A Universal Fix for AI's Physics Problem

In the high-stakes world of scientific computing, neural operators have emerged as a revolutionary tool for solving complex physical systems described by partial differential equations (PDEs). These AI models promise to accelerate simulations that traditionally require massive computational resources, from climate forecasting to materials science. However, researchers from the University of Science and Technology of China and Tsinghua University have uncovered a fundamental flaw: existing neural operators consistently violate the most basic physical laws. Their groundbreaking study reveals that when these models predict systems governed by conservation laws—where quantities like mass or energy must remain constant—they produce physically impossible that accumulate errors over time, severely limiting their real-world applicability.

The research team identified two critical problems plaguing current approaches. First, neural operators lack built-in mechanisms to enforce conservation properties, leading to what they term "physical non-conservation." As shown in their experiments, models like CNO exhibit continuous temporal decay of conserved quantities, with predictions deviating dramatically from ground truth. Second, they discovered that different PDE problems require different optimal neural architectures—UNO performs best for Allen-Cahn problems while FNO excels at adiabatic systems—creating a fundamental tension between specialization and generalization. To address these limitations, the researchers developed the Exterior-Embedded Conservation Framework (ECF), a universal plug-and-play module that can be integrated with any existing neural operator.

ECF operates through an elegant frequency-domain correction mechanism built around two core components. The conserved quantity encoder extracts conservation information from input data using Fourier transforms, specifically capturing the zero-frequency component that represents spatial averages of conserved quantities. The conserved quantity decoder then corrects the neural operator's predictions by replacing the predicted zero-frequency term with the ground truth value from the encoder while preserving all higher-frequency components. This approach is mathematically guaranteed to reduce prediction errors while strictly enforcing conservation laws, as proven in their theoretical analysis. The framework offers two training paradigms: Integrated Mode (ECF I) for end-to-end optimization and Staged Mode (ECF S) for post-training correction of pre-existing models.

The experimental validation demonstrates ECF's remarkable effectiveness across six distinct conservation-law-constrained PDE scenarios, including adiabatic systems, shallow water equations, and Allen-Cahn problems with both double-well and Flory-Huggins potentials. When integrated with leading neural operators like FNO, UNO, CNO, UNet, and Transolver, ECF achieved performance improvements of up to 37.7%, with the +ECF I approach reducing UNO's error on Allen-Cahn problems from 0.140 to 0.0623. Most impressively, the framework reduced conservation errors from levels exceeding 100% in some cases to machine precision (below 1E-6), effectively eliminating the temporal accumulation of conservation violations that plague baseline models. The computational overhead proved negligible, with training time increases of just 1.02% for UNO models.

This research represents a significant advancement in physics-informed machine learning, bridging the gap between data-driven approaches and fundamental physical principles. By providing a theoretically grounded, universally applicable solution to conservation law enforcement, ECF enables neural operators to produce physically realistic predictions while maintaining their computational efficiency advantages. The framework's plug-and-play design means it can immediately enhance existing models across scientific computing applications, from fluid dynamics simulations to quantum system modeling, without requiring architectural changes or sacrificing performance on non-conservative systems where traditional neural operators already excel.