Physics-encoded Recurrent CNNs (PeRCNNs)

• PeRCNNs are deep learning architectures that combine CNN-based spatial extraction and RNN-based temporal modeling with explicit physics encoding to simulate PDE-governed systems.
• They incorporate hardwired inductive biases such as fixed convolution kernels and physics-informed loss functions to achieve rapid, robust surrogate modeling across various physical phenomena.
• PeRCNNs offer enhanced extrapolability, interpretability, and computational efficiency compared to traditional black-box and PINN approaches, making them valuable for complex scientific simulations.

Physics-encoded Recurrent Convolutional Neural Networks (PeRCNNs) are an emerging class of deep learning architectures designed to simulate, forecast, and analyze spatiotemporal physical systems governed by partial differential equations (PDEs). These networks integrate convolutional neural networks (CNNs) for spatial feature extraction with recurrent neural networks (RNNs) for temporal modeling, while embedding explicit or inductive representations of physical laws through either hard-coded stencils, differentiator-integrator blocks, or physics-informed loss functions. PeRCNNs have enabled rapid surrogate modeling of phenomena ranging from fluid flow and reaction–diffusion systems to energetic material response, while offering superior extrapolability and improved treatment of boundary and initial conditions relative to conventional neural PDE solvers.

1. Foundational Principles and Network Architectures

PeRCNNs originate from the recognition that direct encoding of physics into neural architectures can overcome major limitations of end-to-end black-box surrogates and traditional PINN-style fully-connected approaches. At their core, PeRCNNs employ convolution-based operators to extract spatial derivatives, combined with temporal recurrence modules—either explicit (e.g., CNN-based time-stepping, ConvLSTM, differentiator-integrator pairs) or implicit (autoregressive feed-forward updates).

A representative instance is the PhyCRNet architecture (Ren et al., 2021), which utilizes:

• A multi-stage spatial encoder (stacked convolutions, periodic padding) to down-sample high-resolution field inputs (e.g., $u(x,y)$, $v(x,y)$).
• A ConvLSTM as temporal propagator for sequential latent state evolution, governed by standard equations:

$$\mathbf i_t = \sigma(W_i * [\mathbf X_t,\mathbf h_{t-1}] + \mathbf b_i),\;\ldots$$

• A decoder leveraging pixel-shuffle upscaling and a residual (forward Euler) connection:

$$\mathbf u_{i+1} = \mathbf u_i + \delta t \cdot \mathcal{NN}[\mathbf u_i;\theta]$$

Other forms include PARC and its v2 variant (Nguyen et al., 2022, Nguyen et al., 2024), featuring a two-module structure: a CNN “differentiator” approximates spatial PDE right-hand-sides, and a CNN “integrator” advances fields in time through explicit integration—a digital analog of operator splitting schemes.

PhICNet (Saha et al., 2020) further incorporates source-identification cells, augmenting the recurrent convolution cell with learned internal states to capture unobservable and time-varying external source maps.

The architectures are universally designed to operate on field arrays defined over regular grids (typically $64\times64$ to $240\times240$), with options for periodic, Dirichlet, or Neumann boundary conditions enforced by convolutional padding schemes or ghost-point embeddings.

2. Mechanisms of Physics Encoding

Physics encoding in PeRCNNs addresses both the representation of physical laws and the imposition of constraints.

Hardwired Inductive Bias

Several PeRCNN variants, including PARC/PARCv2 and PhICNet, rely on hard-coded convolution kernels for spatial derivatives (e.g., central-difference, Laplacian, gradient stencils), with only physical constants (e.g. diffusivity $\alpha$, advection velocity $u$) as trainable parameters. For instance, PARCv2 evaluates:

$$\frac{\partial x}{\partial t} = k\Delta x - (u\cdot\nabla)x+R_x(x,u,c)$$

where $\Delta$ is realized by fixed kernels:

$K_{\Delta} = \frac{1}{\Delta x^2}\begin{bmatrix}0&1&0\1&-4&1\0&1&0\end{bmatrix}$

and advection terms as:

$$\text{conv}(u_x \cdot \partial_x x + u_y \cdot \partial_y x)$$

Physics-informed Losses

PhyCRNet employs a loss function consisting solely of the summed PDE residual (no direct label supervision):

$$\mathcal L(\theta) = \sum_{k,i,j} \|\mathcal R(x_i,y_j,t_k;\theta)\|_2^2$$

where

$$\mathcal R(\mathbf x, t;\theta) = \partial_t \mathbf u^\theta + \mathcal F[\mathbf u^\theta, \nabla \mathbf u^\theta, \nabla^2 \mathbf u^\theta, \dots]$$

With hard-encoded initial/boundary conditions, solutions are forcibly constrained to satisfy physical laws and domain specifications at all grid points and times.

Differentiator–Integrator Decomposition

PARC (Nguyen et al., 2022) and PARCv2 (Nguyen et al., 2024) formalize physical modeling as the composition of two learned CNNs: one for the time-derivative operator, one for its time-integral. This mirrors numerical ODE solvers:

• Differentiator $f(X,\mu\mid\theta)$ approximates $dX/dt$,
• Integrator $S(\dot X \mid \phi)$ approximates $\int_{t_k}^{t_{k+1}} f(X(t),\mu)\,dt$.

No explicit PDE-residual loss is used; the main inductive bias is architectural.

Source Separation

PhICNet introduces internal recurrent source states $V_t$, isolating homogeneous solutions from unknown exogenous drives ($v(x,y,t)$). Optimization targets both forecast error and source consistency/sparsity.

3. Training Methodologies and Computational Efficiency

PeRCNN training protocols are tailored to exploit the structure enforced by physics encoding:

• PhyCRNet adopts Adam, weight normalization (no batch norm), learning rates decayed by 1% every 50 epochs, and training windows up to 10,000 epochs on NVIDIA V100 GPUs (Ren et al., 2021).
• AR rollouts (autoregressive input/output chaining) underpin both training and inference, forcing explicit one-step predictions and minimizing drift.
• PARCv2 divides the training into a two-stage process:
  • Stage 1: Optimize differentiator weights ($\theta$), freeze integrator.
  • Stage 2: Freeze differentiator, optimize integrator ($\phi$), using data-only $\ell_1$ losses on one-step rollout errors.

Evaluation metrics are problem-specific but include RMSE over field values, signal-to-noise ratios (SNR), Pearson correlation coefficients for source identification, and physical quantities-of-interest (QoIs) such as hotspot temperature and area (Nguyen et al., 2022, Nguyen et al., 2024).

Computational speed comparisons denote orders-of-magnitude improvements over direct numerical simulation (DNS). For instance, PARC reduces a DNS run from $\sim$24 h (1,324 cores + GPU) to 0.5 s on GPU (Nguyen et al., 2022).

4. Representative Applications and Quantitative Outcomes

PeRCNNs have demonstrated state-of-the-art performance in several canonical and applied PDE contexts:

PDE/System	Network	Key Metric(s)	Value(s)	Relative Performance
2D Burgers’	PhyCRNet	a-RMSE ($t=1$–$4$)	< $10^{-2}$	$10^{2}$ lower than AR-DenseED, stable in extrapolation (Ren et al., 2021)
$\lambda$–$\omega$ RD	PhyCRNet	a-RMSE extrapolation	$10^{-3}$–$10^{-2}$	1–2 orders lower than PINN (Ren et al., 2021)
FitzHugh–Nagumo	PhyCRNet	a-RMSE ($t\leq 9$)	< $10^{-2}$	PINN/AR-DenseED: $O(1)$
Heat Equation	PhICNet	SNR ($t+50$)	28 dB	Baselines: 17–24 dB
NS Cylinder Wake	PARCv2	RMSE$_u$	0.0727 m/s	FNO: 0.2147, PhyCRNet: 0.2094
Energetic Materials (EM)	PARC	RMSE$_T$	374 K	Correlation: $>$0.8 for area, $~$0.4 for temp. (Nguyen et al., 2022)

PARCv2 (Nguyen et al., 2024) has further shown stable long-term integration properties in advection-dominated flows (Burgers’, NS), retaining phase lock over 20 vortex-shedding periods, and is able to resolve shock fronts and hotspot dynamics in energetic material simulations.

PhICNet demonstrates concurrent forecasting and identification of highly sparse, evolving source fields with robust correlation to ground truth, and is resilient to additive noise.

5. Advantages, Limitations, and Prospective Extensions

Key strengths of PeRCNNs, as substantiated in published results:

• Extrapolability and Generalization: Error profiles remain flat across both training and extrapolation regimes, with solutions robust to initial condition shifts and domain variations. Hard boundary encoding suppresses boundary blow-up/drift (Ren et al., 2021).
• Computational Speed: Surrogate deployment cuts simulation cost by 10$^3$–10$^4\times$ compared to DNS, with modest hardware requirements (Nguyen et al., 2022).
• Treatment of BCs/ICs: Hard encoding (periodic padding or ghost-point) yields exact performance at boundaries/initialization, a longstanding limitation in soft-penalized PINN methods.
• Interpretability: Architectural separation (e.g., microstructure encoder, differentiator) allows for direct attribution mapping (saliency analyses) and reveals mechanistic links between field evolution and input geometry (Nguyen et al., 2022).
• Source Modeling: PhICNet's internal source state modeling permits both source forecast and identification in systems where exogenous drivers are unknown or unmeasured (Saha et al., 2020).

Limitations include the present restriction to regular grid domains (extension to unstructured meshes plausibly via graph-based convolution), handling only periodic Dirichlet BCs without custom embedding for more general BCs, and potential need for additional hybridization to admit higher-order time stepping or system-specific constitutive laws.

6. Interpretability and Physical Insights

PeRCNN architectures such as PARC provide interpretable internal structures. The separation of microstructure encoding and dynamical operators permits the computation of pixel-level saliency maps, revealing which morphological features contribute most to physical outcomes (e.g., hotspot initiation in EM). Statistical morphometry analyses quantify critical vs noncritical voids and their correlation to shock orientation and hotspot production, matching physical intuition derived from DNS (Nguyen et al., 2022).

PARCv2's hybrid differentiator–integrator design not only improves stability and long-time integration but offers modularity for direct analysis of component contributions—advection, diffusion, reaction—within the modeled field dynamics.

A plausible implication is that PeRCNN frameworks can provide a route to data-driven model discovery and hypothesis testing in complex physical systems, combining predictive accuracy with mechanistic insight.

7. Research Trajectory and Integration in Scientific Computing

PeRCNNs, as exemplified by PhyCRNet (Ren et al., 2021), PhICNet (Saha et al., 2020), PARC (Nguyen et al., 2022), and PARCv2 (Nguyen et al., 2024), reflect an overview of classical operator theory, numerical PDE methods, and deep learning inductive bias strategies. Their adoption addresses the increasing demand for fast, extrapolable surrogates in simulation-based engineering and scientific computing, and enables model parameter identification, inverse problem-solving, and source attribution in challenging dynamical systems.

Ongoing trajectories involve the generalization to irregular/graph-based domains, incorporation of higher-order solvers and multi-physics phenomena, and integration with explicit constraint-enforcement and uncertainty quantification modules. The performance benchmarks substantiate PeRCNNs’ status as key tools in physics-aware deep learning for spatiotemporal modeling.