Mathematical Framework of NeuralFlow

NeuralFlow is a next-generation CFD solver that combines advanced finite-volume methods with artificial intelligence to accelerate fluid simulations. Instead of replacing the CFD solver, NeuralFlow uses a Graph Neural Operator as a trainable correction engine. Starting from a CFD state, the neural operator performs several rollout correction steps on velocity, pressure, and temperature, with the objective of reaching a physically consistent solution in far fewer steps than a conventional solver iteration history.

Finite-Volume Conservation Form

Classical physics-informed neural networks usually enforce the pointwise differential form of the governing equations. This can be difficult for real CFD flows containing shocks, thin boundary layers, steep gradients, or discontinuities.

NeuralFlow instead works with the integral finite-volume form of the conservation laws. For each control volume, the solver measures the net conservation imbalance through the surrounding faces.

\( \mathcal{B}_i(V) = \sum_{f \in \partial \Omega_i} \mathcal{F}_f(V) A_f - S_i(V)\Omega_i = 0 \)

Here, \( \mathcal{B}_i(V) \) is the conservation imbalance of cell \( i \), \( \mathcal{F}_f(V) \) is the numerical flux through face \( f \), \( A_f \) is the face area, \( S_i(V) \) is the source term, and \( \Omega_i \) is the control-volume size.

This formulation directly enforces conservation of mass, momentum, and energy in the same way as a real CFD solver.

Graph Neural Operator Correction

The neural model does not directly replace the CFD solution. Instead, it predicts corrections to the current flow state.

\( V^{k+1} = V^k + \Delta V_\theta^k \)

where \( V^k \) is the current state and \( \Delta V_\theta^k \) is the correction proposed by the Graph Neural Operator. The correction is applied to the primitive flow variables, such as velocity, pressure, and temperature.

The purpose of the network is to learn a compressed convergence path. A classical implicit CFD solver may need many iterations to reach a converged state. NeuralFlow tries to learn how to reach a useful, physically consistent solution in only a small number of learned correction steps.

Multi-Rollout and Episode-Based Training

NeuralFlow uses multi-rollout correction. Starting from a fixed training state, the Graph Neural Operator applies several learned corrections in sequence:

\( V^0 \rightarrow V^1 \rightarrow V^2 \rightarrow \cdots \rightarrow V^m \)

After the rollout sequence, the CFD solver evaluates how well the final field satisfies the conservation laws.

Training is organized into episodes. Each episode starts from a fixed training anchor, performs a limited number of rollout/update steps, and then measures the true online conservation error at the end of the episode.

This gives the method a reinforcement-learning flavor. The CFD state acts as the environment, the neural operator proposes correction actions, and the conservation imbalance provides the physics-based feedback. However, the feedback is not a black-box reward. It is generated by an advanced implicit finite-volume CFD solver.

Learning from the Implicit Solver Methodology

The feedback used by NeuralFlow is not just a scalar residual value. It comes from the numerical methodology of the implicit CFD solver. The solver assembles the conservation imbalance and computes the sensitivity of the equations to the flow variables through flux Jacobians.

\( A = \frac{\partial \mathcal{B}}{\partial V} \)

This matrix contains the linearized finite-volume structure of the solver, including flux coupling, variable interactions, boundary-condition effects, and the sensitivity of the conservation equations to the solution state.

Therefore, the GNO learns not only from conservation laws, but also from the proven numerical structure of an advanced implicit CFD method.

Physics Loss and Solver-Guided Gradient

A physics-based objective can be written as:

\( \mathcal{J}_{phys}(V) = \frac{1}{2} \mathcal{B}(V)^T W \mathcal{B}(V) \)

where \( W \) is a weighting operator used to scale or emphasize different parts of the conservation imbalance.

The gradient with respect to the flow state is obtained through the solver's linearized system:

\( \frac{\partial \mathcal{J}_{phys}}{\partial V} = A^T W \mathcal{B}(V) \)

This is the key connection between the CFD solver and the neural network. The solver provides the physics-based direction in solution space, and the neural network learns parameters that generate better correction steps.

Dual Automatic Differentiation

NeuralFlow uses two forms of automatic differentiation.

Forward-mode automatic differentiation is used inside the CFD solver to compute flux Jacobians and the sensitivity of the conservation equations with respect to the predicted flow variables.

Reverse-mode automatic differentiation is used inside the neural network to update its trainable parameters through backpropagation.

\( \frac{\partial \mathcal{J}_{phys}}{\partial \theta} = \frac{\partial \mathcal{J}_{phys}}{\partial V} \frac{\partial V}{\partial \theta} \)

This dual-AD strategy allows NeuralFlow to combine CFD-grade flux linearization with neural-network training. The solver computes how the conservation equations respond to the flow state, and the neural network learns how to produce corrections that reduce the conservation imbalance more efficiently.

Memory-Efficient Episode Learning

Backpropagating through a long sequence of learned CFD corrections would require storing many intermediate graph states, messages, tensors, flow fields, and solver quantities. For realistic finite-volume meshes, this memory cost can become too high.

Episode-based training controls this cost by using short rollout windows. The model learns from the physics feedback over a limited episode, and the true online conservation error is measured at the end of the episode.

This makes the method practical for large CFD problems while still allowing the GNO to learn accelerated convergence behavior.

NeuralFlow Compared with Classical PINNs

Classical PINNs differentiate the governing equations directly and enforce the pointwise PDE form. This can become fragile near shocks, discontinuities, and thin boundary layers.

NeuralFlow instead integrates fluxes over control volumes. It uses solver-proven numerical flux methods, such as AUSM-type fluxes, and trains with solver-computed Jacobians. This gives the method the robustness of finite-volume CFD while retaining the learning capability of neural operators.

Conclusion

NeuralFlow is designed to learn a compressed convergence path for CFD simulations. It does not replace the governing equations or the CFD solver methodology. Instead, it uses a Graph Neural Operator as a trainable correction engine guided by conservation laws, finite-volume fluxes, exact solver Jacobians, and implicit CFD linearization.

The goal is to reach physically consistent CFD solutions in far fewer steps, while preserving the conservation, robustness, and numerical knowledge of advanced CFD solvers.