Spectropolarimetric Inversion in Four Dimensions with Deep Learning (SpIN4D): Magnetohydrodynamic Modeling and Forward Synthesis Pipeline

NSF's Daniel K. Inouye Solar Telescope (DKIST) provides high-resolution, multiline spectropolarimetric observations that will revolutionize our understanding of the solar atmosphere. Given the massive data volume, novel inversion techniques are required to unlock its full potential. Our SpIN4D project aims to develop deep convolutional neural networks for estimating the physical properties of the solar photosphere based on the time series spectropolarimetric observations from DKIST. We perform a set of small-scale dynamo, quiet-Sun simulations using the MURaM radiative MHD code as our ground truth. The simulation domain covers 25×25×8 Mm with 16×16×12 km spatial resolution with outputs generated at a 40 s cadence and useful physical domain extending up to the temperature minimum. Five cases with different mean vertical magnetic field strengths have been conducted; each case covers six solar-hours, totaling 64 TB in data volume. We forward model the Stokes profile of two Fe I lines (630.2 and 1565 nm) based on the MURaM runs using the Stokes Inversion based on Response functions (SIR) synthesis pipeline. These two lines will be simultaneously observed by the DL-NIRSP instrument at DKIST, and will be used together as the input to our model to better constrain the parameter variations along the line of sight. Our initial training set selects snapshots at a 12-minute cadence to reduce the correlation between individual granules. This allows inference of a reduced set of the MHD variables. Future work will exploit the spatiotemporal coherent patterns of the full-cadence Stokes data to infer the full MHD state vector along each line of sight. The MHD data and neural network models will be publically available.