Physics-encoded convolutional neural operators for parametric PDEs: A convergence-guaranteed framework via...

Operator learning for parametric partial differential equations (PDEs) has emerged as a transformative paradigm, with physics-informed neural networks (PINNs), Deep Operator Networks (DeepONet), and Fourier Neural Operators (FNO) demonstrating remarkable capabilities. However, these methods face a critical tradeoff: data-driven operators (DeepONet, FNO) require 10³-10⁴ expensive solution snapshots for training, while physics-informed approaches (PINNs) lack rigorous convergence guarantees due to non-convex soft constraints computed via automatic differentiation. This creates a fundamental gap: no existing method provides provable O(h^k+1-m) approximation-capacity bounds while simultaneously requiring zero training data and maintaining fast multi-query inference. To address these limitations, we propose Convolutional Neural Operators with Physics-Encoded Kernels (CNO-PEK), a novel operator learning framework that pre-computes differential operators as spatially-adaptive convolutional kernels through the Physics-Informed Kernel Field (PIKF). Unlike data-driven operators requiring thousands of solution snapshots [Formula: see text] or PINNs computing residuals via automatic differentiation at random collocation points, our method encodes the governing physics directly into the network architecture by analytically deriving kernels from polynomial consistency conditions. This architectural design provides provable O(h^k+1-m) approximation-capacity bounds matching classical finite element theory (Theorem 1, subject to polynomial representability; practical convergence of the trained network additionally requires optimization to a global minimum, as stated in Proposition 1)), while simultaneously leveraging neural networks' superior approximation capabilities without requiring any solution training data. Through comprehensive validation on 1D-3D benchmark problems and systematic comparison with PINNs, DeepONet, and classical methods, we demonstrate: (1) O(h^k+1-m) approximation-capacity bounds (Theorem 1) confirmed empirically, (2) 60.9% systematic error reduction compared to linear finite elements (mean error 1.72% vs. 4.40%) without requiring training data, (3) physics-driven zero-shot parameter generalization (trained once at μ=2.5 with no solution data at any μ ∈ [0, 5]) across 1000 unseen parameter values, with 100% of cases achieving  < 3% error, (4) computational speedup of 6.9-13.6 ×  for multi-query scenarios beyond a break-even point of approximately 843 queries. We position CNO-PEK as a theoretically grounded operator learning framework that bridges classical numerical analysis and modern deep learning, particularly suited for parametric studies, uncertainty quantification, and real-time applications where approximation-capacity guarantees and data efficiency are critical.