ACR-PINN: Conflict-Resolved Neural PDE Solver

• ACR-PINN is a neural PDE solver framework that resolves architectural and optimization conflicts via domain-adaptive representations and gradient conflict mitigation.
• It employs finite geometric encoding and layer-wise dynamic attention to inject consistent spatial features and improve gradient flow.
• Empirical benchmarks demonstrate that ACR-PINN achieves significantly lower errors and faster convergence than standard PINN methods.

Architecture-Conflict-Resolved Physics-Informed Neural Networks (ACR-PINN) are a class of neural solvers for partial differential equations (PDEs) that address two main failure modes of classical PINNs: limited representational capacity in complex settings and pathological optimization under multiple competing physical constraints. ACR-PINN frameworks achieve robust and accurate solution of PDEs by unifying advances in network architecture and optimization strategy while preserving the classical PINN loss structure. The central concept underlying ACR-PINN is to explicitly resolve architectural or training conflicts in the PINN pipeline—whether these stem from geometric domain mismatch, physical boundary conditions, or gradient interference—by integrating domain-adaptive representations and gradient conflict mitigation (Niu et al., 19 Jan 2026, Li et al., 2024).

1. Defining Architectural Conflicts in PINNs

Classical Physics-Informed Neural Networks formulate the PDE solution $u(x, t)$ as a neural network $u_\theta(x, t)$ and minimize a composite loss,

$$\mathcal{L}(\theta) = \lambda_{\rm PDE} \,\mathcal{L}_{\rm PDE} + \lambda_{\rm IC} \,\mathcal{L}_{\rm IC} + \lambda_{\rm BC} \,\mathcal{L}_{\rm BC}$$

where each term quantifies residuals or mismatches for the PDE, initial, and boundary conditions. However, this generic approach fails to address two principal sources of conflict:

• Architectural representational conflict: Conventional MLPs struggle to encode solutions over finite, topologically complex geometries found in solid mechanics, as standard inputs and activations are Euclidean and global, while solid domains are typically finite with intricate topology (Li et al., 2024).
• Optimization conflict: PINN training enforces multiple physics constraints by summing their gradients. These can be adversarial (negatively aligned) in the parameter space, which can stall or destabilize optimization (Niu et al., 19 Jan 2026).

The ACR-PINN paradigm addresses each type: geometric/architectural domain conflict is resolved at the input-processing stage; physical constraint interference is resolved via gradient manipulation.

2. Layer-wise Dynamic Attention and Geometric Encoding

Two classes of ACR-PINN methodologies are prominent in recent literature:

Finite Geometric Encoding

For solid mechanics, the “Finite-PINN” model implements a geometric encoding which transforms the neural input space from purely Euclidean to a hybrid Euclidean–topological representation:

• LBO eigenbasis construction: Solve the Laplace–Beltrami operator (LBO) eigenproblem on a finite domain $\Omega$,

$$\begin{cases} -\Delta\phi_k(x) = \lambda_k\,\phi_k(x) & x \in \Omega \ -\nabla \phi_k \cdot n = 0 & x \in \partial\Omega \end{cases}$$

where $\phi_k(x)$ are wavelength-ordered basis functions computed via FEM (weak form).

• Hybrid input construction: At each collocation point, augment $x \in \mathbb{R}^d$ with the first $n_u$ eigenfunctions $\{\phi_1(x), ..., \phi_{n_u}(x)\}$. The input to the network thus becomes $$[x_1, \ldots, x_d, \phi_1(x), ..., \phi_{n_u}(x)] \in \mathbb{R}^{d+n_u}$$ (Li et al., 2024).
• Solution map ansatz: The stress and displacement outputs are represented as

$$\sigma_{ij}(x) = p_{ik}(x)\, \phi_{k,j}(x), \qquad u_i(x) = f_i(x, \phi_1(x), ..., \phi_{n_u}(x))$$

where $p_{ik}$ and $f_i$ are neural networks.

This construction allows the PINN to “see” the correct geodesic distances and domain boundaries, providing domain-awareness not present in standard PINNs (Li et al., 2024).

Layer-wise Dynamic Attention (LDA)

For general PDEs (e.g., Burgers, Helmholtz, Klein-Gordon), the ACR-PINN “Layer-wise Dynamic Attention” network injects original input coordinates at every hidden layer using an attention/gating mechanism:

• Input re-encoding: At each hidden layer $\ell$, two “views” of the original input $x$ are computed as $e_\ell^{(i)} = \phi(E_\ell^{(i)} x + c_\ell^{(i)})$.
• Feature gating: These views are combined with the MLP features via a trainable gating network, yielding feature-wise attention weights. The modulated input is added residually to the hidden layer output.
• Gradient propagation: This architecture distributes gradient signals and alleviates spectral bias by maintaining a persistent flow of pointwise coordinate information throughout the network depth (Niu et al., 19 Jan 2026).

3. Conflict-Resolved Gradient Optimization

The second pillar of ACR-PINN is explicit resolution of optimization conflicts using conflict-aware gradient updates:

• Task decomposition: Each physical loss term (PDE, IC, BC) is treated as a separate “task” generating a gradient $\mathbf{g}_i$.
• Conflict detection: For each task pair $(i, j)$, a conflict is declared if $\mathbf{g}_i^\top \mathbf{g}_j < 0$.
• Orthogonal projection (PCGrad): Upon conflict, project out the conflicting component,

$$\mathbf{g}_i \leftarrow \mathbf{g}_i - \frac{\mathbf{g}_i^\top\mathbf{g}_j}{\|\mathbf{g}_j\|_2^2}\,\mathbf{g}_j$$

• Final update: After random pairwise projections across all tasks, sum the resulting $\widetilde{\mathbf{g}_i}$ for the optimizer step.

No additional hyperparameters or adjustments to the original PINN loss structure are required. When no conflicts are present, the update reduces to ordinary gradient descent (Niu et al., 19 Jan 2026).

4. Integrated ACR-PINN Training Algorithm

The unified ACR-PINN training loop combines architectural (forward pass) and optimization (backward pass) conflict resolution. A typical training iteration, as implemented for benchmark PDEs, proceeds:

1. Sample interior, boundary, and initial collocation batches.
2. Forward pass through an LDA network or domain-encoded PINN for predictions.
3. Compute all composite physics and data residuals as in the standard loss.
4. Compute per-task gradients via automatic differentiation.
5. Apply PCGrad projections to all task gradients to enforce conflict resolution.
6. Aggregate and apply the conflict-resolved update with Adam optimizer.

Hyperparameters and network configurations mirror those used in classical PINNs, e.g., for Burgers equation: 2-in, 4 layers × 20 neurons, tanh activations, $40\,000$ iterations; loss weights $\lambda_{i} = 1$ (Niu et al., 19 Jan 2026, Li et al., 2024).

5. Empirical Performance and Benchmarks

Extensive benchmark results demonstrate that ACR-PINN consistently achieves lower relative $L_2$ and $L_\infty$ errors and faster convergence than standard PINN, LDA-PINN, and GC-PINN ablations (Niu et al., 19 Jan 2026).

Model	Burgers $\varepsilon_{L_2}$ ($\times 10^{-3}$)	Helmholtz $(1,4)$ $\varepsilon_{L_2}$ ($\times 10^{-2}$)	Klein–Gordon $\varepsilon_{L_2}$ ($\times 10^{-2}$)
Std-PINN	9.96  ±  5.59	17.7  ±  5.48	6.23  ±  1.35
LDA-PINN	2.60  ±  1.89	3.06  ±  0.72	2.17  ±  0.41
GC-PINN	4.68  ±  2.38	0.863  ±  0.08	0.871  ±  0.18
ACR-PINN	0.915  ±  0.12	0.816  ±  0.20	0.333  ±  0.12

Similar improvements are observed for the Helmholtz and Klein–Gordon equations, as well as for 2D Navier–Stokes (lid-driven cavity) and solid mechanics tasks with nontrivial topologies. In solid mechanics, Finite-PINN demonstrates order-of-magnitude improvements in convergence and robustness to geometric complexity (Li et al., 2024).

6. Conflict Resolution Mechanisms and Theoretical Basis

• Architectural conflict resolution (via geometric encoding, LDA) ensures that the network’s representation space aligns with the finite, topologically intricate geometry of the physical domain or propagates coordinate information more effectively through the network layers, alleviating spectral bias and stabilizing optimization.
• Optimization conflict resolution (PCGrad) projects out mutually adversarial gradient components corresponding to competing physics or data tasks, thereby improving the effective conditioning of the loss landscape and avoiding pathological training dynamics. These interact synergistically: enhanced representations “spread” task gradients, facilitating stable projection, while gradient conflict mitigation allows the architecture to realize its improved conditioning (Niu et al., 19 Jan 2026, Li et al., 2024).

7. Illustrative Applications and Research Implications

Notable empirical demonstrations include:

• 1D rod dynamics with Neumann boundaries: Finite-PINN captures free-surface reflections absent in standard PINNs.
• 2D and 3D solid mechanics: Topology-adaptive PINNs accurately recover displacement/stress fields and react appropriately to domain notches or springs, with significant gains in data efficiency and fidelity.
• Classical PDE benchmarks (Burgers, Helmholtz, Klein–Gordon): ACR-PINN achieves lower errors and converges an order of magnitude faster than baselines.

The ACR-PINN framework exemplifies architecture–optimization co-design in physics-informed learning. Directly resolving representational and optimization conflicts expands the applicability, robustness, and accuracy of PINN solvers across diverse PDE regimes (Niu et al., 19 Jan 2026, Li et al., 2024).