Physics-Informed Learning for Cross-Domain Generalization

17 concepts17 papers1 proposal

When a standard deep neural network is trained to reconstruct medical images or process sensory data, it typically learns the statistical correlations of its training set. For example, a network trained on healthy brain MRIs learns what a typical brain looks like. However, if that network encounters an unseen pathology, a different species, or a novel material, it often fails catastrophically. The network has memorized semantic shapes rather than the underlying generative process.

A growing body of research addresses this limitation by shifting the learning target from semantic features to physical invariants. Because the laws of nature—such as wave propagation, light scattering, and magnetic resonance—remain constant across domains, networks that internalize these physical principles achieve remarkable out-of-distribution generalization. They do not need to see every clinical scenario or material type during training; they simply need to learn the physics of how the data was captured.

Shifting from Semantics to Physical Invariants

To understand why physics-informed features generalize, consider the task of Quantitative Susceptibility Mapping (QSM) in brain imaging. Traditional deep learning models overfit to healthy anatomy. However, Yoon et al. demonstrated that by training a network (QSMnet) to invert the physical dipole convolution loop, the model successfully generalizes to unseen pathological features like microbleeds and lesions. Because the network learns the physics of magnetic susceptibility rather than a template of a healthy brain, a lesion is processed correctly as just another local magnetic source.

This principle can be pushed to an extreme: a network does not even need real domain-specific data to learn how to reconstruct it. Li et al. trained a super-resolution fluorescence microscopy network entirely on abstract sketches and filtered natural images. Despite never seeing a biological cell during training, the network generalized perfectly to complex biological ultrastructures. It succeeded because the training process forced the network to master the invariant point spread function (PSF) of the optical system, which remains identical whether imaging a sketch or a live cell.

[Abstract Sketches / Natural Images] 
               │
               ▼ (Convolved with Physical Point Spread Function)
     [Synthetic Training Data]
               │
               ▼ (Network learns to invert the PSF)
     [Trained Reconstruction Model]
               │
               ▼ (Zero-Shot Generalization)
[Complex Biological Ultrastructures]

Invariance Across Wave Mechanics

Wave propagation provides another powerful physical invariant. Whether an acoustic wave travels through a human heart, a synthetic gelatin phantom, or a piece of plastic, it obeys the exact same partial differential equations.

In high-frame-rate ultrasound imaging, researchers traditionally train separate models for different organs to avoid artifacts. However, Senouf et al. and Vedula et al. showed that networks trained on dynamic human cardiac data generalize seamlessly to static calibration targets and synthetic phantoms. By optimizing the end-to-end pipeline of transmit beam patterns and receive neural networks, as explored by Vedula et al., the model learns fundamental acoustic principles rather than specific anatomical geometries.

This acoustic consistency extends beyond medical imaging into robotics and audio processing:

  • Granular Flow: Clarke et al. trained a robot to estimate the mass of pouring materials by listening to them. A model trained on a few materials (like rice or pasta) generalized to entirely novel materials like soil because the underlying physics of granular impact produces universal acoustic signatures of flow.
  • Structural Vibrations: Lee developed a contact-localization pipeline (SonicBoom) that maps structural vibrations to collision points on a robot. The model generalized from aluminum links to plastic links because it captured the fundamental equations of vibration propagation rather than material-specific acoustic profiles.
  • Instrument Dynamics: In music transcription, Jeong et al. demonstrated that the relationship between a piano key's physical strike velocity and its spectral evolution follows universal physical laws. Consequently, a model trained on one piano generalizes across entirely different instruments and recording environments.

Optical and Electromagnetic Invariants

Light interaction with matter offers a similarly robust foundation for cross-domain generalization. In the operating room, hyperspectral imaging (HSI) is highly sensitive to changes in illumination. Baumann et al. and Baumann et al. addressed this by training 3D convolutional networks to predict local illumination directly from raw data. Because the networks were constrained to learn the physics of light interference and reflectance, they achieved strong cross-species generalization—performing accurately on human tissue even when trained primarily on simulated or non-human data.

Similarly, in light-field microscopy, Lü et al. incorporated the physical principles of light-field sampling, such as phase correlation between different views. This physical grounding allowed a model trained on a single tissue type to achieve high-resolution volumetric reconstructions of entirely different species and subcellular features without retraining.

When the hardware itself is co-designed with the physics of the model, the generalization becomes absolute. Shi et al. integrated a designable metasurface onto an X-ray scintillator, acting as a physical convolutional encoder. By training a digital decoder to learn the physical propagation model of this optical hardware rather than specific image features, they achieved universal generalization across completely unrelated object categories.

Enforcing Mathematical and Architectural Consistency

How do we practically force a neural network to prioritize physics over semantic memorization? The literature highlights three primary strategies: hybrid optimization, self-consistency priors, and physics-based data synthesis.

Hybrid Optimization Frameworks

Instead of treating the neural network as a black box, researchers embed classical physical operators directly into the network's optimization path. Zhang et al. developed the Efficient Learned Descent Algorithm (ELDA) for low-dose CT reconstruction. By interleaving learned data-driven layers with classical optimization steps that enforce the physics of the Radon transform, the framework guarantees convergence and exhibits high generalizability across diverse anatomical structures without requiring retraining.

Self-Consistency Priors

Generative models can be explicitly constrained by the physical laws of signal acquisition. In accelerated MRI, Cheng integrated the k-space interpolation rule—a universal physical property of magnetic resonance signals—into a diffusion model framework (SPIRiT-Diffusion). Because the generative process is bound by this physical self-consistency prior, a model trained exclusively on brain scans can successfully reconstruct knee images. The anatomy changes, but the physics of k-space remains identical.

                       ┌─────────────────────────┐
                       │  Generative Diffusion   │
                       │          Model          │
                       └────────────┬────────────┘
                                    │
                                    ▼  Produces candidate image
                       ┌─────────────────────────┐
                       │   k-Space Consistency   │◀─── Enforces universal
                       │       Constraint        │     interpolation physics
                       └────────────┬────────────┘
                                    │
                                    ▼  Generalizes across anatomy
                       ┌─────────────────────────┐
                       │ Brain-Trained Model     │
                       │ Reconstructs Knee Scan  │
                       └─────────────────────────┘

Pure Physics-Based Synthesis

When real-world training data is scarce or biased, networks can be trained entirely on synthetic data generated by precise physical simulators. Wang et al. broke the training data bottleneck in multi-shot diffusion MRI by training a network entirely on 100,000 synthetic samples derived purely from physical laws. Because the network internalized the mathematical rules of signal acquisition rather than real-world anatomical biases, it generalized flawlessly across different scanner hardware, clinical protocols, resolutions, and noise levels.

Conclusion

The success of physics-informed feature learning stems from a simple truth: the physical laws governing sensor data are far less complex than the semantic variations of the objects being sensed. By anchoring network architectures, loss functions, and data synthesis in physical invariants—such as wave equations, phase correlations, and k-space consistency—the deep learning model stops acting as a fragile pattern recognizer and begins acting as an inversion operator for the physical world.

Key Foundational Work

Learning beamforming in ultrasound imaging (2019) — contributed 1 concept elements to this direction.
Quantitative susceptibility mapping using deep neural network: QSMnet (2018) — contributed 1 concept elements to this direction.
High frame-rate cardiac ultrasound imaging with deep learning (2018) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_physics_informed_generalization_physics__16f46034,
  title = {physics-informed generalization (physics-informed feature learning)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physics-informed-generalization-physics-informed-feature-learning/rdr_16f4603496ca4b42a4e55223ca6f7270},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_physics_informed_generalization_physics__16f46034_v1,
  title = {physics-informed generalization (physics-informed feature learning)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physics-informed-generalization-physics-informed-feature-learning/rdr_16f4603496ca4b42a4e55223ca6f7270/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake