Stellar age determination using deep neural networks: Isochrone ages for 1.3 million stars, based on BaSTI, MIST, PARSEC, Dartmouth and SYCLIST evolutionary grids

This paper introduces NEST, a model-driven deep learning framework that rapidly estimates stellar ages for over 1.3 million stars across multiple evolutionary grids with high accuracy and a 60,000-fold speedup compared to traditional Bayesian methods, thereby enabling large-scale galactic archeology studies.

The Gist

Here is an explanation of the paper, translated from "Astrophysicist" to "Human."

The Big Picture: The Cosmic Clock Problem

Imagine the Milky Way galaxy as a giant, bustling city. To understand how this city was built, when its neighborhoods were formed, and how it has changed over billions of years, astronomers need to know one thing above all else: How old are the stars?

Knowing a star's age is like knowing the birth date of a person. It tells you if they are a baby, a teenager, or an elder. But here's the catch: stars don't have ID cards. They don't wear watches. To figure out their age, astronomers have to look at their "vital signs"—how bright they are, what color they are, and what chemicals they are made of—and compare them to a theoretical map of how stars should look at different ages.

For a long time, doing this math was like trying to solve a massive, complex puzzle by hand. It was slow, expensive (in terms of computer power), and if you wanted to date a million stars, it could take your computer years to finish.

The New Solution: The "Cosmic Speed Reader"

This paper introduces a new tool called NEST (Neural Estimator of Stellar Times). Think of NEST as a super-fast, super-smart "Cosmic Speed Reader."

Instead of solving the complex puzzle for every single star from scratch, the authors taught a computer program (a Deep Neural Network) to recognize the patterns of stellar ages.

Here is the analogy:

• The Old Way (Bayesian/Isocrone Fitting): Imagine you are trying to guess the age of a stranger. You pull out a thick encyclopedia of human growth, find the page that matches their height and weight, and spend 20 minutes carefully calculating the probability of their age. You do this for every person you meet. It's accurate, but incredibly slow.
• The New Way (Deep Learning): Imagine you spent a few weeks studying that same encyclopedia until you memorized every pattern. Now, when you see a stranger, you can instantly guess their age with a glance. You don't need the book anymore; your brain has become the encyclopedia.

How They Built the "Brain"

The authors didn't teach the computer by showing it real stars with known ages (which can be biased or wrong). Instead, they fed the computer theoretical models.

1. The Textbooks: They used five different "textbooks" of stellar physics (called grids: BaSTI, MIST, PARSEC, Dartmouth, and SYCLIST). These are massive mathematical simulations of how stars evolve.
2. The Training: They showed the computer millions of simulated stars from these textbooks, teaching it: "If a star looks like this (color, brightness, metal content), it is this old."
3. The Result: The computer learned the rules of the universe so well that it could predict the age of a real star in a fraction of a second.

The "Speed Run" Comparison

The paper highlights just how fast this new method is.

• The Old Method (SPInS): To calculate the age of one star, it takes about 20 seconds.
• The New Method (NEST): In those same 20 seconds, it can calculate the age of 60,000 stars.

That is a 60,000x speedup. It's the difference between walking across a continent and teleporting there. This means astronomers can now date the entire Milky Way's star population in a single afternoon, rather than waiting decades.

Testing the Tool: The "Class Photo" and the "Crowd"

To make sure their new "Speed Reader" actually worked, the authors tested it in two ways:

1. The Class Photo (Star Clusters):
They looked at groups of stars born at the same time (like a school class). Since everyone in the class should be the same age, this is a perfect test.

• The Result: Their method guessed the ages of these "classes" with an error margin of only 0.2 billion years. This is incredibly precise, proving the tool is reliable.

2. The Crowd (Field Stars):
They applied the tool to 1.3 million individual stars from massive surveys (like LAMOST, GALAH, and APOGEE).

• The Result: The ages matched up very well with other existing catalogs, but with much less "noise" (random errors).
• The Surprise: They discovered that different "textbooks" (evolutionary models) sometimes give slightly different answers. For example, one model might say a star is 10 billion years old, while another says 8 billion. This proves that the biggest source of error isn't the computer speed, but the underlying physics models we use to understand stars.

Why This Matters: Galactic Archaeology

The authors call this field Galactic Archaeology. Just as archaeologists dig up pottery shards to understand ancient civilizations, astronomers look at stars to understand the history of our galaxy.

• The "Thin" and "Thick" Disks: By dating millions of stars, they confirmed that our galaxy has two main "neighborhoods" (disks). The older stars (the "Thick Disk") are metal-poor and formed early in the universe's history. The younger stars (the "Thin Disk") are metal-rich and formed more recently.
• Future Surveys: Upcoming telescopes will find billions of new stars. Without a tool like NEST, we would be overwhelmed. With NEST, we can instantly turn that data into a history book of the Milky Way.

The Takeaway

This paper isn't just about a faster computer program; it's about unlocking the history of our galaxy. By teaching a neural network to read the "life story" of stars, the authors have given astronomers a time machine. They can now look at a star, glance at its color and brightness, and instantly know when it was born, allowing us to reconstruct the evolution of the Milky Way with unprecedented speed and detail.

In short: They built a cosmic speed-reader that can date 1.3 million stars in the time it used to take to date just one, helping us finally write the biography of our home galaxy.

Technical Summary

Here is a detailed technical summary of the paper "Stellar age determination using deep neural networks" by Boin et al. (2026).

1. Problem Statement

Determining stellar ages is a fundamental yet notoriously difficult task in Galactic archaeology. Traditional methods, such as Bayesian isochrone fitting (e.g., using the SPInS pipeline), are computationally expensive, often requiring the loading of full evolutionary grids and Markov Chain Monte Carlo (MCMC) sampling for every star. This limits their application to massive spectroscopic catalogs (millions of stars) from surveys like LAMOST, GALAH, and APOGEE.

Furthermore, existing machine learning approaches for age determination are predominantly data-driven. They rely on training sets of stars with pre-existing age estimates (derived from asteroseismology, gyrochronology, or previous isochrone fits). This introduces a "bias cascade," where errors and systematic biases from the original age estimation methods are inherited and potentially amplified by the neural network, restricting the model's applicability to specific subsets of data.

2. Methodology

The authors propose a model-driven deep learning approach that bypasses the need for pre-dated training data. Instead, the neural networks are trained directly on theoretical stellar evolutionary grids.

• Neural Network Architecture:

  • Type: Multilayer Perceptrons (MLP).
  • Structure: 4 hidden layers with 32, 64, 64, and 32 neurons respectively.
  • Activation: ReLU.
  • Optimizer: Adam solver with a constant learning rate of $1 \times 10^{-3}$.
  • Inputs: The network takes three primary observables: Metallicity ($[M/H]$), Absolute Magnitude ($M_G$), and Color ($G_{BP} - G_{RP}$). The authors found that adding separate $G_{BP}$ and $G_{RP}$ magnitudes or effective temperature ($T_{eff}$) did not improve performance and sometimes degraded it due to observational inconsistencies between $T_{eff}$ and color.
  • Output: Stellar age ($\tau$).

• Training Data:

  • The networks were trained on five distinct stellar evolutionary grids: BaSTI, MIST, PARSEC, Dartmouth, and SYCLIST (Geneva).
  • The training sets included all evolutionary phases (except pre-main sequence) to create a general tool, though the authors note that Main Sequence (MS) ages inherently carry higher uncertainties due to isochrone degeneracy.
  • Preprocessing: Input parameters were standardized (mean subtraction and standard deviation division) based on the training set statistics. Metallicity offsets were applied to align different grids to a common solar mixture reference (Caffau et al. 2011).

• Uncertainty Estimation (Monte Carlo Sampling):

  • Since the NN outputs a single deterministic value, the authors employ a Monte Carlo (MC) sampling method to estimate uncertainties.
  • For each star, 10,000 synthetic samples are generated by perturbing the observed inputs ($x_{obs}$) according to their Gaussian uncertainties ($\Delta x_{obs}$).
  • These samples are passed through the trained network to generate a distribution of ages.
  • The median of this distribution is taken as the final age estimate, and the standard deviation provides the uncertainty.

• Cluster Age Estimation:

  • For star clusters, the global age Probability Density Function (PDF) is derived by multiplying the individual star PDFs. To avoid vanishing probabilities due to outliers, a small $\epsilon$ term is added, and a bootstrapping approach (100 iterations) is used to determine the cluster age and its random uncertainty.

3. Key Contributions

1. Model-Driven Paradigm: The paper introduces a method where neural networks learn the physics of stellar evolution directly from theoretical grids rather than observational age labels. This eliminates biases inherent in the training data of previous ML methods.
2. Computational Efficiency: The method achieves a 60,000x speedup compared to traditional Bayesian isochrone fitting (e.g., SPInS). While SPInS takes ~20 seconds per star, the NN can process 60,000 stars in the same timeframe (including MC uncertainty estimation).
3. Cross-Grid Analysis: The authors systematically compare age estimates derived from five different evolutionary grids (BaSTI, MIST, PARSEC, Dartmouth, Geneva) applied to the same massive datasets, quantifying the systematic differences arising from the underlying physics of the models.
4. NEST Package: The release of NEST (Neural Estimator of Stellar Times), a Python package and web interface, allowing the community to apply these trained models to new data.

4. Results

• Validation against SPInS: When trained on the BaSTI grid, the NN results match the SPInS (Bayesian) results almost perfectly (median absolute deviation < 0.05 Gyr for most ages), confirming that the NN successfully learns the theoretical grid.
• Grid Comparisons:
  • MIST and Dartmouth: Show very little systematic offset compared to BaSTI (median std dev ~0.3–0.4 Gyr).
  • PARSEC: Systematically yields younger ages than BaSTI by approximately 1.5 Gyr.
  • Geneva (SYCLIST): Shows significant divergence, yielding ages up to 3–4 Gyr younger than BaSTI for old stars.
  • Conclusion: The choice of evolutionary grid is a primary source of systematic uncertainty (1–2 Gyr) in age determination.
• Cluster Ages: Applied to 13 open clusters and 1 globular cluster. The method achieves a Median Absolute Deviation (MAD) of 0.20 Gyr against literature ages, outperforming the SPInS-based results from Casamiquela et al. (2024) which had a MAD of 0.27 Gyr. Younger clusters showed slightly higher estimated ages, likely due to MS isochrone density.
• Field Star Catalogs: Applied to 1.3 million stars from LAMOST DR10, GALAH DR3/DR4, and APOGEE DR17.
  • Results show good agreement with other major age catalogs (StarHorse, Nataf et al., Xiang et al., Kordopatis et al.).
  • The Xiang et al. (2025) catalog showed the best agreement (lowest dispersion, ~0.79 Gyr).
  • LAMOST DR10 MRS data showed the highest dispersion, suggesting potential issues with the specific metallicity estimates or selection cuts in that dataset.
• Age-Chemistry Relations: The derived catalogs reveal clear Age-Metallicity Relations (AMR) and Age-$\alpha$ relations. A "knee" is observed at ~10 Gyr, separating the thin disk (flat metallicity, low $\alpha$) from the thick disk (steeper metallicity decline, high $\alpha$).

5. Significance

This work represents a paradigm shift in stellar age determination for large-scale surveys. By decoupling the age estimation process from the computational cost of MCMC fitting and the biases of data-driven training sets, the authors provide a scalable solution for the upcoming era of massive spectroscopic surveys (e.g., 4MOST, SDSS-V, WEAVE).

The ability to generate age catalogs for millions of stars in a fraction of the time required by traditional methods enables high-resolution studies of Galactic evolution, chemical enrichment, and kinematic history that were previously computationally prohibitive. The release of NEST democratizes access to these model-driven age estimates, allowing researchers to explore the impact of different evolutionary physics on Galactic archaeology conclusions.