StellarF: A Physics-Informed LoRA Framework for Stellar Flare Forecasting with Historical & Statistical Data

TL;DR

StellarF tackles the unpredictability of stellar flares by integrating physics-informed priors with a LoRA-finetuned large language model and multimodal representations of light curves, historical flare records, and flare statistics. It introduces a unified light-curve preprocessing pipeline, two textual information modules (FHR and FSI), and a physics-informed cross-entropy loss that imposes a minimum rising-rate prior on positive flare predictions, all trained via LoRA for efficiency. Empirical results on Kepler and TESS surpass state-of-the-art baselines across accuracy and AUC, with ablations showing the critical contributions of first-order differencing, historical/statistical multimodal inputs, and the physics prior. This framework enables physically interpretable, data-efficient forecasting of transient stellar events and provides a reproducible, open-source approach for AI in astrophysics.

Abstract

Stellar flare forecasting represents a critical frontier in astrophysics, offering profound insights into stellar activity mechanisms and exoplanetary habitability assessments. Yet the inherent unpredictability of flare activity, rooted in stellar diversity and evolutionary stages, underpins the field's core challenges: (1) sparse, incomplete, noisy lightcurve data from traditional observations; (2) ineffective multi-scale flare evolution capture via single representations; (3) poor physical interpretability in data-driven models lacking physics-informed priors. To address these challenges, we propose StellarF, a physics-informed framework synergizing general Al with astrophysical domain knowledge via three core components: a unified preprocessing pipeline for lightcurve refinement (missing-value imputation, temporal patch partitioning, adaptive sample filtering); a Low-Rank Adaptation (LoRA)-finetuned large language model (LLM) backbone enhanced by first-order difference augmentation, flare statistical information, and flare historical record modules for multimodal fusion instead of only simple representations; and a novel physics-informed loss embedding a minimum rising rate prior, appended to the cross-entropy loss, to align with flare physics. Extensive experiments on Kepler and TESS datasets show StellarF achieves state-of-the-art performance across key metrics, setting new benchmarks for flare forecasting. This work bridges general AI with astrophysics, offering a practical, physically interpretable paradigm for transient event forecasting in time-domain astronomy.

Figures (8)

• Figure 1: Some observed images of stellar flares.
• Figure 2: Light Curves of Three Stars with Flare Eruption Timings Marked. Visualizes observation period variations and light curve pattern differences across stars.
• Figure 3: Overall Architecture of StellarF: It consists of a unified preprocessing pipeline (linear interpolation, patch-based sampling, adaptive sample filtering) for lightcurve refinement; a LoRA-tuned LLM backbone enhanced by first-order difference augmentation and dual multimodal fusion modules (statistical + historical record) for multimodal representation learning, followed by an MLP prediction head; and is further constrained by our novel physics-informed loss term, which embeds a minimum rising rate prior and is added to the cross-entropy (CE) loss to align with flare physics.
• Figure 4: Textual Representations of Structured Data: Historical Flare Eruption Records (upper) and Statistical Summaries (count, median flux) (bottom) for Example Stellar Flares. Note: The model prompt used in the experiment was in Chinese.
• Figure 5: IG Attribution Maps (Upper: Positive Samples; Lower: Negative Samples)