PENCO: A Physics-Energy-Numerics-Consistent Operator for 3D Phase Field Modeling

Abstract

Accurate and efficient solutions of spatiotemporal partial differential equations (PDEs), such as phase-field models, are fundamental for understanding interfacial dynamics and microstructural evolution in materials science and fluid mechanics. Neural operators (NOs) have recently emerged as powerful data-driven alternatives to traditional solvers; however, existing architectures often accumulate temporal errors, struggle to generalize over long temporal horizons, and require large training datasets. To overcome these limitations, we propose PENCO (Physics-Energy-Numerics-Consistent Operator), a hybrid operator-learning framework that integrates physical laws with data-driven neural operator methods, using either the Fourier Neural Operator (FNO-4D) or the Multi-Head Neural Operator (MHNO) architecture as the backbone. The formulation introduces an enhanced L^2 Gauss-Lobatto collocation residual around the temporal midpoint that robustly enforces the governing dynamics and significantly improves accuracy, a Fourier-space numerical consistency term that captures the balanced behavior of semi-implicit discretizations, and an energy-dissipation constraint that ensures thermodynamic consistency. Additional low-frequency spectral anchoring and teacher-consistency mechanisms further stabilize learning and suppress long-term error growth. This hybrid design enables PENCO to preserve governing physics while mitigating long-term error growth. Through extensive three-dimensional phase-field benchmarks on periodic cubic domains, covering phase ordering, crystallization, epitaxial growth, and complex pattern formation, PENCO demonstrates superior accuracy, stability, and data efficiency compared to state-of-the-art neural operators, including FNO-4D and MHNO, while maintaining physically consistent evolution. The associated dataset and implementation are available at github.com/MBamdad/PENCO.