Physics-Informed Transport Guidance (PITG)

• Physics-Informed Transport Guidance is a design principle that couples a physically meaningful guidance object with a constrained transport process to ensure model accuracy.
• It underpins diverse applications including neural rendering, UAV traffic control, optimal transport, and adaptive sample refinement.
• PITG employs structured loss functions and iterative consistency to enforce physical semantics, improving simulation robustness and reliability.

Searching arXiv for the supplied PITG-related papers to ground the article in current literature. Physics-Informed Transport Guidance (PITG) denotes, across several recent research contexts, a class of physics-informed formulations in which a guidance representation is coupled directly to a transport process and constrained by governing physical structure. In the most explicit usage, RenderFormer++ introduces PITG as a mechanism that injects rendering-equation inductive biases into a Transformer-based light transport model for global illumination (Du et al., 29 Jun 2026). Closely related usages appear in macroscopic UAV traffic, where a value field guides conservative density transport under wind and obstacle constraints (Liang et al., 1 Apr 2026), and in optimal-transport, porous-media, stochastic-traffic, inverse-diffusion, and device-optimization settings, where the “guidance” object may be a convex transport map, an adaptive collocation distribution, a flow-informed operator surrogate, a residual-driven inverse model, or a waveform computed from self-consistent electrostatics (Caboussat et al., 17 Jan 2025, Tadiparthi et al., 2021, Kara et al., 27 Aug 2025, Xin, 10 Mar 2026, Zapf et al., 2022, Sokolov et al., 8 Oct 2025). Taken together, these works suggest that PITG is best viewed as a design principle rather than a single formalism: the guidance signal should be physically meaningful, and the induced transport should preserve consistency, conservation, admissibility, or boundary semantics.

1. Conceptual structure and recurring elements

A common PITG pattern is the decomposition of a transport problem into a guidance object and a transport mechanism, followed by a coupling rule that prevents the guidance from becoming merely geometrically plausible or numerically convenient. In RenderFormer++, the guidance object is the explicit separation between emitted radiance and transported/scattered radiance; in the UAV framework, it is the minimum-time value field $\phi(\bm{x})$; in Monge–Ampère optimal transport, it is the convex potential $u$ whose gradient yields the transport map; in OT-refined PINNs, it is the transport plan that reallocates collocation points toward high-error regions; and in stochastic traffic modeling, it is the probability-flow velocity induced by a one-point Fokker–Planck equation (Du et al., 29 Jun 2026, Liang et al., 1 Apr 2026, Caboussat et al., 17 Jan 2025, Tadiparthi et al., 2021, Xin, 10 Mar 2026).

Across these formulations, the transport stage is not left unconstrained. Instead, the literature imposes explicit structure: transport consistency loss in neural rendering, conservative finite-volume fluxes in UAV density transport, convexity and transport-boundary conditions in Monge–Ampère PINNs, mass-preserving constraints in OT-based sampling, and score-form Fokker–Planck residuals in stochastic traffic generative modeling (Du et al., 29 Jun 2026, Liang et al., 1 Apr 2026, Caboussat et al., 17 Jan 2025, Tadiparthi et al., 2021, Xin, 10 Mar 2026). This suggests that PITG is centrally concerned with the distinction between a guidance field that is merely predictive and one that remains operationally valid under the underlying transport law.

A related theme is that “physics-informed” does not mean the same thing in every instance. In some papers it refers to a governing PDE residual, as in PINNs for diffusion, phonon BTE, and Buckley–Leverett transport; in others it refers to hard-encoded target semantics, conservative flux discretization, self-consistent Poisson–Schrödinger evaluation, or explicit rendering-equation structure (Zapf et al., 2022, Li et al., 2021, Gasmi et al., 2021, Liang et al., 1 Apr 2026, Sokolov et al., 8 Oct 2025, Du et al., 29 Jun 2026). The commonality is not a single optimizer or architecture, but the insertion of domain equations, admissibility structure, or boundary semantics into the guidance-to-transport pipeline.

2. Neural rendering: PITG as a learned light transport solver

In RenderFormer++, PITG is the mechanism that injects explicit rendering-equation structure into the Transformer-based light transport model. The starting point is the radiance equilibrium condition

$\mathbf{L}_o = T\mathbf{L}_o + \mathbf{L}_e,$

where $\mathbf{L}_o$ is the outgoing radiance field, $\mathbf{L}_e$ is emitted radiance, and $\mathbf{L}_s = T\mathbf{L}_o$ is the transported/scattered component. The paper introduces PITG because RenderFormer’s unconstrained Transformer formulation,

$$\mathbf{L}_o = T_\theta(\mathbf{F}^{\text{global}}),$$

does not explicitly separate emitted radiance from transported/scattered radiance and does not enforce physical consistency across iterative multi-bounce transport updates (Du et al., 29 Jun 2026).

PITG combines an emission encoder, a neural light transport operator built on self-attention, and a transport consistency loss. Emission features $\mathbf{E}$ are aligned with the scene tokens, with non-emissive entries zero-padded. The paper defines

$$\mathbf{F}_e^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{E})\big), \qquad \mathbf{F}_s^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{0})\big),$$

so that $\mathbf{F}_e^{\text{global}}$ is the feature analogue of emitted radiance and $u$0 is the feature analogue of transported/scattered radiance. Transport is then written in feature space as

$u$1

and iterated with shared parameters:

$u$2

In implementation, the number of transport iterations is set to $u$3 (Du et al., 29 Jun 2026).

The transport consistency loss regularizes the iteration toward a fixed-point-like solution:

$u$4

with full training objective

$u$5

where $u$6 and $u$7. PITG operates in the view-independent stage after Hierarchical Object-Centric Tokenization (HOCT) and before the geometry-guided view-dependent decoder, producing

$u$8

HOCT is the scalability mechanism; PITG is the physics-grounding mechanism (Du et al., 29 Jun 2026).

The ablations isolate its contribution. On the small-scale dataset, adding PITG to RenderFormer improves over the fine-tuned baseline from L1 $u$9, MAPE $\mathbf{L}_o = T\mathbf{L}_o + \mathbf{L}_e,$0, LPIPS $\mathbf{L}_o = T\mathbf{L}_o + \mathbf{L}_e,$1, and SSIM $\mathbf{L}_o = T\mathbf{L}_o + \mathbf{L}_e,$2. On the large-scale dataset, removing PITG degrades performance to L1 $\mathbf{L}_o = T\mathbf{L}_o + \mathbf{L}_e,$3, MAPE $\mathbf{L}_o = T\mathbf{L}_o + \mathbf{L}_e,$4, LPIPS $\mathbf{L}_o = T\mathbf{L}_o + \mathbf{L}_e,$5, and SSIM $\mathbf{L}_o = T\mathbf{L}_o + \mathbf{L}_e,$6, compared to the full model with L1 $\mathbf{L}_o = T\mathbf{L}_o + \mathbf{L}_e,$7, MAPE $\mathbf{L}_o = T\mathbf{L}_o + \mathbf{L}_e,$8, LPIPS $\mathbf{L}_o = T\mathbf{L}_o + \mathbf{L}_e,$9, and SSIM $\mathbf{L}_o$0. Removing the transport consistency loss (“w/o TCL”) yields L1 $\mathbf{L}_o$1, MAPE $\mathbf{L}_o$2, LPIPS $\mathbf{L}_o$3, and SSIM $\mathbf{L}_o$4. The iteration-depth study reports that $\mathbf{L}_o$5 give very similar numbers, with $\mathbf{L}_o$6 chosen as the best efficiency-quality trade-off (Du et al., 29 Jun 2026).

The paper does not claim a formal guarantee of convergence beyond empirical consistency regularization, and its broader limitations affect PITG’s setting: the dataset does not include textured objects, and material properties are shared across triangles within each object. This suggests that the transport refinement is physically structured but still contingent on the scene and material assumptions encoded in the learned features (Du et al., 29 Jun 2026).

3. Guidance fields, transport maps, and conservative semantics

In macroscopic UAV traffic, PITG appears as a value-function-driven transport pipeline. The domain is a bounded 3D airspace $\mathbf{L}_o$7 with obstacles $\mathbf{L}_o$8, free space $\mathbf{L}_o$9, an absorbing target region $\mathbf{L}_e$0, static wind $\mathbf{L}_e$1, and a congestion-dependent speed bound $\mathbf{L}_e$2. The framework outputs a value field $\mathbf{L}_e$3, an induced motion field $\mathbf{L}_e$4, and a steady density field $\mathbf{L}_e$5. The guidance field is derived from the anisotropic HJB/Eikonal problem

$\mathbf{L}_e$6

leading to

$\mathbf{L}_e$7

with induced velocity

$\mathbf{L}_e$8

Transport is then enforced conservatively through

$\mathbf{L}_e$9

optionally with flux $\mathbf{L}_s = T\mathbf{L}_o$0, strict no-flux boundaries on $\mathbf{L}_s = T\mathbf{L}_o$1, and an absorbing target implemented by setting $\mathbf{L}_s = T\mathbf{L}_o$2 inside $\mathbf{L}_s = T\mathbf{L}_o$3 at every pseudo-time step. The coupled system is advanced through an outer Picard iteration with under-relaxation,

$\mathbf{L}_s = T\mathbf{L}_o$4

which the paper describes as a pragmatic stability strategy rather than a formal convergence proof (Liang et al., 1 Apr 2026).

The key reliability claim is that the guidance field is physically coupled to the induced mass transport and that conservation and boundary semantics are imposed directly in the numerical method rather than learned approximately. The reported qualitative outcomes are value slices without spurious local minima away from the target or obstacle geometry, induced-motion corridors that curve with wind and split around obstacles, and steady density bands and bottlenecks shaped by those corridors. By contrast, the end-to-end PINN baseline can fit a plausible $\mathbf{L}_s = T\mathbf{L}_o$5, but its $\mathbf{L}_s = T\mathbf{L}_o$6 tends to collapse to near zero, whereas the hybrid route produces a non-degenerate corridor and matches a traditional FSM-FVM reference qualitatively (Liang et al., 1 Apr 2026).

A different but related PITG formulation appears in the Monge–Ampère optimal transport setting. There, the guidance object is the transport map itself. The paper formulates the quadratic-cost Monge problem

$\mathbf{L}_s = T\mathbf{L}_o$7

with measurable map $\mathbf{L}_s = T\mathbf{L}_o$8 satisfying $\mathbf{L}_s = T\mathbf{L}_o$9. By Brenier’s theorem, the optimal map has the form

$$\mathbf{L}_o = T_\theta(\mathbf{F}^{\text{global}}),$$0

where $$\mathbf{L}_o = T_\theta(\mathbf{F}^{\text{global}}),$$1 is convex and satisfies the generalized Monge–Ampère equation

$$\mathbf{L}_o = T_\theta(\mathbf{F}^{\text{global}}),$$2

together with the transport boundary condition

$$\mathbf{L}_o = T_\theta(\mathbf{F}^{\text{global}}),$$3

The use of Input Convex Neural Networks (ICNNs) is essential because the approximation must remain in the convex function class so that $$\mathbf{L}_o = T_\theta(\mathbf{F}^{\text{global}}),$$4 remains a valid transport guidance map (Caboussat et al., 17 Jan 2025).

The loss combines the PDE residual with a transport-boundary term based on a Hausdorff-distance idea. Discretely, the paper defines $$\mathbf{L}_o = T_\theta(\mathbf{F}^{\text{global}}),$$5 so that one term promotes injectivity of the boundary mapping and the other surjectivity, and solves

$$\mathbf{L}_o = T_\theta(\mathbf{F}^{\text{global}}),$$6

after which the transport map is recovered by differentiation,

$$\mathbf{L}_o = T_\theta(\mathbf{F}^{\text{global}}),$$7

Numerical experiments include disk-to-ellipse transport with componentwise errors around $$\mathbf{L}_o = T_\theta(\mathbf{F}^{\text{global}}),$$8, ellipse-to-rotated-ellipse transport with pointwise errors of order $$\mathbf{L}_o = T_\theta(\mathbf{F}^{\text{global}}),$$9, Gaussian-to-uniform and Gaussian-to-Gaussian redistribution on the unit square, bimodal Gaussian-to-uniform transport, and a 3D Gaussian-to-uniform case on the unit cube (Caboussat et al., 17 Jan 2025).

These two formulations differ substantially in numerics, but both clarify a central PITG idea: the guidance object is operational only if its induced transport respects the correct semantics. In the UAV case those semantics are conservation and no-flux boundaries; in the Monge–Ampère case they are convexity and transport-boundary satisfaction. A common misconception is that PITG reduces to learning a plausible field. The cited work instead treats plausibility as insufficient without conservation, convexity, or boundary correctness (Liang et al., 1 Apr 2026, Caboussat et al., 17 Jan 2025).

4. Transport-guided learning, adaptive sampling, and admissibility

In OT-based refinement of PINNs for the stationary Fokker–Planck-Kolmogorov equation, transport guidance does not act on a physical state variable directly; it acts on the training measure. A nominal PINN is first trained on a relatively small set of collocation points, then evaluated on a testing grid $\mathbf{E}$0, after which the top $\mathbf{E}$1 points with the largest PDE residuals define an error set $\mathbf{E}$2. An OT problem maps a uniformly weighted source set into the error-weighted target set, producing transported samples $\mathbf{E}$3, which are added to the training set:

$\mathbf{E}$4

The transport matrix $\mathbf{E}$5 solves

$\mathbf{E}$6

subject to

$\mathbf{E}$7

The guidance signal is therefore the residual distribution itself, reinterpreted as a transport target for collocation density (Tadiparthi et al., 2021).

The reported gains are substantial. For the Van der Pol–Rayleigh system, a baseline PINN with $\mathbf{E}$8 points yields $\mathbf{E}$9, whereas OT-PINN with $$\mathbf{F}_e^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{E})\big), \qquad \mathbf{F}_s^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{0})\big),$$0 points yields $$\mathbf{F}_e^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{E})\big), \qquad \mathbf{F}_s^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{0})\big),$$1. For the Van der Pol system, a baseline PINN with $$\mathbf{F}_e^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{E})\big), \qquad \mathbf{F}_s^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{0})\big),$$2 points yields $$\mathbf{F}_e^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{E})\big), \qquad \mathbf{F}_s^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{0})\big),$$3, whereas OT-PINN with $$\mathbf{F}_e^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{E})\big), \qquad \mathbf{F}_s^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{0})\big),$$4 points yields $$\mathbf{F}_e^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{E})\big), \qquad \mathbf{F}_s^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{0})\big),$$5. The mechanism is not transport in physical space, but transport-guided sample refinement for physics-informed learning (Tadiparthi et al., 2021).

A different lesson about guidance appears in Buckley–Leverett transport in porous media. There the network predicts saturation $$\mathbf{F}_e^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{E})\big), \qquad \mathbf{F}_s^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{0})\big),$$6 under the hyperbolic conservation law

$$\mathbf{F}_e^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{E})\big), \qquad \mathbf{F}_s^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{0})\big),$$7

with inlet and initial conditions

$$\mathbf{F}_e^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{E})\big), \qquad \mathbf{F}_s^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{0})\big),$$8

The paper argues that, for hyperbolic transport, physics alone is not just the PDE in differential form; one must also enforce entropy or shock admissibility. It therefore replaces the raw fractional flow by a Welge convex-hull construction $$\mathbf{F}_e^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{E})\big), \qquad \mathbf{F}_s^{\text{global}} = \mathrm{EEncode}\big(\mathrm{concat}(\mathbf{F}^{\text{global}}, \mathbf{0})\big),$$9, and this, rather than architecture tweaks alone, is what allows recovery of the correct shock. After $\mathbf{F}_e^{\text{global}}$0 iterations, the reported $\mathbf{F}_e^{\text{global}}$1 error is approximately $\mathbf{F}_e^{\text{global}}$2 for the uniform initial-condition case (Gasmi et al., 2021).

The same paper introduces further guidance mechanisms: spatial weighting of the residual,

$\mathbf{F}_e^{\text{global}}$3

and temporary artificial diffusion,

$\mathbf{F}_e^{\text{global}}$4

with $\mathbf{F}_e^{\text{global}}$5 to $\mathbf{F}_e^{\text{global}}$6 and geometric decay during training. In the non-uniform initial-condition case, the spatially weighted loss performs better than diffusion regularization and better matches the Godunov reference (Gasmi et al., 2021).

These studies show that PITG can guide either the computational mesh or the admissible solution class. A plausible implication is that transport guidance in physics-informed learning need not always produce a control law or state-space flow; it can also operate by reshaping where the solver concentrates resolution or by constraining which weak solutions are permitted.

5. Inverse problems, operator surrogates, and stochastic distributional transport

Inverse diffusion in the human brain provides a residual-centered PITG formulation. The transport model is the diffusion equation

$\mathbf{F}_e^{\text{global}}$7

where $\mathbf{F}_e^{\text{global}}$8 is the apparent diffusion coefficient to be inferred from MRI observations at $\mathbf{F}_e^{\text{global}}$9 hours. The PINN minimizes

$u$00

with data loss $u$01 and residual loss

$u$02

No explicit boundary condition is imposed in the loss; the initial condition is supplied through the data at $u$03. The paper shows that with noisy synthetic data, standard training overfits, visually reproduces noise, and drives the learned coefficient to $u$04. Recovery improves by increasing $u$05, switching from $u$06 to $u$07 residuals, using bounded sigmoid parameterization

$u$08

and applying residual-based adaptive refinement (RAR) or residual-based adaptive exchange (RAE). With $u$09, exponentially decaying learning rates, and adaptive refinement, the PINN estimate for the main case is about $u$10, close to the FEM result $u$11 (Zapf et al., 2022).

In phonon transport, PITG takes the form of a data-free surrogate for the mode-resolved phonon Boltzmann transport equation. Under steady-state, no internal heat source, single-mode relaxation time approximation, and small temperature difference, the paper solves

$u$12

using a PINN trained by

$u$13

A distinguishing feature is that the characteristic length scale $u$14 is included as an input, enabling parametric learning across geometry and size. In a 1D thin film, the model matches the Implicit Kinetic Scheme with temperature-profile discrepancy below $u$15 and thermal-conductivity error below $u$16; in a 2D in-plane case, it agrees with Fuchs–Sondheimer theory with heat-flux prediction error below $u$17 and effective-conductivity error below $u$18; and in a 3D cuboid, the $u$19 mm case matches a Fourier-law benchmark with error below $u$20 (Li et al., 2021). Here the guidance is the BTE itself, and the transport object is the multiscale phonon energy distribution.

Operator-learning in porous media yields another PITG variant. The hybrid FEM + physics-informed DeepONet framework solves the steady Darcy system with FEM and injects the resulting velocity field into a DeepONet residual for the time-dependent convection–diffusion equation

$u$21

with residual

$u$22

The training loss is

$u$23

with $u$24, $u$25, and $u$26. A key contribution is adaptive trunk collocation sampling, combining source-centered polar sampling with random points, boundary points, and initial points. In the sampling ablation, hybrid sampling substantially outperforms random-only sampling; for example, in the triple-source fixed-width case at $u$27, hybrid gives $u$28 versus random-only $u$29. Reported inference speedups at $u$30 range from $u$31 to $u$32 relative to FEM (Kara et al., 27 Aug 2025).

A more indirect but still relevant surrogate-modeling contribution is RA-PINN for irregular interfaces and multi-peak transport-like fields. The model uses six residual-attention blocks and physics-informed training on a coupled field $u$33, with total loss $u$34. On three benchmark cases—oblique asymmetric interface, bipolar high-gradient charge layer, and multi-peak Gaussian charge migration field—RA-PINN achieves the best overall averaged RMSE: $u$35, $u$36, and $u$37, respectively, outperforming pure PINN and LSTM-PINN, while incurring the highest computational cost and parameter count (Zhou et al., 24 Mar 2026). The paper does not formulate explicit transport guidance, but it is explicitly presented as PITG-adjacent surrogate modeling for transport-like multiphysics fields.

Finally, stochastic traffic dynamics extends PITG from point estimates to distributions. Starting from an Itô-type stochastic LWR model,

$u$38

the paper derives a one-point marginal equation

$u$39

with conditional drift

$u$40

This is rewritten as a deterministic probability-flow ODE

$u$41

and learned with a score network $u$42 plus advection-closure module $u$43. Training uses denoising score matching together with a Fokker–Planck residual loss. The outputs are local density distributions $u$44, from which one can compute point estimates, conditional standard deviation, credible intervals, and congestion risk (Xin, 10 Mar 2026). This is a distributional PITG formulation in which the guidance is a physics-derived transport law in density space.

6. Device-level transport optimization, limitations, and scope

In semiconductor spin-qubit transport, PITG takes the form of a device-aware optimization workflow for conveyor-mode electron shuttling. The method uses self-consistent Poisson and Schrödinger solvers in QTCAD to design gate-voltage waveforms that maintain a constant ground-state energy and nearly constant velocity. The core equations are

$u$45

and

$u$46

The optimization first flattens the conduction band through iterative numerical solves $u$47, with the rule that the target $u$48 should be kept at least twice as high as the parasitic well depth, then determines voltages giving a target ground-state energy

$u$49

and finally reparameterizes the waveform to enforce nearly constant transport velocity by constructing $u$50 from the simulated dot coordinate $u$51 (Sokolov et al., 8 Oct 2025).

The reported behavior is strongly technology dependent. In FD-SOI with J-gates, the ground-state-to-first-excited-state spacing stays above about $u$52 meV during shuttling and velocity is constant within numerical error; without J-gates, the energy gap collapses to zero around $u$53 a.u., corresponding to a tunnelling event visible as a step in $u$54. In SiMOS, the target ground-state energy is set to $u$55 meV, the trajectory $u$56 is nearly linear, but the splitting $u$57 decreases locally where the dot expands. In Si/SiGe, the shuttling sequence is optimized for constant ground-state energy $u$58 meV, transport remains uniform, and the level spacing gradually decreases from about $u$59 meV to $u$60 meV. Runtime and memory also vary sharply: FD-SOI without J-gates takes $u$61 s and $u$62 MB, SiMOS takes $u$63 s and $u$64 MB, and Si/SiGe takes $u$65 s and $u$66 MB (Sokolov et al., 8 Oct 2025).

These results highlight a general limitation of PITG: adding physically meaningful guidance usually does not remove modeling assumptions, and often introduces new computational or closure burdens. RenderFormer++ does not provide a formal convergence proof for iterative transport refinement and is trained under non-textured, object-wise material assumptions (Du et al., 29 Jun 2026). The UAV framework is limited to steady-state behavior under static wind and uses obstacle treatment in the PINN that is engineered and configuration-dependent (Liang et al., 1 Apr 2026). The Monge–Ampère method requires convexity-preserving ICNN training and meaningful initialization (Caboussat et al., 17 Jan 2025). Brain diffusion recovery depends critically on achieving a small residual after training rather than merely good data fit (Zapf et al., 2022). Porous-media PI-DeepONet is currently most effective in moderately convective regimes, and very high Péclet numbers remain difficult (Kara et al., 27 Aug 2025). The stochastic traffic formulation assumes smooth regimes without shocks and requires closure for the conditional drift $u$67 (Xin, 10 Mar 2026).

A second misconception is that PITG guarantees physical truth once a PDE term appears in the loss. The cited literature argues for a narrower conclusion. In hyperbolic porous-media transport, weak-form PDE enforcement alone is insufficient without entropy or shock admissibility (Gasmi et al., 2021). In MRI-based inverse diffusion, visually acceptable fits can still correspond to misleading parameters when the post-training residual is high (Zapf et al., 2022). In UAV traffic, a plausible value function without conservative density transport can yield numerically meaningless density fields (Liang et al., 1 Apr 2026). In stochastic traffic, the physics constraint itself becomes distributional, and exactness depends on a closure that is not directly known (Xin, 10 Mar 2026). PITG therefore does not eliminate approximation; it relocates approximation into structured components that are intended to preserve the right transport semantics.

Taken broadly, the literature indicates that PITG is a unifying lens for methods that couple a guidance representation to a transport process under explicit physical structure. Depending on the domain, the crucial preserved property may be emission–scattering separation, fixed-point-like consistency, conservative flux balance, convexity of a transport potential, entropy admissibility, residual-driven refinement, physically faithful operator injection, or device-level energy-gap stability (Du et al., 29 Jun 2026, Liang et al., 1 Apr 2026, Caboussat et al., 17 Jan 2025, Gasmi et al., 2021, Kara et al., 27 Aug 2025, Sokolov et al., 8 Oct 2025). The term thus spans rendering, mobility, logistics-style redistribution, reservoir simulation, inverse imaging, quantum-device control, and stochastic traffic inference, while retaining a common methodological core: guidance is not merely predictive structure, but structure constrained to remain compatible with transport physics.