{"id":90,"date":"2025-11-10T12:02:48","date_gmt":"2025-11-10T11:02:48","guid":{"rendered":"https:\/\/cornerstone.web.roma2.infn.it\/?p=90"},"modified":"2025-11-10T12:02:48","modified_gmt":"2025-11-10T11:02:48","slug":"physics-informed-features-for-fair-and-more-interpretable-solar-flare-forecasting","status":"publish","type":"post","link":"https:\/\/cornerstone.web.roma2.infn.it\/?p=90","title":{"rendered":"Physics-Informed Features for Fair and More Interpretable Solar Flare Forecasting"},"content":{"rendered":"\n

A new CORNERSTONE\u2019s recent work explores how physics-driven feature design can make solar flare forecasting not only accurate but also physically meaningful.<\/p>\n\n\n\n

Physics-Informed Learning: Embedding Physical Consistency in AI<\/strong><\/p>\n\n\n\n

In most data-driven approaches to solar flare prediction, models are trained on sets ofStandardized Features (SFs)<\/strong> \u2014 numerical representations of solar magnetic activitynormalized for machine learning. The goal is typically a binary prediction: whether a flare<\/strong>will occur within a given time window or not<\/strong>.<\/p>\n\n\n\n

These features \u2014 though rich in information \u2014 di\ufb00er in their physical units. Applying linear models directly to them means combining quantities that describe fundamentally di\ufb00erent aspects of the magnetic field, which is statistically valid but physically inconsistent.<\/p>\n\n\n\n

In this work, the standardized features are derived from the SHARP data<\/strong> recorded by the Helioseismic and Magnetic Imager (HMI) onboard NASA\u2019s Solar Dynamics Observatory (SDO). To overcome the issue of dimensional inconsistency, these parameters are combined through operations that align their physical dimensions, yielding a new set of Standardized Physics-Informed Features (SPIFs)<\/strong>. All SPIFs share the same unit of energy J.<\/p>\n\n\n\n

This creates a physically coherent feature space<\/strong>, where each variable represents a comparable quantity. Linear operations on this space naturally lead to a clear physical interpretation, something not possible when using the raw SHARP parameters or the standard feature sets alone.<\/p>\n\n\n\n

A Fairer and More Interpretable Framework<\/strong><\/p>\n\n\n\n

To investigate the practical implications of this approach, di\ufb00erent machine learning models \u2014 Lasso, GroupLasso, Random Forest Regressor, and Support Vector Classifier <\/strong>\u2014 were trained and evaluated on both SFs and SPIFs. Performance was assessed using the True Skill Statistic (TSS), <\/strong>a standard metric in space weather forecasting.<\/p>\n\n\n\n

The results were comparable between the two feature sets. However, using SPIFs provides the advantage of interpretability:<\/p>\n\n\n\n