PhysSFI-Net: Physics-Based Neural Models

• PhysSFI-Net is a suite of neural frameworks that embed explicit physics laws into architectures, ensuring high-fidelity and interpretable predictions across diverse domains.
• It employs designs like convolutional RNNs, graph-based models, and neural operators to address applications from facial deformation prediction to fluid-structure interaction analysis.
• The approach rigorously enforces governing equations (e.g., PDEs and Navier–Stokes) within its loss formulations, yielding enhanced stability and performance versus traditional methods.

PhysSFI-Net encompasses a family of distinct methods that leverage explicit physics-based constraints and inductive biases for neural modeling of complex spatiotemporal, structural, or source-driven systems. Across its independent instantiations, PhysSFI-Net architecture incorporates domain-governed equations (such as PDEs), geometric constraints, or interface coupling, while employing neural architectures—convolutional RNNs, neural operators, or geometric deep learning—for super-resolution, source inference, or coupled state estimation. The following provides an in-depth review of PhysSFI-Net within three principal formulations: deep geometric learning for skeletal-facial interactions in orthognathic surgery (Bao et al., 5 Jan 2026), physics-informed neural inference of fluid-structure interaction (Tang et al., 30 Jun 2025), and end-to-end physics-incorporated convolutional RNNs for source identification and forecasting in spatiotemporal PDE systems (Saha et al., 2020).

1. Principal Architectures and Design Rationale

PhysSFI-Net is not a single architecture but denotes several frameworks unified by the concept of embedding physics constraints as primary inductive bias in neural architectures. Each variant is tailored to a specific domain:

• Geometric PhysSFI-Net (Bao et al., 5 Jan 2026): Designed for postoperative facial morphology prediction, input data (preoperative skeleton, post-operative skeleton, preoperative face) are represented as discrete point clouds on $\mathbb{R}^3$ manifolds. The model extracts multi-scale geometric features via a hierarchical graph convolution, fuses semantic and structural information with attention mechanisms, and predicts incremental soft-tissue deformations with an LSTM-based sequential process. A biomechanics-inspired graph-harmonic module reconstructs high-resolution facial surfaces.
• FSI PhysSFI-Net (Tang et al., 30 Jun 2025): Targets the inverse problem of reconstructing both unsteady flow (fluid) and structure states from sparse Lagrangian observations. Three networks—coordinate PINN for fluid state, modal PINN for solid deformation, and hard-constrained advection modules for tracked trajectories—are trained jointly, subject to physics loss (Navier–Stokes), interface kinematics (no-slip), and exact particle advection.
• Spatiotemporal PhysSFI-Net (PhICNet) (Saha et al., 2020): Configured as a convolutional RNN, where the core update enforces a finite-difference approximation to the governing PDE, and a residual encoder-decoder network (RED-Net) learns unknown, time-varying unobserved sources. The entire system is end-to-end trainable, directly coupling physics-based parameter learning and data-driven estimation of hidden forcings.

All architectures implement physics integration structurally (by design), not merely as a soft regularization, ensuring physical feasibility and interpretability of the predictions.

2. Physics-Guided Mechanisms and Loss Formulation

Physics-informed constraints are instantiated distinctly in each PhysSFI-Net variant:

• Graph Laplacian Elasticity (Skeletal–Facial) (Bao et al., 5 Jan 2026):
  • Displacements on high-resolution meshes are solved via a Dirichlet graph-harmonic problem:

  $$L_G \delta = 0,\quad \delta|_S = \delta_{\mathrm{fix}}$$

  The minimization enforces local smoothness and biomechanical stiffness. Incremental deformations are modeled by LSTM recursion, with physical priors encoded in the smoothness and progression loss terms.

• FSI PINN Coupling (Tang et al., 30 Jun 2025):

  • Bulk fluid: incompressible Navier–Stokes equations

  $$\nabla\cdot \mathbf{u} = 0,\quad \rho(\partial_t \mathbf{u} + (\mathbf{u}\cdot\nabla)\mathbf{u}) = -\nabla p + \mu\nabla^2\mathbf{u}$$ - Structural surface: reduced-order modal expansion of deformation, time derivatives yield interface velocities. - Loss:

  $$\mathcal{L}_\mathrm{tot} = \lambda_\mathrm{phys}\mathcal{L}_\mathrm{phys} + \lambda_\mathrm{int}\mathcal{L}_\mathrm{int} + \lambda_\mathrm{trk}\mathcal{L}_\mathrm{trk}$$

  Each term quantifies residuals in PDE, interface, and measurement consistency.

• Convolutional PDE Core (Source Identification) (Saha et al., 2020):

  • State update follows discretized PDE finite-difference with learnable physical parameters $\theta$.
  • Residual source maps $V_t$ are estimated internally, and a residual network learns their unobservable time evolution.
  • Total loss combines forecast MSE, source prediction error, and an $L_1$ sparsity penalty.

These mechanisms guarantee physical plausibility, facilitate source/interaction inference, and stabilize long-horizon predictions.

3. Model Implementation and Workflow

Geometric PhysSFI-Net (Bao et al., 5 Jan 2026)

• Step 1: Downsample pre/post skeletal and facial point clouds, encode each with semantic labels and multi-scale geometric aggregations using graph convolutions.
• Step 2: Encode craniofacial correspondence and facial structure with separate PointNet++-style modules; fuse features by attention mechanism mapping bone plans to facial deformation context.
• Step 3: Sequential LSTM decodes multistep incremental displacements, each summing to final tissue motion.
• Step 4: High-resolution reconstruction enforces Dirichlet constraints (model-predicted displacements) in graph Laplacian smoothers to generate detailed postoperative facial surfaces.

FSI PhysSFI-Net (Tang et al., 30 Jun 2025)

• Fluid and solid networks are evaluated at physical coordinates, and particle-advection is enforced exactly by model construction.
• Losses are computed by automatic differentiation; training requires only particle track data and modal bases, not surface or full-field measurements.

Spatiotemporal PhysSFI-Net (PhICNet) (Saha et al., 2020)

• The RNN core advances state based on physics; residual blocks process recurrent prediction error to estimate source evolution.
• Training alternates between a warm-up (teacher-forced) regime and free recursive forecasting.

4. Benchmarking, Quantitative Results, and Comparative Evaluation

Geometric PhysSFI-Net (Bao et al., 5 Jan 2026):

Metric	PhysSFI-Net (mean ± SD)	ACMT-Net
Point-cloud Hausdorff (mm)	1.070 ± 0.088	1.186 ± 0.080
Surface deviation (mm)	1.296 ± 0.349	1.372 ± 0.351
Landmark error (mm, 19 pts)	2.445 ± 1.326	2.930 ± 1.555

• PhysSFI-Net achieves statistically significant improvements (p < 0.05 or < 0.001) over ACMT-Net across all measures.
• Proportion of predictions within clinically acceptable error (<2 mm) increases from ~40% (ACMT-Net) to ~60% (PhysSFI-Net).

FSI PhysSFI-Net (Tang et al., 30 Jun 2025):

• For 2D flapping beam:
  • Vorticity NRMSE: ~15%
  • Pressure NRMSE: ~8%
  • Surface displacement error: ~20% overall; ~2–3% on leading two modes.
• For 3D pipe flow:
  • Axial velocity NRMSE: ~5%
  • Transverse velocity NRMSE: ~14%
  • Surface radial deflection: ~4.6–8.1%
• Robustness to modal truncation: inclusion of unused or “extra” modes does not degrade performance, indicating no need for regularization via truncation.

Spatiotemporal PhysSFI-Net (Saha et al., 2020):

• PhICNet outperforms PDE-RNN+CNN and ConvLSTM for forecasting SNR by 1–2 dB across 150 steps, with remarkable retention of source-map correlation (>0.9), compared to physical-agnostic or naive methods (drop <0.5 after 50 steps).

5. Domain Applications and Interpretability

• Surgical Outcome Prediction: PhysSFI-Net generates millimeter-precise, interpretable reconstructions for patient-specific orthognathic interventions, supporting both planning and patient communication with rapid inference (~seconds) (Bao et al., 5 Jan 2026).
• FSI Inference in Scarce-Data Regimes: Demonstrated recovery of full spatiotemporal flow and structural states from only a few fluid-phase loci, critical for experimental biomechanics or flows where wall measurements are unavailable or unreliable (Tang et al., 30 Jun 2025).
• Source Identification in Physics-Driven Dynamics: Enables automated amodal source recovery alongside accurate future-state prediction, supporting control, monitoring, and anomaly detection in complex fields (Saha et al., 2020).

Across these applications, the explicit mapping to physical variables (e.g., displacement, velocity, pressure) yields interpretable latent spaces, and the analytic form of priors (graph Laplacians, modal dynamics, finite-difference PDEs) grounds the learned models in mechanistic insight.

6. Limitations and Future Research Directions

• Geometric PhysSFI-Net (Bao et al., 5 Jan 2026): Limited by training data diversity; future extensions should include multi-center datasets, texture prediction, and subject-specific elastic parameters.
• FSI PINN PhysSFI-Net (Tang et al., 30 Jun 2025): Assumes moderate, mode-resolvable structural deformation; future targets include real experimental data, non-smooth/buckling surfaces, and nonlinear constitutive solids.
• Convolutional RNN PhysSFI-Net (Saha et al., 2020): Scalability to very high resolutions and strongly nonlinear (e.g., turbulent) regimes remains a challenge.

Common to all: further integration of more domain physics (e.g., patient-specific tissue mechanics, anisotropic properties, non-Newtonian effects), real-time inference for closed-loop systems, and expansion to broader biomedical, fluid, or structural applications.

7. Relation to Other Physics-Informed Neural Approaches

PhysSFI-Net distinguishes itself by embedding physical constraints at the architectural level (hard graph/finite-difference constraints, modal decomposition, exact interface coupling), rather than as mere penalizations in the loss. In contrast to general-purpose PINNs—which typically employ MLPs to approximate solution fields with PDE soft constraints—PhysSFI-Net variants often leverage task-specific representation (e.g., graph transformers, convolutional RNNs, coupled coordinate networks), and their success in clinical and experimental settings suggests broader viability for neural-physics integration across scientific domains.

---

References:

• "PhysSFI-Net: Physics-informed Geometric Learning of Skeletal and Facial Interactions for Orthognathic Surgical Outcome Prediction" (Bao et al., 5 Jan 2026)
• "Neural inference of fluid-structure interactions from sparse off-body measurements" (Tang et al., 30 Jun 2025)
• "Physics-Incorporated Convolutional Recurrent Neural Networks for Source Identification and Forecasting of Dynamical Systems" (Saha et al., 2020)