Multi-Phase Dual-PINN Framework

• Multi-Phase Dual-PINN is a framework that coordinates multiple neural networks to enforce domain-specific loss terms for solving multiphysics and multiphase PDE problems.
• It employs advanced loss decomposition and adaptive Inexact Restoration techniques to achieve improved stability and up to 90% error reduction in benchmark scenarios.
• The framework is applied across diverse fields such as material design, shock dynamics, and interface tracking, enabling high-fidelity simulations in complex domains.

A Multi-Phase Dual-PINN (Physics-Informed Neural Network) framework refers to a class of physics-informed deep learning architectures and optimization strategies tailored for the solution of multi-domain, multiphysics, or multiphase PDE problems. These schemes coordinate multiple neural networks—each associated with distinct physical domains, phases, or objectives—and enforce coupling and interface conditions via structured loss formulations and training curricula. The approach generalizes both “dual-network” PINNs (interior–boundary specialization, primal–dual variational pairs) and “multi-phase” architectures (per-region/phase subnetworks, generalized IR methods), providing flexibility and enhanced stability for scenarios involving sharp gradients, moving interfaces, discontinuous material parameters, or complex multiphase interactions.

1. Theoretical Foundations and Loss Decomposition

Multi-Phase Dual-PINN frameworks exploit the decomposability of the PINN objective into a sum of non-negative terms, each corresponding to physically distinct residuals, data mismatches, boundary/interface penalties, or region-wise PDEs. Let $L_1(x), \ldots, L_P(x)$ denote $P$ such terms for a parameter vector $x$; e.g., for a standard PINN one may distinguish between bulk physics residuals (e.g., $L_1$) and boundary/data mismatch (e.g., $L_2$) as in

$L_{\mathrm{total}}(x) = L_1(x) + L_2(x)$

for PDE residual and data loss terms (López, 2024). For complex multiphase or multiphysics settings, $L_i$ can correspond to phase-specific residuals, interface continuity, or constraint enforcement. Multi-Phase Dual-PINN generalizations coordinate multiple subnetworks (or variable groups) to minimize such loss decompositions, embedding hierarchical or distributed physical priors.

2. Multi-Phase Dual-PINN Architectures

State-of-the-art frameworks partition the solution field and parameter space in the following ways:

• Domain–Boundary Specialization: $u(x) = u_\mathrm{D}(x) + u_\mathrm{B}(x)$, with $u_\mathrm{D}$ and $u_\mathrm{B}$ neural subnetworks each encouraged to specialize on the interior and near-boundary regions, respectively. Role specialization is softly imposed via distance-weighted priors $P$0 and $P$1, with $P$2 denoting the distance to the boundary. Adaptive annealing is used to modulate the strength of this specialization over training (Abbas et al., 28 Nov 2025).
• Per-Phase or Per-Region Subnetworks: In moving interface and multiphase PDEs, distinct neural networks $P$3 are assigned to each phase or spatial region $P$4; interface conditions (e.g., continuity of field or traction) are enforced through inter-network loss terms (Rodopoulos et al., 29 Jun 2025, Zhang et al., 2024, Kathane et al., 2024).
• Primal-Dual Neural Networks: For variational PDEs and homogenization, “primal–dual” PINNs are constructed by training both a “primal” network (e.g., field variable $P$5) and a “dual” or stream function (e.g., $P$6) in parallel, with loss terms designed to yield guaranteed upper and lower bounds on quantities of interest (e.g., homogenized constants), and to flag convergence failures (Gaynutdinova et al., 9 Sep 2025).

3. Optimization and Training Strategies

A distinctive element of Multi-Phase Dual-PINN frameworks is structured optimization organized into sequential or nested “phases,” drawing on adaptations of the Inexact Restoration (IR) method (López, 2024). The core principles are:

Two-Phase IR Algorithm (for $P$7 minimization):

1. Phase 1 (Restoration): Minimize $P$8 (feasibility term, typically boundary/data loss) such that $P$9 for $x$0.
2. Penalty Adaptation: If sufficient global progress is made in the composite penalty $x$1, update $x$2 using explicit formulas to reflect loss balance.
3. Phase 2 (Descent): Minimize $x$3 (physics residual) starting from $x$4, ensuring monotonic decrease in both $x$5 and the composite penalty.
4. Iteration: Repeat, ensuring monotonicity and adaptivity (López, 2024).

Multi-Phase Extension: For $x$6 losses $x$7, phases proceed recursively, restoring $x$8 and updating a $x$9-simplex penalty parameter $L_1$0. This enables staged control over feasibility and optimality across multiple objectives (e.g., physics, data, interface).

For role specialization, the phase structure is mirrored in “curriculum learning”: early epochs establish network roles via uniform sampling and strong regularization; later, focused collocation and annealing relieve specialization, promoting accurate global solutions (Abbas et al., 28 Nov 2025).

4. Interface, Coupling, and Constraint Handling

Multi-Phase Dual-PINN methodologies address intra- and inter-network coupling using dedicated loss and constraint terms:

• Augmented Lagrangian Boundary Enforcement: Dirichlet and other boundary data are imposed via augmented Lagrangian terms, dynamically updating Lagrange multipliers $L_1$1 and penalty parameters $L_1$2, suppressing the need for manual penalty tuning (Abbas et al., 28 Nov 2025).
• Interface Residuals: For multiphase materials or moving boundary problems, interface conditions (e.g., continuity of displacement and traction, Stefan condition, solute partition) are enforced by adding interface-weighted residuals to the loss:

$L_1$3

for appropriate test functions $L_1$4 (Rodopoulos et al., 29 Jun 2025, Kathane et al., 2024).

• Jump Conditions in Hyperbolic Conservation Laws: At moving shocks/interfaces, the Rankine–Hugoniot condition and continuity are embedded directly in the loss, with subnetwork predictions and interface locations (possibly neural or ODE-based) jointly trained (Zhang et al., 2024).
• Weak/Variational Enforcement: In variational settings, interface and phase constraints can be enforced weakly by integrating against local or global test functions, enabling accommodation of discontinuous coefficients and multi-material systems (Gaynutdinova et al., 9 Sep 2025, Rodopoulos et al., 29 Jun 2025).

5. Application Domains and Quantitative Performance

Multi-Phase Dual-PINN frameworks have been systematically applied to:

• Benchmark PDEs: Laplace, Poisson, Fokker-Planck, Burgers’, and Heat equations, showing $L_1$5 error reduction and $L_1$6 decrease in MAE over standard PINNs, and improved boundary satisfaction by $L_1$7 (Abbas et al., 28 Nov 2025, López, 2024).
• Multiphase Architected Materials: Elastic plates with inclusions, handled by variational dual-PINN with tree-based integration, achieving $L_1$8 relative accuracy and computational cost reductions via adaptive quadrature (Rodopoulos et al., 29 Jun 2025).
• Two-Shock and Rarefaction Flows: Dual-shock Buckley-Leverett equations, employing region-wise subnetworks and Welge’s flux construction for sharp discontinuity tracking, yielding forward errors $L_1$91–5% and reliable inverse parameter identification under noise (Zhang et al., 2024).
• Solidification and Moving Boundary Problems: Separate networks per variable–phase (temperature, concentration, interface), with progressive domain focusing and causal weighting, resolve discontinuous compositions and energy flux at moving boundaries with $L_2$0 temperature error and interface location MSE $L_2$1 (Kathane et al., 2024).
• Microstructure Evolution: Multi-phase field problems, with hierarchical, master–worker PINNs or “pyramidal” training, capturing dynamic triple junctions, curvature-driven evolution, and interface constraint satisfaction with MSE as low as $L_2$2 (Elfetni et al., 2024).

6. Extensions and Multi-Objective Generalizations

The IR-inspired multi-phase curriculum can be extended to any number $L_2$3 of losses/objectives, with phase-wise restoration and nested penalty updates ensuring staged enforcement of feasibility and optimality. In complex multiphysics and coupled field problems, per-variable, per-phase networks and comprehensive interface-coupling terms are readily incorporated. The same principles enable primal–dual PINN approaches, providing guaranteed upper/lower bounds and convergence diagnostics in homogenization and multiscale modeling (Gaynutdinova et al., 9 Sep 2025).

A high-level workflow summarizing the Multi-Phase Dual-PINN paradigm across cited works is given here:

Subfield / Application	Network Decomposition	Interface / Constraint Enforcement
Boundary–interior specialization	Domain and boundary nets	Soft role priors + ALM for boundary (Abbas et al., 28 Nov 2025)
Multi-phase media & inclusions	Per-phase subnetworks	Interface residuals (weak or strong) (Rodopoulos et al., 29 Jun 2025)
Shock/rarefaction conservation laws	Region-wise subnetworks	Jump/Rankine–Hugoniot residuals (Zhang et al., 2024)
Primal–dual variational bounds	Primal/dual networks	Variational constraints, primal–dual gap (Gaynutdinova et al., 9 Sep 2025)
Alloy solidification, moving B.C.	Per-variable/-phase nets	Stefan, partitioning, progressive focusing (Kathane et al., 2024)
Multi-phase field evolution	Phase/batch subnetworks	Master/worker, pyramidal training, partition constraint (Elfetni et al., 2024)

7. Challenges, Diagnostics, and Future Directions

Empirical and theoretical studies highlight:

• Stability and Oscillations: Structured phase-wise descent suppresses loss oscillations seen in naively weighted sum-of-loss PINN training, providing smoother convergence and faster attainment of target accuracy (López, 2024).
• Diagnostics via Primal–Dual Gaps: In homogenization, the gap $L_2$4 serves as a robust indicator for PINN convergence and validity, outperforming loss-based heuristics (Gaynutdinova et al., 9 Sep 2025).
• Adaptive and Focused Sampling: Interfacial and region-aware point selection accelerates training and improves accuracy, especially for sharp features (Abbas et al., 28 Nov 2025, Elfetni et al., 2024).
• Penalty and Role Annealing: Dynamic adjustment of coupling strengths ($L_2$5, $L_2$6, $L_2$7) mitigates the need for manual tuning and fosters specialization, yet requires careful curriculum design (Abbas et al., 28 Nov 2025).

Multi-Phase Dual-PINN frameworks are convergent under suitable IR assumptions, but theoretical guarantees may lapse in heuristic modifications. Ongoing directions include expansion to 3D problems, large-scale parallelization, automatic penalty learning, and integration of semi-analytical or experimental constraints (Rodopoulos et al., 29 Jun 2025, Elfetni et al., 2024). Their principled loss decomposition and modular neural structure underpin their extensibility to multiphysics, multiphase, and multiscale systems.