lfspin: light-front spin dynamics, exact benchmarks and physics-informed neural network solver

Code and data for the Letter N. S. Akintsov, A. P. Nevecheria, S. N. Andreev, Qing-Hua Qin, Rapidity-Coupled Spin Dynamics in Pulsed Laser Fields from Physics-Informed Neural Networks, submitted to Physical Review A. An electron crosses a plane-wave laser pulse with a cos² envelope. Its orbit is the Volkov solution, and its spin obeys the covariant Bargmann–Michel–Telegdi (BMT) equation. For linear polarization the rest-frame polarization angle is an explicit function of the instantaneous vector potential, Σ = 2 arctan(ax/2) + aeax. The net spin rotation after the pulse therefore vanishes for every carrier-envelope phase (CEP), while the peak intra-pulse angle depends on the CEP. The package uses this exact result to benchmark numerical spin solvers and a physics-informed neural network (PINN) trained only on the equation residuals and invariants, and then applies the PINN to elliptical polarization, where no closed form exists. Main results reproduced Closed form against a light-front DOP853 reference: agreement to 10−12; residual net rotation over 16 CEPs below 9×10−13 degrees. CEP spread of the peak polarization angle: 3.02° for a two-cycle and 0.22° for an eight-cycle pulse (49-point scans). Fixed-step-budget convergence of lab-time RK4, Boris–BMT, Higuera–Cary–BMT and light-front RK4 for a co-propagating electron with γ0 = 10. PINN against the reference: maximum error 4.7×10−4 in γ and 3.9×10−5 in the spin components; an ablation of the Fourier-feature embedding, a smaller baseline network, and an elliptically polarized pulse. Contents physics.py: pulse fields, Volkov orbit, BMT right-hand sides, DOP853 reference integrators, closed forms. convergence.py: fixed-step lab-time RK4, Boris–BMT and Higuera–Cary–BMT pushers, light-front RK4. pinn.py, train_pinn.py: the Fourier-feature PINN and its four training configurations. run_exact_checks.py, run_cep_scan.py, run_convergence.py: exact checks, CEP scan, convergence study. make_fig1.py, make_fig2.py, make_figS1.py: Figs. 1, 2 and S1. reproduce_all.py: single entry point; also checks the files against MANIFEST.sha256. results/: computed data (*.json), run transcripts (*.log) and the four trained networks (*_model.pt). Every number in the Letter and its Supplemental Material is read from these files. figs/: the figures as PDF and PNG. Reproduction python -m pip install -r requirements.txt python reproduce_all.py --verify # check the files against MANIFEST.sha256 python reproduce_all.py --skip-training # everything except PINN training python reproduce_all.py # full run, including the four trainings Python 3.14.3, NumPy 2.4.2, SciPy 1.18.1, Matplotlib 3.11.1, PyTorch 2.10.0 (CPU build), Windows 11; no GPU is needed. The ODE computations are deterministic. PINN training is seeded and resumable; the four trainings took 17–87 min each, about 4.4 h in total, on 4–8 CPU threads. A rerun on other hardware can change the PINN errors in their last digits. Scope. Classical BMT spin dynamics of a single electron in an ideal plane-wave pulse: no radiation reaction, no focusing, no ensemble averaging. Each network is trained for one parameter point. License: MIT.