Listening to the Stars: Predicting Massive Star Properties with Machine Learning

Massive stars are some of the brightest and most influential objects in the galaxy. They produce powerful winds, explode as supernovae, and leave behind neutron stars or black holes. But figuring out their fundamental properties, like surface temperature, surface gravity, and a special measure called spectroscopic luminosity, usually requires detailed and time-consuming spectroscopic observations. In contrast, space missions like NASA’s Transiting Exoplanet Survey Satellite (TESS) gather light curves (records of how a star’s brightness changes over time) for hundreds of thousands of stars, though these data do not directly reveal the same physical details. With the upcoming Legacy Survey of Space and Time (LSST) set to monitor millions more stars, Rachel Zhang and collaborators ask: can machine learning bridge the gap between light curves and stellar parameters?

From Spectra to Light Curves

The authors focus on O-type stars, a rare but important group of massive stars. These stars were observed in detail by the IACOB project, which provided accurate spectroscopic parameters for over 250 O stars. To pair these with TESS photometry, Zhang and colleagues selected 285 usable light curves from 106 stars. This combined dataset allowed them to test whether brightness variations contain hidden clues about stellar properties like effective temperature (Teff), surface gravity (g), and spectroscopic luminosity (a combination of Teff and g that places stars on a specialized Hertzsprung–Russell diagram).

Training the Machines

The team compared two strategies. In the first, they used prior work showing that massive stars often display “red noise”, a type of slow, random variation thought to be linked to processes like internal gravity waves or turbulent winds. They extracted four numbers (parameters) describing this noise from each light curve and fed them into a type of neural network called a multilayer perceptron (MLP). In the second approach, they transformed each light curve into a detailed frequency plot called a periodogram and used convolutional neural networks (CNNs), which are better at identifying patterns in complex data. Both models were trained to connect the inputs (red noise or periodograms) to the stellar parameters measured by spectroscopy.

Comparing Results

The MLP models, which relied on just the fitted red noise parameters, performed poorly. They captured some predictive signal, but not enough to reliably match spectroscopic measurements. By contrast, the CNN models achieved substantially higher accuracy, especially when predicting spectroscopic luminosity. For example, CNNs explained over 60% of the variance in luminosity values, while MLPs explained only about 25%. Predictions of effective temperature were harder, but CNNs still performed better, recovering meaningful trends in the data. Importantly, the CNN results showed that the full frequency-domain information from light curves holds far more predictive power than the limited red noise descriptors.

Why It Matters

By demonstrating that light curves contain enough information to estimate fundamental properties of massive stars, this study moves beyond previous correlations and toward actual predictive models. This is especially valuable in the era of large surveys like LSST, which will collect light curves for millions of stars without accompanying spectroscopy. Machine learning can act as a shortcut, providing first estimates of stellar parameters and highlighting the most interesting stars for follow-up observations. Although the dataset in this study was small, the results suggest that larger matched datasets could unlock even more accurate predictions.

Looking Ahead

Zhang and collaborators conclude that convolutional neural networks trained on periodograms are the most promising path forward. They emphasize that their work is not a replacement for spectroscopy, but a complement: a way to sift through enormous amounts of photometric data and prioritize targets for deeper study. In doing so, this method could help astronomers map out the life cycles of massive stars more efficiently, offering new insights into how these cosmic giants evolve and shape the galaxies around them.

Source: Zhang