HyCOP: Hybrid Composition Operators for Interpretable Learning of PDEs
| Authors | Jinpai Zhao et al. |
| Year | 2026 |
| Field | AI / ML |
| arXiv | 2605.00820 |
| Download | |
| Categories | cs.CE, cs.LG |
Abstract
We introduce HyCOP, a modular framework that learns parametric PDE solution operators by composing simple modules (advection, diffusion, learned closures, boundary handling) in a query-conditioned way. Rather than learning a monolithic map, HyCOP learns a policy over short programs - which module to apply and for how long - conditioned on regime features and state statistics. Modules may be numerical sub-solvers or learned components, enabling hybrid surrogates evaluated at arbitrary query times without autoregressive rollout. Across diverse PDE benchmarks, HyCOP produces interpretable programs, delivers order-of-magnitude OOD improvements over monolithic neural operators, and supports modular transfer through dictionary updates (e.g., boundary swaps, residual enrichment). Our theory characterizes expressivity and gives an error decomposition that separates composition error from module error and doubles as a process-level diagnostic.
Engineering Breakdown
Plain English
HyCOP is a framework that learns to solve parametric PDEs by composing simple modular components (advection, diffusion, learned closures, boundary handling) rather than training a single monolithic neural operator. Instead of directly learning a fixed solution map, the system learns a policy that decides which modules to apply and for how long, conditioned on features of the current PDE regime and state statistics. The key result is order-of-magnitude out-of-distribution improvements over existing neural operators, with the added benefit that the learned policies are interpretable as short programs, enabling evaluation at arbitrary query times without expensive autoregressive rollouts. The modular design also supports transfer learning through dictionary updates—you can swap boundary conditions or add residual enrichment without retraining from scratch.
Core Technical Contribution
The core novelty is framing PDE solving as a policy-learning problem over short programs rather than learning a single monolithic map. HyCOP composes heterogeneous modules (both numerical solvers and learned components) in a query-conditioned, adaptive way—the system learns when to apply which operator and for how long based on the current state and PDE regime. This hybrid approach differs fundamentally from prior neural operator work (DeepONet, FNO, etc.) which typically learn a single fixed composition. The interpretability emerges naturally: the learned policies are human-readable decision sequences that reveal which physical processes dominate in different regimes, and the modular structure enables systematic transfer and generalization without full retraining.
How It Works
The system takes as input a PDE specification, a query time/location, and current state statistics. A regime-detection head computes feature embeddings from the input state that characterize which physical processes are dominant (e.g., advection-dominated vs. diffusion-dominated). A learned policy network then outputs a sequence of module selection decisions and duration scalars—essentially a program like 'apply advection for 0.3 time units, then diffusion for 0.2, then the learned closure for 0.5' and so on. Each module in the composition may be a classical numerical solver (with fixed step sizes) or a learned neural component, and they are evaluated sequentially with state passed between them. The final output is the PDE solution at the query time without requiring autoregressive step-by-step rollout, enabling arbitrary temporal resolution and faster inference. The policy is conditioned on both global PDE parameters and local state statistics, allowing the composition strategy to adapt dynamically.
Production Impact
Engineers building surrogate models for scientific computing would see immediate benefits: OOD generalization that's an order of magnitude better than current approaches means fewer domain-specific fine-tuning samples needed and more robust deployment to untrained parameter regimes. The ability to query solutions at arbitrary times without autoregressive unrolling translates to lower latency and reduced memory footprint—critical for interactive simulation or optimization loops. The modular design drastically reduces the cost of adapting to new physics: swapping a boundary condition or adding a residual term becomes a dictionary update rather than full retraining, which is transformative for practitioners who work with evolving or uncertain physics models. Trade-offs include: the policy learning adds complexity to training (you must now learn both module parameters and the selection policy), the framework requires pre-specifying a library of composable modules, and for very smooth, simple PDEs a simpler approach may suffice. Integration is straightforward if you already have numerical solvers available; the system acts as an intelligent scheduling layer on top of them.
Limitations and When Not to Use This
HyCOP's performance depends critically on the pre-specified module library—if the true solution requires a composition strategy not representable by available modules, the approach will fail. The framework assumes PDEs are sufficiently regular and that regime features (computed from state statistics) reliably predict which compositions will work; this may break down for highly chaotic or multiscale systems where the dominant process changes abruptly. The paper doesn't fully address how to systematically design the module library or how to choose regime features for new PDE families, making application to truly novel domains still require domain expertise. Finally, while interpretability is a stated benefit, 'readable programs' are only useful if they align with known physics; if the learned policies exploit implicit features of the training data, they could be misleading.
Research Context
This work builds on the neural operator literature (DeepONet, FNO, Fourier Neural Operators) by recognizing that monolithic maps are inefficient for PDEs where solution structure varies with regime. It connects to meta-learning and program synthesis traditions by framing the problem as learning a distribution over solution strategies rather than a single strategy. The modular composition idea parallels recent work in compositional generalization in language models and object-centric world models. HyCOP opens a research direction toward interpretable scientific machine learning—moving away from pure neural approximation toward hybrid numerical-neural methods that remain scientifically auditable, which aligns with growing concerns about deploying black-box surrogates in high-stakes domains like climate or materials science.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::