Contrast-Source Physics-Driven NN (CSPDNN)

• Contrast-Source-Based Physics-Driven Neural Networks (CSPDNN) are hybrid models that integrate physical contrast-source formulations with deep neural network architectures to solve inverse problems.
• They utilize composite loss functions—state-consistency, data fidelity, and total variation—to embed physical laws such as Maxwell’s equations into the training process.
• CSPDNNs demonstrate enhanced reconstruction accuracy, robust noise handling, and computational efficiency across applications in electromagnetic inverse scattering and heat diffusion.

A Contrast-Source-Based Physics-Driven Neural Network (CSPDNN) is a class of hybrid neural architectures for inverse problems, combining the physical structure of contrast-source formulations with deep neural network parameterization and physics-informed loss functions. CSPDNN is distinguished by using the induced current (contrast source) as the main variable, tightly embedding domain physics (e.g., Maxwell’s equations or diffusion PDEs) into the network workflow. This approach enables highly efficient, robust, and accurate solution of nonlinear inverse scattering and PDE-constrained inference problems across electromagnetics, heat diffusion, and related domains (Sun et al., 14 Aug 2025, Du et al., 27 Jan 2026, Blakseth et al., 2022).

1. Physical Foundations and Contrast-Source Formulation

CSPDNNs are unified by the use of the "contrast source" (Editor’s term): a current-like auxiliary variable $J=\chi E^{\text{tot}}$ (in electromagnetics) or an analogous source in diffusion/PDE settings. The method pivots on a system of coupled equations derived from the governing physics:

• Electromagnetic Example: For two-dimensional, transverse-magnetic (TM) wave scattering, the time-harmonic Maxwell's equations reduce to the Lippmann–Schwinger equation,

$$E^{\rm tot}(r) = E^{\rm inc}(r) + k_0^2\int_{\mathrm{DOI}} G(r, r') \chi(r') E^{\rm tot}(r') dr',$$

with contrast $\chi(r) = \varepsilon_r(r)-1$. The induced contrast-source current is defined as $J(r) = \chi(r) E^{\mathrm{tot}}(r)$.

The coupled system for $J$ is: - State (domain) equation:

$$E^{\rm tot}(r) = E^{\rm inc}(r) + k_0^2\int_{\mathrm{DOI}} G(r, r') J(r') dr'$$

• Data (measurement) equation:

  $$E^{\rm sca}(r) = k_0^2\int_{\mathrm{DOI}} G(r, r') J(r') dr', \;\; r \in S$$

  • Diffusion Example: In transient heat conduction, the physics-driven backbone is

$$\frac{\partial T}{\partial t} = \kappa \Delta T + \sigma(x, y, t)$$

where $\sigma$ is a corrective source term, learned via a DNN to account for divergence from the base model (Blakseth et al., 2022).

In these settings, the contrast source variable forms the “interface” between PDE physics and neural network modeling.

2. Neural Network Architectures

CSPDNN frameworks instantiate the mapping from physical fields and/or spatial coordinates to induced sources using various deep network architectures:

• DeepCSI Implementation: Utilizes a residual multilayer perceptron (ResMLP) with stacked residual blocks. For each frequency, an independent ResMLP is defined. The input vector comprises Fourier-style positional encoding of $(x, y)$, plus transmitter indices. Each residual block includes a 256-wide linear layer, batch normalization, GELU activation, and skip connection. The output is two channels (real/imaginary parts) for $J$ at each node (Sun et al., 14 Aug 2025).
• CSPDNN (2026) Implementation: Adopts a convolutional neural network (CNN) backbone of three Conv layers (16, 32, 64 filters, each $3\times 3$) with residual connections and LeakyReLU, followed by global flattening and two fully connected layers outputting $[\Re\{\mathbf J\}, \Im\{\mathbf J\}]$ for the full domain grid. Initial input combines $\Re, \Im$ parts of the current and permittivity (Du et al., 27 Jan 2026).
• Diffusion/Corrective Source Example: Employs a 4-layer fully-connected feedforward network (hidden width 80, LeakyReLU) that predicts the spatiotemporally varying corrective source term, fed with the predicted temperature field (Blakseth et al., 2022).

A common theme is the use of relatively shallow architectures—enabling rapid convergence while maintaining full differentiability and compatibility with gradient-based optimization.

3. Physics-Informed Training and Composite Loss Functions

CSPDNNs enforce fidelity to physics through a composite loss function, integrating the following components:

• State-Consistency Loss: Measures the residual in the discretized state equation, enforcing that the predicted source and field reconstruct the interior physical model,

$$L^{\rm State} = \frac{\|\mathbf J_{\boldsymbol\theta} - \boldsymbol\chi_{\boldsymbol\theta} \odot \mathbf E^{\rm tot}_{\boldsymbol\theta}\|_2^2}{\|\mathbf E^{\rm inc}\|_2^2}$$

(Du et al., 27 Jan 2026, Sun et al., 14 Aug 2025)

• Data-Fidelity Loss: Penalizes deviation of predicted scattering (measurement) data from observed values,

$$L^{\rm Data} = \frac{ \| \mathbf G_S \mathbf J_{\boldsymbol\theta} - \mathbf E^{\text{sca}}_{\text{meas}} \|_2^2 }{ \| \mathbf E^{\text{sca}}_{\text{meas}} \|_2^2 }$$

• Lower Bound Constraint: Ensures $\Re\{\varepsilon_r\}\geq 1$ for physical admissibility,

$$L^{\rm Bound} = \alpha \|\max\{1-\Re(\varepsilon_{r,\boldsymbol\theta}), 0\}\|_1$$

with $\alpha = 10^{-4}$.

• Total Variation (TV) Loss: Enforces spatial smoothness and edge-preservation in the reconstructed contrast, with an adaptive weighting,

$$L^{\rm TV} = \sum_i \alpha_i \sqrt{ (\nabla_x v_i)^2 + (\nabla_y v_i)^2 }, \qquad \alpha_i = \frac{\beta_0}{\text{mean}(v)}$$

(Du et al., 27 Jan 2026).

• Purely Data-Driven MSE Loss: In the heat diffusion PDE setting, the network is trained exclusively on the mean squared error between the predicted and true corrective sources (Blakseth et al., 2022).

These losses are combined, possibly with scenario-dependent terms (e.g., magnitude-based losses for phaseless data), into an overall objective.

4. Inverse Scattering and Hybrid Model Applications

CSPDNN has been primarily developed and tested in the context of inverse electromagnetic scattering:

• Electromagnetic Inverse Scattering: Directly reconstructs spatially varying permittivity ($\varepsilon_r$) by optimizing both ResMLP weights and a spatial tensor of $\chi$. The method works under diverse measurement scenarios: full-data (complex measurements), phaseless (amplitude-only), and multi-frequency. For each frequency, an independent subnet is trained, and losses are summed across frequencies (Sun et al., 14 Aug 2025, Du et al., 27 Jan 2026).
• Heat Diffusion: The hybrid corrective source-term architecture extends the CSPDNN paradigm to parabolic PDEs. Starting from a first-principles discretization, the DNN corrects for model residuals such as parameter errors or partial physics, and is shown to outperform both pure physics-based and black-box data-driven models (Blakseth et al., 2022).

This breadth highlights the universality of the contrast-source framework as a bridge between physics-based modeling and machine learning.

5. Experimental Performance and Benchmarking

CSPDNNs demonstrate marked gains in reconstruction accuracy, robustness, and computational efficiency:

Scenario	CSPDNN/DeepCSI RMSE	SSIM	Baseline (CSI/MRCSI) RMSE	Inference Time (s)
Synthetic (3GHz)	0.03–0.06	0.86–0.96	0.06–0.08	~27–29 (CSPDNN)
Noisy Data (SNR=1dB)	Preserves boundaries	≈0.5	Rapidly degraded SSIM	~72–138 (baseline)
Multi-frequency	0.03–0.05	>0.92	Higher
Phaseless	0.03–0.05	0.85–0.96	Higher
Experimental (Fresnel/“FoamDielExt”)	0.11–0.17	0.78–0.87	0.18, 0.60–0.75

DeepCSI and CSPDNN reliably exceed classical and unsupervised operator-learning baselines, retaining high-fidelity reconstructions under severe noise, diverse data loss settings, and experimental (not merely synthetic) measurement (Sun et al., 14 Aug 2025, Du et al., 27 Jan 2026). In diffusion, the hybrid corrective source network achieves relative $\ell_2$-errors 1–2 orders of magnitude lower than pure PBM or DDM, with comparable generalization (Blakseth et al., 2022).

6. Advantages, Limitations, and Generalization

CSPDNN offers notable advantages:

• Pipeline Simplification: Physics is incorporated as differentiable loss terms; hand-coded gradients are unnecessary.
• Universality: Supports a range of measurement types (full, phaseless, multi-frequency, broadband, with easy toggling by loss definition) (Sun et al., 14 Aug 2025).
• Speed and Differentiability: Shallow network architectures enable fast convergence (typically under 30 s on modern GPUs), fully integrated with standard optimizers (Du et al., 27 Jan 2026).
• Accuracy and Robustness: Outperforms classical and operator-based methods across noise levels and data incompleteness.

Known limitations and potential directions include:

• Spectral Bias of MLPs: High-frequency features in $J$ require deeper architectures, leading to computational trade-offs for extended domains ($>10\lambda$) (Sun et al., 14 Aug 2025).
• Optimization Landscape: The joint search space over network weights and medium parameters ($J$ and $\chi$) is nonconvex. Proper initialization and tuning remain critical.
• Higher Dimensions and Heterogeneity: 3D or strongly heterogeneous settings may necessitate architectures such as wavelet neural operators, or new preconditioning strategies.
• Hybridization and Extrapolability: The hybrid model (e.g., CoSTA) in heat diffusion exhibits superior generalization, inheriting extrapolation robustness from the physics-based backbone (Blakseth et al., 2022).

7. Related Paradigms and Context

CSPDNN epitomizes the “physics-driven neural network” approach. Unlike purely data-driven frameworks, CSPDNNs typically employ untrained or self-supervised settings: weights are adapted directly from measurement data and physical priors, not from large synthetic datasets. Key related classes include:

• Contrast Source Inversion (CSI): Classical iterative methods framed around $J$ and $\chi$ updates, but lacking neural parameterization or gradient-based end-to-end learning.
• Operator Learning/Untrained Neural Networks (UNNs): Approaches such as physics-driven neural operator regression or self-organizing maps, which offer universality but often incur higher inference times and less flexibility in loss specification (Du et al., 27 Jan 2026).
• Hybrid PDE/DNN Models: The corrective source-term paradigm (Blakseth et al., 2022) exemplifies how a DNN component may compensate explicitly for deficiencies in a physics-based core, leading to compounded improvements in both accuracy and generalizability.

In conclusion, CSPDNNs provide a rigorous, computationally efficient, and highly accurate hybrid framework for solving challenging inverse problems by parameterizing physically-guided source terms with neural networks and training under composite, physics-respecting loss criteria (Sun et al., 14 Aug 2025, Du et al., 27 Jan 2026, Blakseth et al., 2022).