Physics Transfer Learning Framework

• Physics-transfer learning is a method that reuses pre-trained physical models from data-rich sources to adapt to target systems with limited samples and altered conditions.
• It employs techniques such as full fine-tuning and parameter-efficient methods (BitFit, LoRA) to update only critical network components while preserving inherent physics.
• Evaluations on electromagnetic shower simulations and PDE transfer demonstrate substantial improvements—up to a 44% gain in fidelity—with minimal target-domain data.

Searching arXiv for the target paper and closely related transfer-learning frameworks in physics. A physics-transfer learning framework is a class of methods that reuses learned structure from a source physical system and adapts it to a target system with limited data, altered geometry, changed boundary conditions, or different operating regimes. In fast electromagnetic shower simulation, this idea is instantiated as cross-geometry transfer learning: generative calorimeter models are pre-trained on the International Large Detector and then adapted to new detector configurations using point cloud shower representations, avoiding geometry-specific re-voxelization and improving data efficiency. On the CaloChallenge dataset, transfer learning with only 100 target-domain samples yields a $44\%$ improvement on the geometric mean of Wasserstein distance over training from scratch, and parameter-efficient fine-tuning with bias-only adaptation remains competitive while updating only $17\%$ of model parameters (Gaede et al., 28 Nov 2025). Related work applies the same general principle to PDE operators, materials properties, acoustics, morphogenesis, and physics-informed neural networks, but the transferred object and the adaptation mechanism vary substantially across domains.

1. Definition and problem setting

In the recent literature, physics-transfer learning addresses a recurring scientific machine learning problem: the source task contains either abundant simulation data, an approximate governing equation, or a well-characterized domain, whereas the target task is data-scarce, shifted in geometry or operating conditions, or otherwise expensive to solve directly. This pattern appears in calorimeter simulation, where Geant4-quality simulation is computationally prohibitive and detector changes would otherwise require retraining geometry-specific surrogates; in differential equations, where source and target domains differ through distribution bias and operator bias; and in materials prediction, where high-quality target data are scarce relative to auxiliary or related physical data (Gaede et al., 28 Nov 2025, Yang et al., 2 May 2025, Wu et al., 4 Apr 2026).

These studies collectively indicate that physics-transfer learning is not defined by a single architecture. Instead, it is characterized by a transfer pathway that is constrained by physical structure. In some cases, the transferable content is an internal representation of hidden PDE parameters; in others, it is a geometry-agnostic point cloud generator, a harmonic branch learned from elastic data, a latent prior over hidden fluid properties, or a neural surrogate whose frozen layers encode approximate physics (Zhang et al., 2023, Zhu et al., 2024, Chakraborty, 2020). This suggests that the central question is not only how to fine-tune a model, but also where the task-specific physics enters the model and which subspace should remain shared across tasks.

2. Transferred objects and physics representations

A defining feature of these frameworks is the choice of representation in which transfer is attempted. In cross-geometry calorimeter simulation, showers are represented as 3D point clouds $(x, y, z, e)$, decoupled from a specific voxelization or readout geometry. This makes the generative model geometry independent at the representation level and allows direct projection onto arbitrary detector configurations, including planar and cylindrical geometries (Gaede et al., 28 Nov 2025). The reported architecture combines a diffusion-based PointWise Net with ShowerFlow, a normalizing flow that predicts the number of points per calorimeter layer $N_{z,i}$ and conditions the point cloud generator (Gaede et al., 28 Nov 2025).

Other frameworks embed physical structure more explicitly in the network design. KappaFormer separates lattice thermal conductivity into harmonic and anharmonic contributions, pre-training a harmonic branch on elastic-property data and fine-tuning an anharmonic branch on limited experimental $\kappa_\mathrm{L}$ data; the decomposition is motivated by the Slack-model-based separation of sound-velocity-related harmonicity and Grüneisen-parameter-related anharmonicity (Wu et al., 4 Apr 2026). MetaNO argues that hidden task-specific parameter fields in parametric PDEs are best encoded in the first, lifting layer of a neural operator rather than in the final layer, with the iterative kernel integration and projection layers shared across tasks (Zhang et al., 2023). In latent intuitive physics, the transferable quantity is neither an explicit scalar parameter nor a final embedding, but a probabilistic latent feature sequence generated by a physical prior learner and distilled from visual posterior inference through a neural renderer and particle simulator (Zhu et al., 2024).

A related but distinct strategy appears in brain morphogenesis and high-strength alloy discovery. In the former, nonlinear elasticity learned from simple geometries such as spheres and ellipsoids is embedded in graph neural network weights and transferred to anatomically complex brain geometries using a digital library of continuum-mechanics simulations (Zhao et al., 22 Aug 2025). In the latter, a neural network learns the mapping from elastic constants and $\gamma$-surface information to Peierls stress from abundant low-fidelity force-field simulations, then applies the learned mapping to high-fidelity DFT-derived descriptors without retraining (Zhao et al., 2024). These cases indicate that transfer can target either the input representation, the latent mechanism, or the learned constitutive map.

3. Adaptation mechanisms

The adaptation stage varies from conventional fine-tuning to strongly structured transfer rules. In calorimeter simulation, the basic mechanism is unsupervised domain adaptation from a single-detector pre-training phase to a new detector geometry and particle domain, without paired data or explicit geometric correspondences. Full fine-tuning updates all parameters, while parameter-efficient fine-tuning includes BitFit, where only bias terms are trainable, Top2, where only the last two and time-embedding layers are updated, and LoRA, which adds low-rank adaptation layers. For ShowerFlow, a layer mismatch between source and target calorimeters is handled by modelling new target layers with log-normal priors and updating them during fine-tuning while retaining transferred knowledge in the existing layers (Gaede et al., 28 Nov 2025).

Comparable selectivity appears in several other frameworks. KappaFormer freezes the harmonic branch after pre-training and trains only the anharmonic branch and downstream layers on scarce $\kappa_\mathrm{L}$ data (Wu et al., 4 Apr 2026). The near-field acoustic holography framework pre-trains a complex-valued convolutional neural network on labeled rectangular-plate data and then performs purely physics-informed fine-tuning on a single target sample using a Kirchhoff-Helmholtz-integral-based pressure loss, without source-velocity labels in the target domain (Luan et al., 15 Jul 2025). MF-PIDNN first trains a low-fidelity physics-informed network from an approximate governing equation and then updates only the last one or two layers using sparse high-fidelity data, keeping earlier layers frozen so that they preserve features learned from physics (Chakraborty, 2020). One-shot transfer learning of PINNs pushes this logic further for linear ODEs and PDEs: after bundle pre-training, the hidden layers are frozen and only the final linear output layer is solved analytically for a new equation in the same family (Desai et al., 2021).

The diversity of adaptation strategies can be summarized as follows.

Framework	Transfer locus	Adaptation mode
Cross-geometry calorimeter simulation (Gaede et al., 28 Nov 2025)	Point cloud generator and ShowerFlow	Full fine-tuning, BitFit, Top2, LoRA
MetaNO (Zhang et al., 2023)	First lifting layer of neural operator	Per-task adaptation of $\mathcal{P}$ with shared iterative/projection layers
KappaFormer (Wu et al., 4 Apr 2026)	Harmonic/anharmonic branches	Freeze harmonic branch; fine-tune anharmonic branch
Physics-preserved OT for DEs (Yang et al., 2 May 2025)	Source-to-target product distribution	Learn a physics-regularized transport map and fine-tune on pushed-forward data
NAH transfer learning (Luan et al., 15 Jul 2025)	Pre-trained CV-CNN weights	Single-sample physics-informed fine-tuning via KH integral
One-shot PINN transfer (Desai et al., 2021)	Frozen hidden representation	Closed-form solution for output weights

A common misconception is that transfer in physics-based models is equivalent to updating only the final layer. MetaNO explicitly argues against this for parametric PDE families, because the hidden parameter field acts as a modifier of the input representation and is therefore more naturally captured in the first layer (Zhang et al., 2023). The calorimeter study likewise reports that low-rank LoRA underperformed even at rank 106, with a layerwise SVD analysis indicating that the required physics weight updates have high intrinsic dimensionality (Gaede et al., 28 Nov 2025).

4. Physics constraints, losses, and evaluation

Physics-transfer learning frameworks generally combine transfer with physically structured objectives. In the calorimeter setting, evaluation emphasizes distributional fidelity rather than pointwise reconstruction. The study reports both Kullback-Leibler divergence and Wasserstein-1 distance,

$$KL(P \,\|\, Q) = \sum_i P_i \log\left(\frac{P_i}{Q_i}\right),$$

$$W_1(P, Q) = \min_{\pi \in \Pi(P,Q)} \sum_{i, j} |x_i - x_j| \pi(x_i, x_j),$$

together with a geometric mean across six observables,

$17\%$0

namely voxel energy spectrum, energy ratio, visible energy, occupancy, longitudinal profile, and radial profile (Gaede et al., 28 Nov 2025).

In differential-equation transfer, POTT formalizes source and target data domains as product distributions over inputs and solutions and learns a transport map whose push-forward matches the target while preserving the governing operator through a regularization term,

$17\%$1

with physics preservation expressed through

$17\%$2

This formulation is designed to address both distribution bias and operator bias under domain shift (Yang et al., 2 May 2025).

Other domains impose physics through residuals, constitutive structure, or frequency constraints. In near-field acoustic holography, fine-tuning minimizes a pressure discrepancy after propagating the predicted source velocity through the Kirchhoff-Helmholtz integral, thereby enforcing consistency with the forward wave physics (Luan et al., 15 Jul 2025). In bearing-fault diagnosis, the loss penalizes predictions that conflict with expected Ball Pass Frequency Outer and Ball Pass Frequency Inner signatures derived from bearing geometry and shaft speed, with BPFO and BPFI computed from

$17\%$3

and added to cross-entropy as a physics-informed penalty (Alam et al., 11 Aug 2025). These cases show that “physics-informed” transfer can refer either to the source of the transferable representation or to the form of the target-domain adaptation objective.

5. Reported gains, robustness, and failure modes

The calorimeter study provides a clear low-data benchmark. With 100 target-domain samples, transfer learning reaches a geometric mean Wasserstein distance of $17\%$4 versus $17\%$5 when training from scratch, corresponding to the reported $17\%$6 improvement. The gain is largest in low-data regimes of at most $17\%$7 samples and is particularly visible for geometric observables such as longitudinal and radial shower profiles. BitFit remains within $17\%$8 of full fine-tuning while updating only $17\%$9 of parameters, and Top2 shows comparable results (Gaede et al., 28 Nov 2025).

Several adjacent literatures report analogous behavior. MetaNO achieves less than $(x, y, z, e)$0 prediction error with only 4 samples on new PDE tasks, whereas single or baseline transfer neural operators need more than 100 samples, and it reports up to $(x, y, z, e)$1 reduction in required measurements for comparable relative error (Zhang et al., 2023). In one-shot transfer learning for PINNs, the final-layer solve yields mean test MSE values of $(x, y, z, e)$2 for first-order ODEs and $(x, y, z, e)$3 for second-order ODEs, while avoiding full retraining for each new equation (Desai et al., 2021). In methane sorption modeling, hydrogen pre-training produces $(x, y, z, e)$4 lower RMSE and $(x, y, z, e)$5 faster convergence than random initialization, achieving $(x, y, z, e)$6 on held-out coal samples (Nooraiepour et al., 15 Apr 2026).

The same literature also identifies important limits. LoRA can fail when the required adaptation is not effectively low rank, as observed in the calorimeter study (Gaede et al., 28 Nov 2025). Domain adaptation methods for differential equations can align features while destroying or blurring physical structure, which is precisely the failure mode POTT is designed to avoid (Yang et al., 2 May 2025). In methane sorption, deep ensembles degrade under shared physics constraints because the admissible solution manifold becomes too narrow, whereas Monte Carlo Dropout remains well calibrated at minimal overhead (Nooraiepour et al., 15 Apr 2026). These observations suggest that transfer gains depend not only on task similarity, but on whether the chosen adaptation family matches the dimensionality and rigidity of the underlying physics shift.

6. Interpretation, scalability, and broader significance

A notable consequence of physics-transfer learning is that it often yields interpretable adaptation pathways rather than opaque parameter reuse. In cross-geometry calorimeter simulation, the reported analysis is presented as insight into adaptation mechanisms for particle shower development and as a baseline for future progress of point cloud approaches in calorimeter simulation (Gaede et al., 28 Nov 2025). In KappaFormer, interpretability follows from the explicit harmonic/anharmonic decomposition and from atom-wise and element-wise analyses linked to vibrational mechanisms (Wu et al., 4 Apr 2026). In brain morphogenesis, reduced-dimensional evolutionary representations such as vertex and ridge networks, together with latent-space uncertainty analysis, are used to characterize the essential physics of cortical folding and to support digital-twin deployment with medical images (Zhao et al., 22 Aug 2025).

Several papers also frame physics-transfer learning as an enabling technology for scalable surrogate modeling. The calorimeter work presents single-detector pre-training, combined with point clouds and parameter-efficient fine-tuning, as a pathway toward modular, generalizable, and scalable surrogate models for future collider experiments (Gaede et al., 28 Nov 2025). The multi-site emission-control framework reports that adaptation across 13 municipal solid waste incineration plants occurs through structured re-weighting of operating regimes rather than complete model re-learning, and extends the learned representation into an interpretable digital twin for regime-aware operational navigation (Ying et al., 29 Apr 2026). In alloy discovery, the transfer from low-fidelity multiscale simulations to DFT-derived descriptors is used to screen materials from a limited number of single-point calculations with chemical accuracy (Zhao et al., 2024).

This broader record clarifies two points. First, physics-transfer learning is not merely a data-efficiency technique; it is also a method for deciding which parts of a model should remain invariant under domain shift. Second, it does not eliminate the need to characterize the shift itself. The calorimeter paper explicitly notes that future work should test hadronic showers, larger detector shifts, and alternatives such as multi-detector pre-training or meta-learning (Gaede et al., 28 Nov 2025). The larger implication is that successful transfer depends on a correct decomposition of what is universal, what is geometry- or material-specific, and what must be re-estimated in the target domain.