Neural networks trained to satisfy a differential equation, known as physics-informed neural networks (PINNs), are scored on how well they obey the equation at thousands of sample locations called collocation points. These networks break down on boundary layers. A boundary layer is a thin strip near a wall where the solution changes extremely fast, and it is the same structure whose breakdown under rising downstream pressure causes flow separation and stall. Inside the layer the training error is huge, and those few points pull gradient descent away from fitting everything else. Two published fixes take opposite routes. One builds the layer's known analytic shape into the network. The other deliberately ignores the hardest points until the rest is learned. Our proposal starts from the observation that this is a heterogeneous classroom. Different regions of the domain are "students" that learn at very different speeds. Educational measurement already knows how to schedule practice for such a class: it neither drills only the strugglers nor only the easy cases, but targets whoever is ready to improve next.
We propose to put a small dynamic learner model next to the PINN. Each region's probability of being "mastered" is tracked over training as a shared trend plus a region-specific deviation, borrowing the decomposition from dynamic item-response models. Training points are then sampled where the predicted gain in mastery per step is largest. The model also answers a question that global loss cannot: has the network actually learned the layer, or has it collapsed to the smooth outer solution while reporting a low average error? That second failure is the solver's version of what tutoring research calls false mastery. The core claim is testable. Scheduling by predicted learning velocity should match or beat loss-thresholded curricula on singularly perturbed benchmarks without needing an analytic corrector, and it should detect layer collapse that aggregate loss misses.
Core Claim and Positioning
The closest prior work is Less Emphasis on Difficult Layer Regions. It trains PINNs on convection-diffusion-reaction problems down to perturbation parameter $\epsilon = 10^{-9}$ by down-weighting points whose loss exceeds an adaptive threshold $\beta(t)$, so the smooth part is learned first. That paper supplies our benchmarks and our strongest learned-schedule competitor. Its schedule reacts only to each point's current loss. It has no notion of how fast a region is improving, and it has no per-region convergence test. Singular layer PINN takes the other route: it adds an asymptotically derived corrector function to the architecture and reaches relative $L^2$ error $2.0 \times 10^{-3}$ at $\epsilon = 10^{-6}$, where a standard PINN gets $9.21 \times 10^{-1}$. It constrains our proposal. Any schedule-only method must be compared against a method that knows the answer's structure, and it must justify itself where no corrector can be derived, which is the setting of separated flows.
The educational side supplies the method and the motivation. D-BIRD models each student's ability as a cohort trend plus an individual deviation, each following its own random walk. We reuse this split: the network's overall progress is the cohort trend, and each region's lag or lead is the deviation. The Impact of Individualizing Student Models shows on ASSISTments what a population-average learner model costs. Over 20% of students did twice the necessary practice, and 17% were declared mastered at a true mastery of 60% or less. These are exact analogues of wasted compute on already-fit regions and of a PINN declared converged while missing its layer. The adaptive testing survey supplies the selection principle: pick items whose information is highest at the learner's current state, not the hardest items.
The research file's own summary frames these two sides as opposed. The PINN curriculum avoids hard points, while tutoring insists on mastering hard prerequisites. Our claim is that both are special cases of one rule: target the units with the highest expected near-term gain. Hard layer points have low expected gain early in training and high expected gain later, once the outer solution is stable.
Research Program
We partition the domain into regions $r$. These can be grid cells, or bands at increasing distance from the wall, so that layer and outer regions are separated. At each training checkpoint $t$, a point counts as "correct" if its residual is below a tolerance $\tau$. We model the fraction of correct points in region $r$ with
$$p_{r,t} = \sigma(\mu_t + \beta_{r,t}),\qquad \mu_t = \mu_{t-1} + \eta_t,\quad \beta_{r,t} = \beta_{r,t-1} + \delta_r + \xi_{r,t}.$$
Here $\sigma$ is the logistic function. $\mu_t$ is the shared progress of the whole network, and $\beta_{r,t}$ is region $r$'s deviation from it. $\delta_r$ is region $r$'s own learning velocity, and $\eta_t, \xi_{r,t}$ are Gaussian innovations with separate variances, as in D-BIRD. We use a Kalman filter on a logit-linearized version for speed, rather than D-BIRD's Gibbs sampler, because the model runs inside a training loop. Its overhead will be reported.
The sampling weight for region $r$ is its predicted one-step mastery gain:
$$w_{r,t} \propto \sigma(\mu_t + \beta_{r,t} + \hat\delta_r) - \sigma(\mu_t + \beta_{r,t}).$$
This gain is small for regions already near mastery, where $p \approx 1$. It is also small for regions that are not moving yet, either because $\hat\delta_r \approx 0$ or because they sit deep in the flat low-$p$ tail. It peaks at regions that are partly learned and actively improving, in the same way item information $p(1-p)$ peaks at matched difficulty. A per-region mastery test stops training only when every region's posterior $p_{r,t}$ exceeds a threshold. This test is also a diagnostic: it flags layer collapse whenever outer regions are mastered and wall-adjacent regions have stalled.
Baselines
The naive reference point is a standard PINN with uniform collocation and the same network, optimizer, point budget and iteration count. The cited reported figures are NRMSE $4.47 \times 10^{-1}$ on the 1D $\epsilon = 10^{-9}$ problem (Less Emphasis on Difficult Layer Regions, Table 3) and relative $L^2$ $9.21 \times 10^{-1}$ on the 2D square at $\epsilon = 10^{-6}$ (Singular layer PINN).
The strongest learned-schedule competitor is the loss-threshold curriculum from Less Emphasis on Difficult Layer Regions, which reports NRMSE $1.47 \times 10^{-4}$ on the 1D $\epsilon = 10^{-9}$ benchmark. We will rerun it from its description on all six of its benchmarks, with its threshold schedule tuned under the same number of hyperparameter trials we give our method. Beating it means lower NRMSE on a majority of the six problems at matched iterations, or equal NRMSE with at least 30% fewer iterations to reach it. We also require no loss on the 1D case beyond its reported value.
The structured competitor is SL-PINN from Singular layer PINN, which reports relative $L^2$ $2.0 \times 10^{-3}$ on the square domain at $\epsilon = 10^{-6}$. It uses information our method does not have. We treat it as an upper reference on its square, circle and ellipse problems, and success there means coming within a factor of three of it without a corrector.
Two ablations isolate our key ingredient. The first is a "population-average" version that keeps the learner model but drops the region deviations $\beta_{r,t}$, so every region shares one schedule. It mirrors the population-average knowledge-tracing condition. The second is a "static band" sampler that draws points whose current loss lies in a middle quantile. It captures "neither easiest nor hardest" with no learning-velocity estimate. A hardest-first sampler that up-weights high-loss points tests the claimed inversion directly.
Why not just…?
Why not just use the loss-threshold curriculum, which already works? On the 1D problem it reaches $1.47 \times 10^{-4}$, so the room for improvement there is small, and we say so. Our case rests on two things it cannot provide. The first is the stopping and collapse diagnostic, measured as the false-mastery rate defined below. The second is robustness of the threshold schedule across the 2D and 3D problems, where a fixed rule for $\beta(t)$ may need per-problem tuning. If our method ties it everywhere and the diagnostic adds nothing, the proposal reduces to a monitoring tool, and we will report it that way.
Why not just build the boundary layer into the network, as SL-PINN does? Where an asymptotic corrector exists, SL-PINN will probably win, and we do not expect to beat $2.0 \times 10^{-3}$. Its correctors come from closed-form expansions, which do not exist for separated flows with recirculation. A schedule needs no such knowledge. Our separation test case, described below, is where this matters.
Why not just sample from a middle loss band, with no learner model at all? This is the most serious threat, because it captures most of the intuition cheaply. The static-band ablation is run under identical budgets. If it matches our method on error and on collapse detection, the learning-velocity term $\hat\delta_r$ is not earning its place, and we would drop the dynamic model and publish the simpler sampler. We expect the difference to show at very small $\epsilon$. There, layer points sit at high loss for a long time and then drop quickly, and only a velocity estimate anticipates that drop.
Proposed Experiments
The main suite is the six singularly perturbed convection-diffusion-reaction benchmarks of Less Emphasis on Difficult Layer Regions, covering 1D, 2D and 3D problems down to $\epsilon = 10^{-9}$. We add the square, circular and elliptical problems of Singular layer PINN at $\epsilon$ from $10^{-6}$ to $10^{-8}$. We report NRMSE and relative $L^2$ error to match each paper's metric. We also report error restricted to the layer region alone, iterations to reach a target error, and wall-clock overhead of the filter, all with five seeds.
The decisive check is a false-mastery rate. It is the fraction of runs in which aggregate training loss falls below its stopping criterion while the layer-region error exceeds ten times the outer-region error. We compare how often aggregate-loss stopping versus our per-region mastery test is fooled. A second decisive check sweeps $\epsilon$. The predicted signature is that hardest-first sampling degrades fastest as $\epsilon$ shrinks, the static band degrades next, and our scheduler degrades least.
Finally, to reach flow separation, we construct a steady laminar two-dimensional boundary layer under an imposed adverse pressure gradient strong enough to produce a small recirculation bubble, with a fine finite-difference reference solution. No reported result was found on Lacuna for PINNs on this kind of case. It therefore serves as an exploratory test of whether the per-region diagnostic flags a network that misses the reversed-flow region near the wall, rather than as a leaderboard comparison.
References
- Less Emphasis on Difficult Layer Regions: Curriculum Learning for Singularly Perturbed Convection-Diffusion-Reaction Problems
- Singular layer physics-informed neural network method for convection-dominated boundary layer problems in two dimensions
- Dynamic Bayesian Item Response Model with Decomposition (D-BIRD): Modeling Cohort and Individual Learning Over Time
- The Impact on Individualizing Student Models on Necessary Practice Opportunities
- Survey of Computerized Adaptive Testing: A Machine Learning Perspective