Physics-Informed Deep Learning (PIDL)

• Physics-Informed Deep Learning (PIDL) integrates deep neural networks with governing physical laws, such as PDEs, to enhance data efficiency and predictive accuracy.
• PIDL employs a composite loss function that balances observational data with physics residuals via automatic differentiation, ensuring robust constraint enforcement.
• This framework is applied across disciplines like traffic estimation, fluid dynamics, and materials science, resulting in reduced sensor requirements and improved uncertainty quantification.

Physics-Informed Deep Learning (PIDL) is a hybrid paradigm that integrates deep neural networks (DNNs) with physical knowledge, most commonly in the form of partial differential equations (PDEs) or other mechanistic system constraints. Unlike purely data-driven DNNs, PIDL explicitly encodes governing laws in the training or architecture of machine learning models. This approach unifies the expressivity and generalization of DNNs with the data efficiency, extrapolation, and physical fidelity of structured models. PIDL has proven effective across disciplines—from engineering (traffic estimation, materials, reliability, plasticity) and geophysics to fluid dynamics, uncertainty quantification, and scientific machine learning—by enabling the efficient assimilation of sparsely observed or noisy data with domain knowledge.

1. Core Principles and Mathematical Foundations

PIDL posits that the state variables $u(x,t)$ of a physical system can be modeled as neural network outputs $\hat{u}_\theta(x,t)$, where $\theta$ are trainable weights, and embeds the governing physical laws as constraints on these outputs. The dominant mechanism is the construction of a composite loss functional,

$$\mathcal{L}_{\mathrm{total}} = \mathcal{L}_{\mathrm{data}} + \lambda_{\mathrm{phys}}\mathcal{L}_{\mathrm{phys}} + \lambda_{\mathrm{BC/IC}}\mathcal{L}_{\mathrm{BC/IC}}.$$

Here, $\mathcal{L}_{\mathrm{data}}$ penalizes discrepancies between $\hat{u}_\theta$ and observations $u_{\mathrm{obs}}$ at known points, while $\mathcal{L}_{\mathrm{phys}}$ enforces satisfaction of PDEs or other constraints (as residuals over collocation points) via automatic differentiation. $\mathcal{L}_{\mathrm{BC/IC}}$ encodes essential (Dirichlet, initial, boundary) conditions.

For example, in traffic state estimation, the Lighthill–Whitham–Richards (LWR) PDE for density $\rho(x,t)$ is embedded as

$\hat{u}_\theta(x,t)$0

while system reliability is encoded by the Kolmogorov master equation residual (Zhou et al., 2021). For fluid dynamics, Navier–Stokes (N-S) residuals are imposed in diffusion-based models (Qiu et al., 2024). Hyperparameters $\hat{u}_\theta(x,t)$1, $\hat{u}_\theta(x,t)$2 balance data fit and physics regularization.

PIDL can also be used for parameter discovery by treating physical parameters (e.g., diffusivity, relaxation time, FD curve parameters) as trainable variables jointly optimized with network weights (Shi et al., 2021, Shi et al., 2021).

2. Architectural Patterns and Computational Graphs

PIDL architectures are categorized by the locus and mechanism of physics integration (Di et al., 2023):

• Physics-Guided Loss (“soft” regularization): The canonical structure (“PINN”) has a vanilla DNN whose output is subjected to physics residuals via autodiff. This enables enforcement of arbitrary PDEs/ODEs, including nonlocal and stochastic variants (Huang et al., 2023, Wang et al., 2024).
• Physics-Embedded Networks (“hard” embedding): Physical operators are implemented as fixed-weight convolutional layers (e.g., finite-difference stencils), “baked” into the network, as in PDE-preserved neural networks (PPNN) (Liu et al., 2022). This approach stabilizes long time-rollouts and allows for mesh discretization within DNNs.
• Hybrid or Modular Designs: Composite models (e.g., PIDL+FDL) decouple the system into state-predicting DNNs and neural surrogates for unknown closure laws (e.g., FD relations in traffic or free energies in materials), combined through a physics-informing layer (Shi et al., 2021, Bahtiri et al., 2024).
• Uncertainty Quantification (UQ) Extensions: GANs or normalizing flows supply stochastic predictions, with physics constraints enforced in generator, discriminator, or both (Mo et al., 2022, Daw et al., 2021).

The architecture may be parametric (accepting variable PDE coefficients as input for generalizability (Boudec et al., 2024)), ensemble-based (for UQ (Agata et al., 2024)), or variational (VAEs for distributional calibration in stochastic physical regimes (Wang et al., 2024)).

3. Training Methodologies, Loss Engineering, and Optimization

PIDL models are typically trained on mini-batches comprising both supervised (measurement) points and unsupervised collocation points. Loss balancing is often handled by hyperparameter search, but recent work proposes joint or neural-learner schemes for adaptive scheduling (Boudec et al., 2024). Training regimes commonly use Adam and/or second-order optimizers (L-BFGS) for global and local convergence.

When physics loss includes high-order derivatives or stiff systems (Tyagi et al., 2024), mixed formulations or additional outputs can reduce gradient pathologies. For example, in strain-gradient plasticity, auxiliary outputs for fourth-order terms can lower the order of PDEs imposed via residuals, improving numerical convergence.

For inverse problems and UQ (e.g., seismic inversion), ensemble variational training (e.g., function-space ParVI) can yield calibrated posteriors over physical fields and propagate uncertainty into downstream inference (Agata et al., 2024).

In GN/GAN approaches, adversarial losses and mutual-information regularizations mitigate issues of mode collapse and insufficient diversity (Mo et al., 2022, Daw et al., 2021), while physics-informed discriminators provide robust guidance even with imperfect or partial knowledge.

4. Representative Applications and Empirical Results

Transport Systems and Traffic Estimation

PIDL has demonstrated marked superiority over both purely model-driven (e.g., Kalman filtering) and data-driven baselines (NN/LSTM) for traffic state estimation, especially in the small-data regime (Shi et al., 2021, Shi et al., 2021, Di et al., 2023). For LWR-based highway TSE on the NGSIM dataset, PIDL achieves a target relative error (Err < 0.06) with only 8 detectors (vs. 12 for EKF and >12 for NN/LSTM); with 5 detectors, Err ≈ 0.04 (≈30–50% reduction in sensor density) (Shi et al., 2021).

Extensions to nonlocal traffic flow models further improve estimation accuracy by embedding realistic driver look-ahead effects, with optimal window sizes empirically determined (Huang et al., 2023). Stochastic PIDL models, fusing distributional (Beta, percentile-based) FDs, accurately reproduce macroscopic scattering while delivering tight confidence bands on state estimates (Wang et al., 2024).

Materials Science and Continuum Mechanics

In thermodynamically consistent constitutive modeling, PIDL architectures utilize LSTM+feedforward blocks to encode viscoelastic/plastic history, internal variables, and free energy, enforcing the Clausius–Duhem inequality via tailored loss terms. These models achieve stress prediction errors ≈0.008 even outside the training window and preserve positivity of dissipation (Bahtiri et al., 2024). For strain-gradient plasticity, the PIDL approach robustly resolves stiff microforce PDEs and captures concurrence with FE results across different hardening regimes (Tyagi et al., 2024).

Fluid Dynamics and Scientific Machine Learning

For fluid mechanics, physics-informed diffusion models (Pi-fusion) couple DDPM/score-based generative frameworks with PDE guidance in both training and sampling, achieving 3–10× lower errors than classic PINNs or domain-specific NN architectures (e.g., NSFnets, PIPN) (Qiu et al., 2024). Embedded physical constraints in the loss and in the diffusion process ensure high accuracy and interpretability for both synthetic and real 3D flow datasets.

PPNN approaches, which embed discretized operators as CNN residuals, prevent error accumulation in long rollouts on high-dimensional grids (reaction–diffusion, Burgers', N-S equations), outperforming black-box models by more than an order of magnitude in full-field $\hat{u}_\theta(x,t)$3 error (Liu et al., 2022).

UQ, Reliability, and Inverse Problems

PIDL enables continuous solutions to reliability models, e.g., multi-state Markov reliability, using PINN or PI-GAN architectures for both forward prediction and uncertainty quantification with significant acceleration over Monte Carlo (Zhou et al., 2021). In geophysics, ensemble-based PIDL propagates epistemic uncertainty from seismic velocity structures to earthquake hypocenter localization, rectifying bias and underestimation prevalent in conventional approaches (Agata et al., 2024). Bayesian and GAN-based extensions supply explicit UQ and fusion with sparse or noisy observations (Mo et al., 2022, Daw et al., 2021).

Biomedical Imaging

In PET kinetic modeling, physics-informed convolutional neural regressors for input function estimation (PIDLIF) maintain robustness in severe motion blur, with test $\hat{u}_\theta(x,t)$4 degradation from 0.94 to 0.89 (vs. 0.94 to 0.79 for baselines) and stable kinetic parameter estimation under out-of-distribution noise (Salomonsen et al., 24 Oct 2025).

5. Model Discovery and Scientific Insight

PIDL frameworks can be directly applied to closed-form discovery of governing laws in noisy, sparse-data regimes. Given observation data $\hat{u}_\theta(x,t)$5, model variables are regressed onto a library of candidate terms (e.g., polynomial, derivative, interaction) via sparse regularization (e.g., STRidge), yielding accurate PDE recovery for Burgers’, Kuramoto-Sivashinsky, N-S, and reaction–diffusion systems (Chen et al., 2020). This mechanism extends to systems with multi-IBC and experimental measurements (e.g., cell migration assays).

6. Limitations, Challenges, and Theoretical Insights

• Non-smooth/Hyperbolic PDEs: PINNs exhibit pronounced “spectral bias” and convergence failure for hyperbolic PDEs with shocks, unless a small diffusive regularization (parabolic augmentation) is included (Huang et al., 2023).
• Ill-conditioning: High-order derivatives in physics losses and multi-scale/stiff dynamics can severely impede optimization. Learned solver networks and mixed-form residuals alleviate this by conditioning gradients and lowering equation order (Boudec et al., 2024, Tyagi et al., 2024).
• Hyperparameter Sensitivity: Loss balancing, network depth/width, and collocation sampling all require careful tuning. Several works advocate for meta-learning, ADMM-like alternating minimization, or modular architectures to address these challenges (Di et al., 2023).
• Scalability: Large-scale, high-dimensional, or networked physical systems (traffic networks, 3D flows) necessitate advanced architectures (e.g., graph neural operators, domain-decomposition, or modular surrogates) for feasible PIDL deployment (Liu et al., 2022, Di et al., 2023).
• Distributional Generalization: Deterministic FDs in traffic models fail to capture observed scattering; stochastic and distributional priors in SPIDL remedy this while providing actionable UQ (Wang et al., 2024).

7. Future Directions and Extensions

Emerging themes for PIDL research encompass:

• Adaptive Physics Coding: Automatic inference/discovery of governing laws or closure terms through neural-architecture/search or symbolic regression (Chen et al., 2020).
• Modular and Meta-learning Solvers: Universal, PDE-parametric solver networks that generalize across families of equations or domains, meta-trained for rapid adaptation (Boudec et al., 2024).
• Hard and Soft Constraint Integration: Blending variational, energy-based, and weak-form formulations with flexible neural structures for improved stability and interpretability (Tyagi et al., 2024, Liu et al., 2022).
• Surrogate and Operator Learning: Multi-fidelity, operator-theoretic surrogates for fast emulation and inverse design, particularly where direct data is costly or sparse (Boudec et al., 2024, Sellke, 2023).
• Real-time and Edge Deployment: Integration with fog computing and IoT infrastructure to deliver traffic or physics field reconstructions on minimal data in real time (Huang et al., 2023).
• Multimodal and Heterogeneous Data Fusion: End-to-end hybrid models that incorporate image, trajectory, point cloud, and sensor data directly in physics-informed learning (Di et al., 2023).
• Probabilistic and Stochastic Physics Priors: Fully Bayesian or stochastic knowledge-infusion for robust uncertainty quantification and decision making under aleatoric and epistemic uncertainty (Agata et al., 2024, Wang et al., 2024, Mo et al., 2022).

Physics-Informed Deep Learning constitutes a versatile framework for fusing domain knowledge with machine learning, unlocking data-driven discovery, efficient inference, and robust prediction across diverse scientific and engineering domains. Its ongoing evolution is characterized by increasingly sophisticated architecture, robust optimization, uncertainty awareness, and broadening applicability to real-world, noisy, and complex systems.