Hybrid-Cooperative Learning (HYCO)

• Hybrid-Cooperative Learning (HYCO) is a framework that combines a physics-based model and a data-driven model through mutual regularization.
• It employs a game-theoretic, alternating optimization scheme where each model retains its own learning objective while aligning predictions.
• HYCO enhances performance in sparse, noisy, or heterogeneous data regimes by leveraging complementary strengths of both model types.

Hybrid-Cooperative Learning (HYCO) is a hybrid modeling framework in which a physics-based model and a data-driven model are trained side by side and coupled through a mutual regularization mechanism. Rather than imposing physical constraints directly on a single synthetic model, HYCO treats the physical and synthetic components as co-trained agents that are nudged toward agreement while each retains its own learning objective. In the control-and-machine-learning survey literature, HYCO is positioned as a methodology for hybridizing mechanics-based modeling with data-driven learning in a game-theoretic setting; in the dedicated PDE-modeling papers, it is developed as a framework for static and time-dependent inverse problems under sparse, noisy, localized, or heterogeneous observations (Liverani et al., 17 Sep 2025, Zuazua, 30 Sep 2025).

1. Conceptual origin and methodological identity

HYCO appears in the broader literature on the interface of control theory, neural networks, PDEs, and scientific machine learning as a response to a specific limitation of single-architecture hybrid methods. The survey formulation states that, unlike physics-informed neural networks (PINNs), which enforce physical and data constraints directly on a single synthetic model, HYCO trains two models in parallel: one grounded in physical principles, typically formulated as an ODE or PDE, and another driven by data, using a neural-network ansatz. Their interaction is described as being “much like two experts exchanging insights before reaching consensus,” and the overall setup is explicitly framed as game-theoretical (Zuazua, 30 Sep 2025).

This distinction is definitional rather than cosmetic. In HYCO, the physical model is not merely a regularizer attached to a neural network, and the neural network is not merely a surrogate constrained by PDE residuals. The two components remain separate learners. The physical branch contributes inductive bias, interpretability, and extrapolative structure; the synthetic branch contributes flexibility and data adaptation. The cooperative mechanism is the requirement that their predictions agree more closely over the domain.

The detailed HYCO papers formulate this paradigm for data-driven PDE modeling. They emphasize that the framework is intended for settings in which pure data-driven models extrapolate poorly or become fragile in low-data regimes, while pure physics-based solvers may be rigid, difficult to calibrate, or underdetermined when model form or coefficients are uncertain. A plausible implication is that HYCO is best understood as a two-model regularization architecture for inverse and reconstruction problems rather than as a variant of residual-based physics-informed learning (Liverani et al., 17 Sep 2025).

2. Two-model formulation

HYCO assumes two outputs: the prediction $u_{\mathrm{phy}}$ from a physical model with parameters $\Lambda$, and the prediction $u_{\mathrm{syn}}$ from a synthetic model with parameters $\Theta$. The physical model is generally a PDE or ODE solver, often implemented through finite differences or finite elements, with unknown quantities grouped as

$\Lambda := \{\kappa(x), \lambda(x), \ldots\}.$

The synthetic model is any trainable data-driven approximator. Although the experiments use neural networks, the formulation explicitly allows alternatives such as a Gaussian mixture, least-squares fit, or a learned dynamical system (Liverani et al., 17 Sep 2025).

In the formalism paper, the component losses are defined as

$$L_{\mathrm{phy}}(\Lambda) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{phy}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Lambda),$$

$$L_{\mathrm{syn}}(\Theta) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{syn}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Theta),$$

and the coupling term

$$L_{\mathrm{int}}(\Theta, \Lambda) := \int_{\Omega} \|u_{\mathrm{syn}}(x) - u_{\mathrm{phy}}(x)\|^2\,dx.$$

The overall objective is

$$\min_{\Theta, \Lambda}\, \alpha L_{\mathrm{syn}}(\Theta) + \beta L_{\mathrm{phy}}(\Lambda) + L_{\mathrm{int}}(\Theta, \Lambda),$$

with $\alpha,\beta \ge 0$. Either weight may be set to zero, so one model can be trained only through interaction with the other (Liverani et al., 27 Feb 2026).

The time-dependent formulation extends the same structure over space-time. In the extended preprint, the data-fit losses are written over $\Lambda$0 sampled spatial locations and $\Lambda$1 time instants, while the interaction term is expressed as

$\Lambda$2

This makes HYCO applicable to both static inverse problems and dynamic PDE identification (Liverani et al., 17 Sep 2025).

3. Cooperative optimization and game-theoretic interpretation

The dedicated HYCO papers interpret the coupled training process as a two-player game. The physical model minimizes

$\Lambda$3

while the synthetic model minimizes

$\Lambda$4

A Nash equilibrium $\Lambda$5 is defined by the usual unilateral optimality conditions: $\Lambda$6 for all admissible $\Lambda$7 (Liverani et al., 17 Sep 2025).

This interpretation leads directly to alternating optimization. In the formalism paper, the update scheme is written as

$\Lambda$8

The extended preprint states the same logic procedurally: fix $\Lambda$9 and update $u_{\mathrm{syn}}$0, then fix $u_{\mathrm{syn}}$1 and update $u_{\mathrm{syn}}$2, using stochastic gradient descent or Adam (Liverani et al., 27 Feb 2026, Liverani et al., 17 Sep 2025).

The computational structure is deliberately modular. The interaction loss is not evaluated densely over the full grid at each step. Instead, it is approximated using random “ghost points,” auxiliary points that are not part of the observational dataset and serve only to estimate agreement between the two models. In the formalism paper, the approximation is written as

$u_{\mathrm{syn}}$3

Because the two branches are independent apart from this synchronization step, the framework is described as naturally parallelizable; the models need not exchange raw data or parameters, only predictions or residual summaries (Liverani et al., 27 Feb 2026).

For parameter-identification experiments, the extended preprint uses early stopping based on stabilization of the physical parameters: $u_{\mathrm{syn}}$4 with $u_{\mathrm{syn}}$5 and $u_{\mathrm{syn}}$6. This operationalizes convergence in practice without claiming a general convergence theorem (Liverani et al., 17 Sep 2025).

4. Data regimes, modeling flexibility, and implementation

A central advertised property of HYCO is that the two models need not observe identical data. One model may receive solution values, the other source terms or fluxes, or one model may even receive no direct data at all. The datasets may be partially overlapping or distinct; the only required communication is via model outputs. This is a substantive departure from architectures in which all information sources are fused into a single constrained network (Liverani et al., 17 Sep 2025).

The framework is designed for several challenging observation regimes. In the Helmholtz and heat experiments, the observation region is progressively reduced from the full domain $u_{\mathrm{syn}}$7 to

$u_{\mathrm{syn}}$8

The intended effect is that the interaction term propagates information from the observed region into the unobserved one. In the Darcy-flow appendix, multiplicative uniform noise up to $u_{\mathrm{syn}}$9 is added to the observations. The reported conclusion is that HYCO remains competitive with FEM parameter fitting and significantly outperforms PINNs in that noisy setting (Liverani et al., 17 Sep 2025).

The implementation details are correspondingly heterogeneous. For Gray–Scott, the synthetic model is a feedforward neural network with 4 hidden layers, 128 neurons per layer, and ReLU activation. For Helmholtz, Heat, and Darcy, the synthetic models use 2 hidden layers of width 256, ReLU, and residual connections. PINN baselines use the same depth and width but $\Theta$0, because PINNs require smooth activations for PDE residual differentiation. Adam is used throughout. Typical training schedules are 600 epochs for HYCO and NN, and 3500 for PINN in Gray–Scott; 3000 epochs in Helmholtz and Heat; and 2000 epochs in Darcy (Liverani et al., 17 Sep 2025).

A recurring misconception is to treat HYCO as requiring direct supervision of both branches. The Gray–Scott experiment explicitly sets $\Theta$1, so the physical model receives no direct data and is trained only through the interaction term: $\Theta$2 with $\Theta$3. This demonstrates that HYCO’s coupling can operate even when one branch is unsupervised in the standard empirical sense (Liverani et al., 17 Sep 2025).

5. Benchmark problems and reported performance

The empirical literature evaluates HYCO on both static and time-dependent PDE problems. The flagship dynamic example is the Gray–Scott reaction-diffusion system,

$\Theta$4

with the task of reconstructing the full solution and identifying $\Theta$5 from $\Theta$6 irregularly sampled space-time observations on a $\Theta$7 reference simulation. HYCO uses $\Theta$8 ghost points. The final normalized $\Theta$9 solution errors are reported as $\Lambda := \{\kappa(x), \lambda(x), \ldots\}.$0 for HYCO Physical, $\Lambda := \{\kappa(x), \lambda(x), \ldots\}.$1 for HYCO Synthetic, $\Lambda := \{\kappa(x), \lambda(x), \ldots\}.$2 for PINN, and $\Lambda := \{\kappa(x), \lambda(x), \ldots\}.$3 for NN. The recovered diffusivities are $\Lambda := \{\kappa(x), \lambda(x), \ldots\}.$4 and $\Lambda := \{\kappa(x), \lambda(x), \ldots\}.$5, compared with the ground truth $\Lambda := \{\kappa(x), \lambda(x), \ldots\}.$6 and $\Lambda := \{\kappa(x), \lambda(x), \ldots\}.$7; the PINN estimates are $\Lambda := \{\kappa(x), \lambda(x), \ldots\}.$8 and $\Lambda := \{\kappa(x), \lambda(x), \ldots\}.$9 (Liverani et al., 17 Sep 2025).

The static inverse benchmark is a heterogeneous Helmholtz equation on $$L_{\mathrm{phy}}(\Lambda) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{phy}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Lambda),$$0,

$$L_{\mathrm{phy}}(\Lambda) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{phy}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Lambda),$$1

with

$$L_{\mathrm{phy}}(\Lambda) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{phy}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Lambda),$$2

The task is to recover

$$L_{\mathrm{phy}}(\Lambda) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{phy}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Lambda),$$3

from only $$L_{\mathrm{phy}}(\Lambda) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{phy}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Lambda),$$4 sensor measurements. On the hardest regime $$L_{\mathrm{phy}}(\Lambda) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{phy}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Lambda),$$5, the parameter errors are $$L_{\mathrm{phy}}(\Lambda) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{phy}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Lambda),$$6 for FEM, $$L_{\mathrm{phy}}(\Lambda) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{phy}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Lambda),$$7 for HYCO Physical, and $$L_{\mathrm{phy}}(\Lambda) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{phy}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Lambda),$$8 for PINN. The corresponding solution errors are $$L_{\mathrm{phy}}(\Lambda) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{phy}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Lambda),$$9 for HYCO Physical, $$L_{\mathrm{syn}}(\Theta) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{syn}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Theta),$$0 for HYCO Synthetic, $$L_{\mathrm{syn}}(\Theta) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{syn}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Theta),$$1 for PINN, and $$L_{\mathrm{syn}}(\Theta) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{syn}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Theta),$$2 for FEM (Liverani et al., 17 Sep 2025).

A second time-dependent inverse problem uses the heat equation

$$L_{\mathrm{syn}}(\Theta) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{syn}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Theta),$$3

with $$L_{\mathrm{syn}}(\Theta) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{syn}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Theta),$$4 time steps and $$L_{\mathrm{syn}}(\Theta) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{syn}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Theta),$$5 spatial samples per time step. On the hardest regime $$L_{\mathrm{syn}}(\Theta) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{syn}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Theta),$$6, the reported parameter errors are $$L_{\mathrm{syn}}(\Theta) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{syn}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Theta),$$7 for FD, $$L_{\mathrm{syn}}(\Theta) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{syn}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Theta),$$8 for HYCO Physical, and $$L_{\mathrm{syn}}(\Theta) := \frac{1}{M} \sum_{i=1}^M \ell\!\left(u_{\mathrm{syn}}(x_i), u^D(x_i)\right) + \mathscr{P}(\Theta),$$9 for PINN; the solution errors are $$L_{\mathrm{int}}(\Theta, \Lambda) := \int_{\Omega} \|u_{\mathrm{syn}}(x) - u_{\mathrm{phy}}(x)\|^2\,dx.$$0 for FD, $$L_{\mathrm{int}}(\Theta, \Lambda) := \int_{\Omega} \|u_{\mathrm{syn}}(x) - u_{\mathrm{phy}}(x)\|^2\,dx.$$1 for HYCO Physical, $$L_{\mathrm{int}}(\Theta, \Lambda) := \int_{\Omega} \|u_{\mathrm{syn}}(x) - u_{\mathrm{phy}}(x)\|^2\,dx.$$2 for HYCO Synthetic, and $$L_{\mathrm{int}}(\Theta, \Lambda) := \int_{\Omega} \|u_{\mathrm{syn}}(x) - u_{\mathrm{phy}}(x)\|^2\,dx.$$3 for PINN. In the Darcy appendix with $$L_{\mathrm{int}}(\Theta, \Lambda) := \int_{\Omega} \|u_{\mathrm{syn}}(x) - u_{\mathrm{phy}}(x)\|^2\,dx.$$4, $$L_{\mathrm{int}}(\Theta, \Lambda) := \int_{\Omega} \|u_{\mathrm{syn}}(x) - u_{\mathrm{phy}}(x)\|^2\,dx.$$5, and $$L_{\mathrm{int}}(\Theta, \Lambda) := \int_{\Omega} \|u_{\mathrm{syn}}(x) - u_{\mathrm{phy}}(x)\|^2\,dx.$$6 noise, HYCO and FEM are described as broadly comparable, with HYCO slightly improving solution approximation relative to FEM as noise increases, while both HYCO and FEM outperform PINNs in parameter recovery (Liverani et al., 17 Sep 2025).

These results support the papers’ stated interpretation of HYCO as a mutual regularization loop: the synthetic model supplies a flexible target for the physical model, and the physical model supplies PDE structure that stabilizes and extrapolates beyond sparse observations.

6. Theoretical status, scope, and related usages

The theoretical status of HYCO remains explicitly open. The survey paper that first positions HYCO in the control–machine-learning interface states that the framework “opens a new perspective” but also “raises fundamental questions regarding convergence guarantees with rates that merit careful investigation.” In that survey venue, no HYCO-specific optimization equations, variational principles, convergence theorem, error bound, or stability result are given. The connection to neural approximation theory is also only conceptual: the data-driven component uses a neural-network ansatz, but no HYCO-specific universal approximation or approximation-rate result is proved there (Zuazua, 30 Sep 2025).

The later papers provide the formal optimization framework and numerical evidence, but the main theoretical narrative remains interpretive rather than theorem-driven. The game-theoretic reading is cooperative rather than adversarial: both players minimize losses that share the same interaction term, and the equilibrium notion is invoked to justify alternating optimization, not to introduce competitive payoffs in the usual adversarial-learning sense (Liverani et al., 27 Feb 2026).

The term “hybrid” and the language of cooperation also appear in other domains, but those usages are methodologically distinct from HYCO as defined for PDE modeling. “Grouped Hybrid Q Learning” (GHQ) addresses heterogeneous cooperative multi-agent reinforcement learning by grouping agents, using hybrid value factorization, and maximizing inter-group mutual information (Yu et al., 2023). “HC$$L_{\mathrm{int}}(\Theta, \Lambda) := \int_{\Omega} \|u_{\mathrm{syn}}(x) - u_{\mathrm{phy}}(x)\|^2\,dx.$$7L” denotes “Hybrid and Cooperative Contrastive Learning” for zero-shot cross-lingual spoken language understanding, combining one unsupervised and three supervised contrastive mechanisms (Xing et al., 2024). In model-based reinforcement learning, “Nested Mixture of Experts” is described as conceptually aligned with a hybrid-cooperative perspective because it uses competitive specialization across modes and cooperative white-box/black-box blending within each mode, though it does not use the term HYCO explicitly (Ahn et al., 2020). These parallels suggest a broader methodological family centered on structured interaction among heterogeneous learners, but HYCO proper refers to the physics-model/synthetic-model co-training framework developed for data-driven PDE modeling.

Within that scope, HYCO’s defining features are stable across the literature: separate physical and synthetic branches, mutual regularization through an interaction loss, an alternating optimization scheme with a game-theoretic interpretation, and a focus on sparse, noisy, localized, or heterogeneous scientific data. The central unresolved issue is not empirical viability but theory: convergence guarantees with rates remain the most clearly identified open problem (Zuazua, 30 Sep 2025).