Fast Machine Learning for Science Conference 2026

An Asteroseismic Inversion Framework Using Physics-Informed Neural Networks to Estimate Mixing Length Parameters in Evolved Stellar Grids

Not scheduled

12m

Presentation Contributed Talks

Speaker

Pritam Singh (Wichita State University)

Description

Determining the internal physical structures of evolved stars remains a central challenge in stellar astrophysics. Standard procedures involve combining 1D stellar evolution tracks from codes like MESA with linear perturbation data from GYRE to reproduce observed stellar oscillation spectra. Although these numerical techniques are highly accurate, mapping observable surface frequencies back to unobserved interior properties remains computationally intensive. Conventional parameter estimation workflows rely heavily on dense grid-searches or iterative Markov Chain Monte Carlo (MCMC) simulations. Because these methods require executing thousands of resource-intensive forward models for each target, they introduce a computational bottleneck that prevents real-time processing of high-throughput data streams from next-generation wide-field astronomical surveys.

To resolve this limitation, we present a computationally efficient inversion framework based on Physics-Informed Neural Networks (PINNs) designed to map observed stellar acoustic spectra directly to hidden envelope properties. We validate this architecture using the open-source Joyce et al. (2024) grid of over 3,000 Thermally Pulsing Asymptotic Giant Branch (TP-AGB) stellar models. Rather than relying on traditional data-driven regression, which fails to generalize when a key physical parameter lacks variance in the training data, our approach incorporates radius-density scaling laws and structural identities directly into the neural network's loss function via automatic differentiation.

The network uses basic atmospheric features alongside a detailed spectrum of the first ten radial pressure modes ($f_{1}$​ through $f_{10}$​) computed by GYRE. By penalizing structural configurations that violate mass conservation and hydrostatic equilibrium constraints, the physics layer forces the network to deduce the latent mixing length parameter ($\alpha_{MLT​}$) variations required to preserve mechanical consistency within the convective envelope. The trained framework accurately solves the inversion task, achieving precise parameter retrieval with sub-millisecond inference times on standard hardware. By substituting repetitive forward numerical evaluations with a physics-guided neural network architecture, this methodology establishes a fast, automated pipeline for characterizing evolved variable populations observed across large-scale astronomical surveys.