陈国银
The Sun's magnetic field is written in polarised light. I build the diagnostics that read it: inversion codes, neural solvers, field extrapolations. Then I test what we think the solar atmosphere is doing.
Five threads, one question asked at five heights: what is the field doing in the solar atmosphere, and how would we know?
A Stokes profile is a fingerprint with the field, the temperature and the velocity tangled together. I write inversion codes that untangle them: Milne–Eddington solved in batch on a GPU, node-based nonlinear least squares, and now physics-informed networks that return a whole atmosphere in a single pass.
The inverse problem is expensive because it is iterative. A network trained on radiative-transfer solutions does not replace the physics; it makes the physics affordable, returning a full field of view without iterating per pixel.
The photosphere supplies a boundary condition and the corona supplies the consequences. Extrapolation lifts a measured surface field into the volume above it, force-free in the classical formulation and now learned with a physics-reinforced GAN.
A simulation is only useful if it can be observed. I run MHD models of the global corona constrained by measured fields, and synthesise what an instrument would see, so that model and observation can be compared in the same units.
Loops oscillate, and the period carries density, field strength and geometry inside it. Spatial seismology turns that oscillation back into the parameters responsible for it: a measurement made of time instead of light.
PI2NN. A physics-informed network trained cyclically: a Milne–Eddington forward model turns the network’s own output back into spectra, so the loss is measured in Stokes space rather than in parameter space. It returns a starting atmosphere; the classical inversion refines it.
SPOT. An operator transformer: Stokes IQUV in, two Fourier neural-operator blocks, four transformer blocks, a depth decoder, and eight parameter heads out. One forward pass replaces the per-pixel iteration.
Chen et al. 2026, A&A 710, A227
Five of 11, chosen for what they show about the method. The full record, 2021–2026 and 4 as first author, carries abstracts and BibTeX.
Plate IVCoronal loops
Schematic, not an observation: a family of loops standing on the limb, drawn from the geometry an extrapolated field produces. The same shape is what PF2nlfff has to recover from a photospheric boundary.
The codes behind the papers, and the small tools that keep the literature readable while the papers get written.
PyTorch library for Stokes IQUV spectral synthesis and inversion, with a pretrained neural-operator solver.
GPU-batched Milne–Eddington inversion of solar Stokes profiles with gradient-bias annealing and a physics-informed network.
Physics-reinforced GAN that extrapolates a photospheric boundary into a 3D nonlinear force-free magnetic field.
Agent skill that watches chosen arXiv categories and files relevant new papers into a three-level Zotero tree.
Agent skill that turns Zotero PDFs into illustrated Markdown reading notes mirrored into an Obsidian vault.
Overleaf workbench embedded as a fourth session tab in DeepSeek Harness Web, with selection quoting and caret insertion.
Everything else, agent skills and editor plugins, lives on GitHub.
I work between the observation and the model. The instrument measures polarisation; the atmosphere answers with a magnetic field, a temperature and a line-of-sight velocity. Everything I do is an attempt to get from the first to the second without losing the physics on the way.
I am a direct-entry PhD student in astronomy at Nanjing University, where the coronal-seismology work began, and a guest researcher at the Rosseland Centre for Solar Physics in Oslo, where I work on inversion, extrapolation and radiation synthesis for data-constrained MHD models of the corona.
The codes are public, SPOT and CuSP for inversion and PF2nlfff for extrapolation, because a diagnostic is only as reproducible as the code that produces it and because the fastest way to be corrected is to publish the tool.