HyCOP and the Return of Modular Structure in Neural Operators

Scientific machine learning has reached an inflection point. While monolithic neural operators like Fourier Neural Operators and DeepONets have demonstrated impressive interpolation capabilities within training distributions, their behavior under distribution shift remains opaque and often unreliable. When faced with novel boundary conditions, unseen parameter regimes, or discontinuous initial data, these end-to-end learned surrogates frequently fail catastrophically, offering little insight into the nature of their errors or how to rectify them.

The paper "HyCOP: Hybrid Composition Operators for Interpretable Learning of PDEs" by Zhao et al. presents a compelling alternative to this black-box paradigm. Rather than learning a direct map from input functions to solution operators, HyCOP treats PDE solving as a program synthesis problem. It learns a compact policy that composes reusable physical primitives (advection solvers, diffusion operators, learned closures) into query-specific computational graphs. This approach represents not merely an architectural variation, but a conceptual shift from function approximation to algorithm learning.

The Mechanics of Compositional Surrogates

At the heart of HyCOP lies a surprisingly compact learned policy πθ, typically parameterized by only 50 to 100 weights. This policy operates over a dictionary D of computational primitives, which may include classical numerical schemes (spectral methods, finite volume solvers), pretrained neural components, or hybrid combinations thereof. Given a query x = (u0, µ, T) consisting of initial conditions, parameters, and target time, the policy outputs a program: a sequence of (primitive, duration) pairs that compose to approximate the solution at time T.

This architecture circumvents the autoregressive rollout that plagues many neural surrogates. Rather than stepping forward iteratively through time, HyCOP produces a direct mapping to arbitrary query horizons by orchestrating the computational budget across primitives. The framework supports non-uniform time allocation; for instance, the policy might assign extended durations to diffusion primitives in viscous regimes while rapidly cycling through advection operators in hyperbolic regions.

The authors provide a theoretical error decomposition that separates composition error (how well the policy orchestrates primitives) from module error (how accurately individual primitives represent their respective physical processes). This separation serves a dual purpose. Theoretically, it characterizes the expressivity of the framework, showing that composition policies can approximate complex operators through simpler components. Practically, it functions as a diagnostic tool, enabling scientists to identify whether poor performance stems from inadequate primitive modeling or suboptimal orchestration.

Scientific Workflows and Modular Transfer

HyCOP distinguishes between two operational regimes that mirror realistic scientific workflows. In Regime A, where governing equations are fully specified, the framework serves as a fast surrogate constructor. The scientist encodes known physics as dictionary primitives, and HyCOP learns only the composition policy. The results here are striking: on shallow water equation benchmarks, HyCOP demonstrates order-of-magnitude improvements in out-of-distribution generalization compared to monolithic baselines. Specifically, policies trained on smooth periodic boundary conditions transfer zero-shot to dam-break shocks and wall boundaries, outperforming traditional neural operators by factors exceeding 10×. Swapping the boundary primitive for a specialized wall-boundary module without retraining the policy yields additional 4× improvements.

Regime B addresses the more nuanced scenario of incomplete physics or heterogeneous data sources. Here, HyCOP implements a compose-diagnose-enrich loop. When composing known primitives produces structured residuals (errors concentrating in specific spatiotemporal regions), these patterns reveal missing physics. The authors demonstrate this on an advection-diffusion-reaction benchmark where initially only advection and diffusion are known. The composition policy reveals systematic failures in reaction-dominated regions, prompting the inclusion of a UNO (U-Net Neural Operator) closure trained on the residual. Relearning only the policy then recovers accuracy while providing interpretable activation maps of where the learned closure engages.

This modular transfer capability represents a significant departure from conventional transfer learning in neural operators. Rather than fine-tuning massive parameter counts, scientists can update the primitive dictionary through targeted additions or swaps, requiring only policy retraining with minimal computational overhead.

Your Take: Program Synthesis and the Scientific Method

Beyond the technical specifications, HyCOP embodies a philosophical stance regarding the role of structure in scientific machine learning. The work challenges the prevailing narrative that numerical PDE solvers are prohibitively expensive and must be replaced wholesale by neural surrogates. As the authors note, this conflates the cost of full coupled solvers with their constituent processes. Individual primitives (a spectral diffusion step, an upwind advection scheme) often execute faster than neural forward passes at comparable resolution. The true bottleneck lies in determining how to compose these elements, a cognitive and algorithmic challenge that HyCOP addresses through learned policies.

This framework situates itself at the intersection of neuro-symbol