airfoil cases at the reported few-percent level under sparse and noisy data (Zhu et al., 31 Mar 2025).

6. Interpretation, limitations, and research directions

The literature makes clear that “physics-informed” rarely means exact enforcement of first principles. Many systems use soft penalties, regularized latent fields, or prompt-mediated priors rather than hard constraints. UTOPYA speaks of soft physics constraints; SpinFlow softly enforces mass conservation and smoothness; Pi-Transformer regularizes the prior to evolve smoothly and distills it weakly toward dataset-level statistics; Pi-DPM optionally applies only a small physics-informed correction during denoising (Pessoa et al., 18 May 2026, Deng et al., 22 May 2026, Maleki et al., 24 Sep 2025, Sharma et al., 8 Jun 2025). A plausible implication is that practical physics-informed detection is usually a calibration problem between mechanistic bias and empirical flexibility rather than a direct embedding of an exact simulator.

The same literature also identifies clear failure modes. TPIDM notes that quality depends on the accuracy of the physics prior and on finite-difference derivative estimates, and that the weight-schedule hyperparameters require tuning per domain (Soni et al., 15 Aug 2025). The radar digital-twin framework requires an accurate geometric or EM model and a one-shot unlabeled empty-room capture; more complex sim-to-real discrepancies may need additional calibration (Trinh et al., 25 Jan 2026). Hidden-boundary PINNs report hours of training for 2D problems and note that extension to 3D sharply increases collocation budgets and network size (Zhu et al., 31 Mar 2025). FDTRImageEnhancer attributes the large grain-boundary error to insufficient resolution (Odufisan, 8 Aug 2025). These are not peripheral issues: they define the operational envelope of the frameworks.

A further misconception is that adding more regularization or more standard training heuristics necessarily improves generalization. UTOPYA reports a training ablation across 14 design choices in which instance normalisation, Mixup, ensembling, test-time augmentation, and stochastic weight averaging fail to improve or actively degrade generalisation, and it explicitly identifies a fundamental tension between smoothing-based regularisation and anomaly detection (Pessoa et al., 18 May 2026). That result is consistent with the broader observation that anomaly detectors often require preserving small, physically meaningful deviations rather than suppressing them.

Current directions suggest continued convergence between inverse modeling, multimodal reasoning, and deployable detection. Multi-turn VLM prompting demonstrates that physical priors can be delivered in natural language (Gu et al., 16 Mar 2026). PILLM shows that rule generation itself can be embedded in a physics-informed evolutionary loop (Lin et al., 20 Oct 2025). PIGTN and the radar digital twin extend the idea to sensing design and sim2real calibration (Elnour et al., 16 Feb 2026, Trinh et al., 25 Jan 2026). POU-PINN, residual-loss anomaly analysis, and hidden-boundary PINNs show that detection can be recast as identifying where governing physics changes across space or time (Rodriguez et al., 2024, Bai et al., 28 Apr 2026, Zhu et al., 31 Mar 2025). Taken together, these works suggest that the field is moving from simple “physics-regularized classifiers” toward systems that jointly infer latent mechanisms, physical regimes, and operational decisions.