Experiment Note
The practical starting point was a Kovasznay-flow PINN notebook. Kovasznay flow was a useful first system because the exact steady Navier-Stokes solution made it possible to inspect velocity, pressure, vorticity, and residual errors over the whole domain. The central observation was that a single global loss could look satisfactory even while one physical quantity remained wrong in a small part of the domain. A pressure defect, velocity-component error, or vorticity error could be averaged away by regions that were already easy.
The initial research question was therefore not simply, "How can a PINN be made to train faster?" It was closer to: "Can the training algorithm remember both what is failing and where it is failing, and spend its next unit of compute on that particular failure?"
Technical Essay
This question led naturally to a two-axis state representation. Most adaptive sampling methods preserve location but collapse variable identity into a single residual score. Most loss-balancing methods preserve objective identity but use one global weight over the domain. The emerging VARA idea was to preserve the joint coordinate:
physical channel x spatial region.
In a multi-field flow problem, the channel might represent u, v, pressure, pressure gradient, vorticity, continuity, momentum, aggregate PDE residual, or boundary violation. The domain could be divided into patches. A diagnostic tensor would then score each channel-patch pair, allowing the training system to act on a specific weak pair rather than applying a uniform global change.
The architectural intuition was a feedback-control loop around an ordinary PINN: observe the current field, diagnose a localized failure, propose a small training intervention, measure its effect, and retain or undo it. The neural network remained the field approximator; VARA would be an online controller of the training allocation rather than a replacement network architecture.
Reading Note
The literature map later recorded in the report and paper suggests four bodies of work shaped the initial problem framing:
- Foundational PINNs and automatic-differentiation residual training established the base formulation and the appeal of learning from physics with sparse data.
- Work on PINN pathologies emphasized gradient imbalance, stiffness, spectral bias, and neural-tangent-kernel behavior, explaining why a low aggregate loss does not imply uniformly accurate fields.
- Residual adaptive refinement, importance sampling, self-adaptive attention, gradient-enhanced PINNs, causal curricula, and global loss balancing showed several ways to redirect training, but each retained only part of the state.
- XPINNs, cPINNs, VPINNs, hp-VPINNs, and FBPINNs demonstrated the value of spatial locality, while also clarifying that VARA should not be presented as a new domain-decomposition architecture. VARA keeps one trial network and uses regions for diagnosis and resource allocation.
Progress Update
The project hypothesis was formed: localized PINN failures should be treated as structured control signals rather than compressed into a scalar objective. The first milestone was to turn the notebook idea into a modular codebase capable of training, diagnosing, intervening, rolling back, logging, and evaluating.