
Reconstructing Hamiltonian parameters from noisy scattering data is an ill-posed inverse problem in non-Hermitian quantum systems. We propose an S-matrix Constrained Equivariant Graph Neural Network (SCEGNN) that embeds two architectural hard constraints to reduce the solution ambiguity. An equivariant message-passing encoder incorporates lattice point-group symmetry and permutation-invariant aggregation (for the Aubry–André–Harper quasiperiodic chain), thereby suppressing symmetry-induced parameter variability. A deterministic S-matrix solver serves as a fixed decoder that enforces the scattering equations within the assumed model, eliminating the need for physics-based residual loss terms. Inference is performed via a two-stage hybrid optimisation: offline pre-training on synthetic data, followed by online single-sample fine-tuning in a low-dimensional latent space. On a 48-site Kagome supercell and a 50-site Aubry–André–Harper quasiperiodic chain, using a synthetic dataset of 120,000 samples (100,000 training, 10,000 validation, 10,000 test) with additive white Gaussian noise (0–10%), the mean absolute percentage error of reconstructed parameters remains below 5% for both systems under 10% noise, with per-parameter standard deviations below 7% across 20 independent repetitions. The transmission spectrum reconstruction RMSE is 0.026±0.003 (Kagome) and 0.029±0.003 (Aubry–André–Harper) at 10% noise. Single-sample CPU inference takes 23 s (offline simulation scenario), achieving a 5.6-fold speedup over least-squares fitting with five restarts. Ablation studies confirm that each architectural component contributes to the reported performance. For the AAH quasiperiodic chain, we consider two scenarios: a uniform-dissipation scenario with two independent parameters (V0 and γ), which serves as our primary benchmark. Unless otherwise stated, all numerical results reported in the main text for the AAH chain correspond to the uniform-dissipation scenario.All experiments are conducted on synthetic data; the method is validated as a proof-of-concept for the two benchmark systems under the wide-band approximation and does not claim general applicability to arbitrary non-Hermitian systems or experimental real-time control.
We introduce a reduced-order learning framework that couples a convolutional autoencoder to a structured latent neural ordinary differential equation. A learned scalar generates the latent dynamics through an antisymmetric operator and a positive-definite dissipative operator. The antisymmetric component permits motion along scalar level sets, while the dissipative component decreases the scalar. Adding a coercive quadratic term gives bounded continuous-time latent trajectories and approach to the scalar’s critical set under the stated assumptions. The scalar provides a structural constraint rather than an approximation of physical energy. The encoder, decoder, scalar, and operators are trained jointly on short snapshot windows without time-derivative supervision, physical-energy labels, or dissipation penalties. At inference, one initial state is encoded and evolved continuously in latent space, with decoding only at requested times. We demonstrate the framework on micromagnetic magnetization trajectories from NIST $\mu$ MAG Standard Problem 4 and compare it with dissipative-only dynamics and an unstructured latent neural differential equation. Repeated training runs show competitive observed-horizon accuracy, while selected longer rollouts illustrate improved retention of approximately stationary states relative to the unstructured baseline. These results motivate scalar-generated latent dynamics as an inductive bias for reduced-order prediction and other learning tasks where bounded evolution and equilibrium behavior are important.
Physics-informed neural networks (PINNs) are difficult to train for radio-frequency capacitively coupled plasma fluid-Poisson systems because missing initial conditions, steep sheath gradients, and strongly imbalanced residuals can lead to unstable or nonunique reconstructions. We propose a Hierarchical Adaptive Fourier-feature PINN (HA-FPINN) for such coupled plasma models. The method combines Fourier feature mapping to improve the representation of small-scale sheath structures, sparse physical anchors to provide interior constraints, and a two-level adaptive weighting scheme to balance the PDE, boundary, and anchor losses as well as individual equation residuals. The framework is tested on an 8-equation helium discharge model and a 23-equation electronegative $\mathrm{Ar}/\mathrm{O}_2$ model. In the helium case, HA-FPINN reconstructs the space-time evolution of the main physical variables with $1\%$ anchors and converges for all tested network architectures, with convergence times of $110$ – $246$ s. In the $\mathrm{Ar}/\mathrm{O}_2$ case, HA-FPINN achieves stable reconstruction with $5\%$ anchors, whereas the non-adaptive FPINN and standard PINN fail to converge under the same setting. The best performance in both systems is obtained at the tested Fourier frequency parameter $\sigma = 2.20$ . Noise-injection tests show that HA-FPINN maintains the lowest reconstruction error in the helium case and preserves the main space-time structures in the $\mathrm{Ar}/\mathrm{O}_2$ case at anchor noise levels of up to $15\%$ . These results show that HA-FPINN provides a robust PINN formulation for strongly coupled plasma fluid models with missing initial conditions.
Phase-sensitive optical time-domain reflectometry (Φ-OTDR) is widely used for distributed acoustic sensing, yet effective event recognition in multi-source interference environments remains a significant challenge. To address these issues, this paper proposes a lightweight and efficient separation framework based on the time-frequency interleaved gain extraction and reconstruction (TIGER) architecture. A uniform band-split strategy replaces the speech-oriented frequency allocation, while multi-scale processing and attention model the temporal and spectral structure of the mixtures. Experiments cover six event classes, two-source and three-source synthetic mixtures, and real simultaneous two-source recordings. Comparisons include ConvTasNet Base and Tiny, DPRNN, SuDoRM-RF, TF-GridNet, and TIGER with the adapted original band split. Uniform-band TIGER achieves per-source classification accuracies of 97.17 ± 0.16% and 94.74 ± 0.01% with 0.822 M and 0.779 M parameters in the two-source and three-source tasks, respectively. Relative to the original-band variant, it improves accuracy by 2.29 and 1.79 percentage points and SI-SDR by 0.509 and 0.430 dB. In the three-source task, TIGER uses 9.5% of the parameters of TF-GridNet, with an accuracy difference of 0.87 percentage points. On 21 real two-source windows, it achieves an 87.30 ± 2.75% both-events-correct rate despite measurable differences between real and synthetic input features. Repeated training and paired test-mixture analysis support the performance comparisons. On an RTX 4090, the three-source separator processes a 0.5 s input in 73.69 ms, giving a real-time factor of 0.147. The results establish uniform-band adaptation as an effective approach to compact separation-based multi-event perception.
We introduce a probabilistic physics-informed learning framework that couples quantumstate reconstruction with dataset-level identification of globally shared Hamiltonian parameters from bound-state spectra. The Woods–Saxon inverse problem provides a physically interpretable test case involving multiple nuclei, neutron and proton sectors, and Coulomb and spin–orbit contributions. WaveNet represents separated radial, polar, and azimuthal wavefunction components, while ParamNet combines normalized radial probes, observed energies, nuclear descriptors, and quantum numbers to infer a distribution over six global parameters. Training jointly enforces Rayleigh-energy consistency, separated Schr ̈odinger residuals, boundary and phase conditions, explicit normalization, selective orthogonality, spin–orbit splitting consistency, and Kullback–Leibler regularization. The physics-informed formulation supports learning from limited spectral observations without supervised wavefunction labels, using only 35 experimental energies from four nuclei. Spectral evaluation uses the physical-space distribution mean. The learned standard deviations describe modeldependent parameter variability. In synthetic closure tests, all six generating parameters are recovered with sub-percent relative errors. Reintroducing the inferred means into an independent finite-difference solver gives spectral mean absolute errors of 0.0133 MeV for the Seminole parameterization and 0.0071 MeV for Wahlborn. For experimental spectra, the inferred Seminole interaction gives an aggregate mean absolute error of 0.7812 MeV, compared with 0.7969 MeV for the reference interaction and 0.8492 MeV for a finite-difference least-squares fit using the same identification levels. Identification within the Wahlborn expression reduces the corresponding error from 1.0783 to 0.9256 MeV. The framework provides a reusable and robust scientific-machine-learning formulation for inverse quantum eigenvalue problems in which globally shared parameters and physically constrained states must be identified jointly.