Physics-Informed Masking Curriculum

• Physics-Informed Masking Curriculum is a training strategy that selectively masks and reweights physics-based components, enabling smoother optimization and improved convergence.
• It employs various mechanisms like collocation support expansion, boundary-to-interior loss activation, and gradient gating to target specific physical constraints.
• The approach integrates fixed schedules and adaptive curricula to gradually expose challenging physics, thereby reducing error rates and boosting PDE and PINN performance.

“Physics-Informed Masking Curriculum” is best understood as an umbrella label for training strategies that selectively mask, gate, or reweight physically meaningful parts of a learning problem, and that do so either progressively or adaptively over optimization. In the strictest usage, the masked object is part of a physics-constrained objective—collocation support, PDE residuals, boundary terms, or physics gradients—and the curriculum determines when and where those constraints are active (Münzer et al., 2022, Marcandelli et al., 2 Feb 2026). In broader usage, related ideas appear in measurement-space masking for inverse problems, hidden-feature masking inside PINNs, and physics-based control or simulation pretraining, although several of those papers explicitly stop short of claiming either a full curriculum or a strict physics-informed formulation (Golooba et al., 25 Mar 2026, Yang et al., 19 May 2026).

1. Conceptual scope and terminological boundaries

The term is not standardized across the literature. Some papers implement an explicit curriculum over the support of the physics loss. Others contribute only a masking mechanism, only a curriculum mechanism, or a domain-informed analogue. The distinction matters because “physics-informed” is used narrowly in PINNs and operator learning to denote direct use of governing equations, residuals, or physically meaningful operators, but more loosely elsewhere to denote alignment with structured system behavior.

This narrow–broad distinction is explicit in several adjacent works. The operational-technology cyber-defence study on IPMSRL uses action masking and curriculum learning in a maritime OT simulator, but states that the method is not “physics-informed” in the classical sense of embedding physical laws, differential equations, conservation constraints, or mechanistic process models; its masking is more accurately described as environment-informed or domain-informed (Wilson et al., 2024). By contrast, SpecTM deterministically masks diagnostically meaningful hyperspectral bands based on established phytoplankton optics, which is physics-informed in a domain-prior sense, but it does not introduce a progressive masking schedule and therefore is not a curriculum in the strict sense (Imtiaz et al., 23 Mar 2026). The MRI multi-mask training paper similarly applies masking to the acquired measurement set inside physics-guided data-consistency units, but explicitly presents random multi-masking rather than easy-to-hard task sequencing (Yaman et al., 2020).

A useful synthesis is that a strict physics-informed masking curriculum requires two elements at once: first, masking or selective activation of quantities that are themselves tied to governing physics or physically meaningful operators; second, staged or adaptive control over that masking during training. The clearest instantiations in the current literature are expanding collocation supports in PINNs (Münzer et al., 2022), boundary-to-interior loss activation in physics-informed operator learning (Marcandelli et al., 2 Feb 2026), gradient-level gating of physics losses in epidemiological PINNs (Golooba et al., 25 Mar 2026), and residual-cluster-based soft weighting in CGMPINN (Yang et al., 19 May 2026).

2. Main forms of masking and what they act upon

Across the cited papers, masking operates on different mathematical objects rather than on a single canonical target.

Masked or reweighted object	Representative formulation	Curriculum character
Collocation support	Expanding cuboid or cylinder support for PINN collocation points (Münzer et al., 2022)	Explicit fixed schedule
Boundary vs interior physics	$$\lambda_{bd}\mathcal{L}_{bd}+\lambda_{res}\mathcal{L}_{res}$$ with early $\lambda_{res}=0$ (Marcandelli et al., 2 Feb 2026)	Explicit multi-stage continuation
Physics gradient	$$\lambda(t)=\frac{\|\mathbf{g}_{data}\|}{\|\mathbf{g}_{phy}\|+\epsilon}\sigma(\kappa S_{cos})$$ (Golooba et al., 25 Mar 2026)	Adaptive emergent curriculum
Residual clusters / collocation points	GMM-derived sample weights $w_i$ on PDE residual samples (Yang et al., 19 May 2026)	Smooth easy-to-hard curriculum
Difficult layer regions	Residual-threshold weights $w(t,\mathbf{x}_i)$ that cap hard-region loss (Wang et al., 2022)	Dynamic soft masking
Measurements or channels	MRI subsets $\Theta_j\subset\Omega$ or diagnostic spectral bands $\mathcal{D}$ (Yaman et al., 2020, Imtiaz et al., 23 Mar 2026)	Usually masking without explicit curriculum

The first class masks where physics is enforced. In the MHD PINN collocation paper, the physics loss is evaluated only on collocation points sampled from an active subset of space-time, initially a narrow cuboid in time or a cylinder around observed trajectories, and later on the full domain (Münzer et al., 2022). The second class masks which constraints are emphasized. In PhIS-FNO, the early stages emphasize boundary conditions while residual enforcement is absent or weak, then progressively unmask the interior PDE residual (Marcandelli et al., 2 Feb 2026).

A third class masks the strength of physics in gradient space rather than in input or coordinate space. CGGS computes the cosine similarity between data and physics gradients and attenuates the physics term when the two objectives conflict, producing a continuous gate on physical supervision (Golooba et al., 25 Mar 2026). A fourth class masks hard residual regions softly. The singularly perturbed PINN paper and CGMPINN both retain all collocation points but downweight points or clusters judged difficult, thereby acting as continuous masks over residual contributions rather than binary exclusions (Wang et al., 2022, Yang et al., 19 May 2026).

Finally, several neighboring works mask measurements, channels, nodes, or hidden features. These are highly relevant building blocks, but many of them do not themselves define a curriculum. Multi-mask MRI training masks acquired $k$-space samples within the data-consistency operator (Yaman et al., 2020); SpecTM masks diagnostically selected hyperspectral bands (Imtiaz et al., 23 Mar 2026); Ambient Physics additionally masks already-observed measurements during training from partial observations (Majid et al., 14 Feb 2026); Mask-PINNs applies a learnable pointwise mask to hidden activations (Jiang et al., 9 May 2025); MeshMask removes graph nodes and their incident edges during CFD pretraining (Garnier et al., 15 Jan 2025).

3. Curriculum mechanisms: fixed schedules, progressive unmasking, and adaptive gates

The most explicit curriculum designs are fixed schedules over the support of the physics loss. In the MHD reconstruction paper, the number of collocation points per epoch is kept constant while the sampling region expands. Training lasts at most 5000 epochs, the curriculum occupies the first 30% of training, the cuboid support expands five times every 300 epochs, and the cylinder support expands fifteen times every 100 epochs before switching to full-domain sampling (Münzer et al., 2022). This makes the curriculum a progressive domain reveal rather than a density increase.

A second fixed-schedule design is staged constraint activation. In PhIS-FNO, the stage loss is written as $$\mathcal{L}_{(\lambda_{bd},\lambda_{res})}=\lambda_{bd}\mathcal{L}_{bd}+\lambda_{res}\mathcal{L}_{res}$$, and early stages can literally set $\lambda_{res}=0$. For Poisson, the three stages are $\lambda_{res}=0$0, $\lambda_{res}=0$1, and $\lambda_{res}=0$2; for Navier–Stokes, the five-stage schedule begins with two boundary-only phases before progressively increasing residual weight (Marcandelli et al., 2 Feb 2026). The optimizer is reinitialized at each stage, which the paper interprets as a continuation mechanism that avoids stagnation in Adam’s accumulated moments.

Adaptive curricula replace hand-designed stage transitions with training-state-dependent gates. CGGS defines

$\lambda_{res}=0$3

where $\lambda_{res}=0$4 is the cosine similarity between data and physics gradients (Golooba et al., 25 Mar 2026). When gradients are anti-aligned, $\lambda_{res}=0$5 is small and the physics term is nearly shut off; when they align, the gate opens. The paper explicitly interprets the resulting “data-first, then physics-strong” pattern as an autonomous curriculum learning effect.

CGMPINN implements a different adaptive curriculum. A GMM is fit to the empirical PDE residual distribution, each component receives a difficulty score and a precision score, and sample weights are built by combining responsibilities with a shared curriculum parameter

$\lambda_{res}=0$6

Early in training, component weights favor easy and low-variance residual clusters; later they shift toward harder clusters and remove variance-based suppression (Yang et al., 19 May 2026). This is not a hand-crafted spatial schedule, but an easy-to-hard progression in residual space.

4. Canonical PDE and PINN exemplars

The clearest soft-masking curriculum for hard PDE regions appears in the singular perturbation paper. It replaces the standard residual average by

$\lambda_{res}=0$7

with

$\lambda_{res}=0$8

The threshold $\lambda_{res}=0$9 is updated from points whose residual-squared gradients satisfy $$\lambda(t)=\frac{\|\mathbf{g}_{data}\|}{\|\mathbf{g}_{phy}\|+\epsilon}\sigma(\kappa S_{cos})$$0, so likely layer regions are attenuated rather than removed (Wang et al., 2022). Empirically, for the 1D benchmark with $$\lambda(t)=\frac{\|\mathbf{g}_{data}\|}{\|\mathbf{g}_{phy}\|+\epsilon}\sigma(\kappa S_{cos})$$1, the reported NRMSE is $$\lambda(t)=\frac{\|\mathbf{g}_{data}\|}{\|\mathbf{g}_{phy}\|+\epsilon}\sigma(\kappa S_{cos})$$2 for the curriculum method versus $$\lambda(t)=\frac{\|\mathbf{g}_{data}\|}{\|\mathbf{g}_{phy}\|+\epsilon}\sigma(\kappa S_{cos})$$3 for vanilla PINN; for the 2D benchmark with $$\lambda(t)=\frac{\|\mathbf{g}_{data}\|}{\|\mathbf{g}_{phy}\|+\epsilon}\sigma(\kappa S_{cos})$$4, $$\lambda(t)=\frac{\|\mathbf{g}_{data}\|}{\|\mathbf{g}_{phy}\|+\epsilon}\sigma(\kappa S_{cos})$$5 versus $$\lambda(t)=\frac{\|\mathbf{g}_{data}\|}{\|\mathbf{g}_{phy}\|+\epsilon}\sigma(\kappa S_{cos})$$6; and for the 3D benchmark with $$\lambda(t)=\frac{\|\mathbf{g}_{data}\|}{\|\mathbf{g}_{phy}\|+\epsilon}\sigma(\kappa S_{cos})$$7, $$\lambda(t)=\frac{\|\mathbf{g}_{data}\|}{\|\mathbf{g}_{phy}\|+\epsilon}\sigma(\kappa S_{cos})$$8 versus $$\lambda(t)=\frac{\|\mathbf{g}_{data}\|}{\|\mathbf{g}_{phy}\|+\epsilon}\sigma(\kappa S_{cos})$$9 (Wang et al., 2022).

CGMPINN pushes this logic from hand-crafted layer attenuation to probabilistic residual clustering. Its full method consistently outperforms both standard PINN and ablations without curriculum. In the 1D Poisson case, CGMPINN reports $w_i$0, compared with $w_i$1 for standard PINN and $w_i$2 for the GMM-only ablation without curriculum; across six PDEs, the paper reports relative $w_i$3 reductions of up to 97.8% over standard PINN at comparable cost (Yang et al., 19 May 2026). A notable theoretical feature is that the weighted PDE loss remains uniformly equivalent to the standard PDE loss, while the generalization bound makes the weighting-induced bias explicit through a term $w_i$4 (Yang et al., 19 May 2026).

PhIS-FNO demonstrates that the same masking-curriculum logic carries over from PINNs to operator learning. Its multi-stage training with optimizer reset decisively outperforms single-stage training. For Poisson, the reported interior mean residual is $w_i$5 for multi-stage training, compared with $w_i$6 for single-stage and $w_i$7 for multi-stage without reset (Marcandelli et al., 2 Feb 2026). For Navier–Stokes with $w_i$8, PhIS-FNO reports $w_i$9 under multi-stage training, versus $w(t,\mathbf{x}_i)$0 for single-stage and $w(t,\mathbf{x}_i)$1 without reset; supervised PhIS-FNO is $w(t,\mathbf{x}_i)$2, showing that the staged unsupervised method can approach supervised performance (Marcandelli et al., 2 Feb 2026).

The MHD collocation-curriculum paper and the epidemiological CGGS paper illustrate two other poles of the design space. The former reports that curriculum collocation required about 35% fewer epochs to converge on average, with improvements of up to $w(t,\mathbf{x}_i)$3 and $w(t,\mathbf{x}_i)$4 on average, and that the cylinder curriculum performed slightly better overall than the cuboid method (Münzer et al., 2022). The latter shows that pure magnitude balancing is insufficient when physics and data gradients conflict: LRA undershoots the infection peak by about 15%, whereas CGGS avoids the local minimum, reaches a final error an order of magnitude lower, and exhibits a “dip-and-rise” adaptive physics weight, with $w(t,\mathbf{x}_i)$5 near $w(t,\mathbf{x}_i)$6 in steps 0–250 and above $w(t,\mathbf{x}_i)$7 in steps 250–600 (Golooba et al., 25 Mar 2026).

5. Inverse problems, scientific representation learning, and adjacent simulation settings

Several papers contribute the masking part directly while leaving the curriculum component implicit or absent. In accelerated MRI, multi-mask supervised PG-DL repeatedly trains the same unrolled network on random subsets $w(t,\mathbf{x}_i)$8 of the acquired measurements, with $w(t,\mathbf{x}_i)$9 and $\Theta_j\subset\Omega$0 (Yaman et al., 2020). The method is explicitly described as a physics-consistent stochastic masking mechanism in measurement space, not as a curriculum. On knee MRI, the best variant is $\Theta_j\subset\Omega$1: for uniform sampling it reports median SSIM $\Theta_j\subset\Omega$2 and PSNR $\Theta_j\subset\Omega$3, compared with $\Theta_j\subset\Omega$4 and $\Theta_j\subset\Omega$5 for conventional supervised PG-DL; for random sampling it reports $\Theta_j\subset\Omega$6 and $\Theta_j\subset\Omega$7, compared with $\Theta_j\subset\Omega$8 and $\Theta_j\subset\Omega$9 (Yaman et al., 2020).

Mask-PINNs shifts masking inside the network rather than over samples or domain regions. It defines a smooth learnable mask

$\mathcal{D}$0

and applies it pointwise to hidden activations as

$\mathcal{D}$1

The paper argues that this preserves the deterministic pointwise map required by PINNs while regulating feature spread (Jiang et al., 9 May 2025). It is explicitly not a curriculum method, but it is one of the clearest instances of genuinely physics-compatible masking. Empirically, on the convection equation with SoftPlus, the relative $\mathcal{D}$2 error drops from $\mathcal{D}$3 for Vanilla PINN to $\mathcal{D}$4 for Mask-PINN, roughly two orders of magnitude; on the Helmholtz equation, Mask-PINN remains accurate across Tanh, GELU, SiLU, and SoftPlus while several baselines largely fail for some activations (Jiang et al., 9 May 2025).

In simulation pretraining, MeshMask masks graph nodes rather than PDE losses. It randomly removes up to 40% of input mesh nodes during pretraining, along with incident edges, and reports improvements of up to 60% in long-term prediction accuracy over previous best models across seven simulation datasets (Garnier et al., 15 Jan 2025). Its main value for the present topic is architectural and curricular by implication: the paper finds that moderate-to-high masking, around 25%–40% node masking, is best, and that harder masking may require $\mathcal{D}$5-hop edge augmentation. This suggests, but does not itself implement, an easy-to-hard masking curriculum over mesh information (Garnier et al., 15 Jan 2025).

SpecTM and Ambient Physics provide two further scientific exemplars. SpecTM deterministically masks 28 of 122 PACE OCI bands—about 23%—covering 615–640 nm, 660–680 nm, and 695–720 nm, and reports $\mathcal{D}$6 for current-week and $\mathcal{D}$7 for 8-day-ahead microcystin prediction, with targeted masking improving downstream performance by $\mathcal{D}$8 over random masking and yielding 2.2x superior label efficiency under extreme scarcity (Imtiaz et al., 23 Mar 2026). Ambient Physics, by contrast, trains PDE reconstruction models from partial observations by additionally masking a subset of already-observed entries and supervising on the full originally observed set; it reports a 62.51% reduction in average overall error while using 125$\mathcal{D}$9 fewer function evaluations than prior diffusion-based methods, and identifies a “one-point transition,” in which masking a single already-observed point is enough to enable learning from partial observations (Majid et al., 14 Feb 2026). Both papers are powerful masking designs; neither is an explicit curriculum, but both supply mechanisms that can be staged.

6. Limitations, misconceptions, and likely directions

A central misconception is that every masking technique in scientific machine learning is already a physics-informed masking curriculum. The literature is more segmented. Some papers are clearly physics-informed but not curricular, such as Mask-PINNs and multi-mask MRI training (Jiang et al., 9 May 2025, Yaman et al., 2020). Others are clearly curricular but use masking only in an implicit sense, such as boundary-to-interior loss activation in PhIS-FNO (Marcandelli et al., 2 Feb 2026). Still others are domain-informed rather than strictly physics-informed, as the OT cyber-defence paper explicitly states (Wilson et al., 2024).

Another limitation is that many masks or schedules are hand-designed. SpecTM fixes its diagnostic band set; it is domain-grounded but static (Imtiaz et al., 23 Mar 2026). PhIS-FNO uses manually selected stage weights and fixed 100-epoch stage lengths (Marcandelli et al., 2 Feb 2026). Ambient Physics shows that masking geometry matters and that mismatched masks may need stronger additional masking, especially in structured observation settings such as super-resolution lattices (Majid et al., 14 Feb 2026). In control or deployment-oriented settings, handcrafted masks may also create dependencies at inference time: the OT paper notes that “the masking conditions used during training need to be present when querying the trained policy” (Wilson et al., 2024).

Theoretical coverage is also uneven. CGGS proves the $k$0 convergence-rate guarantee only for the instantaneous gate with $k$1, while experiments use EMA with $k$2 (Golooba et al., 25 Mar 2026). CGMPINN provides stronger theory for time-varying weighted losses, but its own analysis makes clear that the curriculum-weighted objective may retain a non-vanishing bias relative to the original PDE risk through $k$3 (Yang et al., 19 May 2026). This means that some methods are better interpreted as principled reweightings than as temporary warm-start masks that disappear entirely by the end of training.

Taken together, the literature suggests a converging design principle rather than a settled recipe. Physics-informed masking is most effective when the masked object is genuinely tied to physical structure—collocation support, PDE residuals, measurement operators, diagnostically meaningful bands, or physically valid action spaces—and curriculum is most effective when it exposes those structures in an optimization-friendly order. The clearest empirical lesson is not merely “mask more,” but “control when and where the learner sees hard physics.” In current work that control is implemented by expanding active domains, progressively unmasking residual terms, suppressing conflicting gradients, or downweighting difficult residual clusters (Münzer et al., 2022, Marcandelli et al., 2 Feb 2026, Golooba et al., 25 Mar 2026, Yang et al., 19 May 2026). This suggests that future systems described by the term will likely combine explicit physics-derived masking with adaptive, state-dependent progression rather than relying on static masks or fixed uniform enforcement.