Fall 1999 MIRA Quarterly Newsletter: Decoding the Mysteries of Binary Stars

Fall 1999

Decoding the Mysteries of Binary Stars
Dr. William Bruce Weaver, MIRA Director

binary.jpg (32382 bytes)
Chesley Bonestell’s painting of the interacting binary star
system Beta Lyrae as viewed from an imaginary planet.
Most of the stars in our Galaxy are binaries.

At least 30% of stellar objects are composed of two or more stars. Some astronomers feel that all stars have at least a giant planet as a companion; a second massive body may be needed in forming stars so that the initial angular momentum of the system can be shared. Binaries provide invaluable information about stellar masses, luminosities, interiors, sizes, and several other fundamental properties.

However, most binaries are close enough to each other or far enough away from the Earth that they appear as a single object through a telescope. These close binaries can be detected in a number of ways: if the Earth lies in the orbital plane of the system, the stars may periodically eclipse each other, the object may show one or two sets of spectral lines that shift back and forth due to the doppler shifts as the stars orbit one another, or the spectrum may show the characteristics of stars of two different temperatures or luminosities.

The spectra of single stars are quite complicated; when the spectra of two or more are combined, the possibilities are seemingly endless. For example, Figure 1 shows the components of a common combination: an A dwarf star (A0 V) and a K giant (K 5 III). This is a common combination because, in old age, A stars become K giants; so, a binary star system originally composed of two A stars will become an A dwarf-K giant combination when the more massive of the two A stars evolves first into the cooler giant. The combined spectrum may be further complicated by the fact that the relative brightnesses of the two components can also vary.

Figure 1. The MIRA near infrared spectra of Vega and Bright Star Catalog 6705 (g Draconis)
are combined in the computer to generate a synthetic binary spectrum (top tracing).

In order to understand these systems one must be able to classify the spectra of the component stars. A few years ago, Dr. Ana Torres-Dodgen and I extended the standard stellar classification system to the near infrared. Later we applied Artificial Neural Networks (ANNs) to the problem of classifying these spectra. ANNs are computerized analogues to the neural networks of living brains (wetware). We published two research articles showing that ANNs could classify stars as well as the best human experts.

For the last two years, I have been applying this same technique to the classification of combined (binary) star spectra. Starting with thousands of synthetic binary spectra made up by combining appropriate single star spectra that we observed, ANNs learn to classify the components by practicing on known spectra. A typical ANN architecture is shown in Figure 2. Because there can be over 25,000 interconnections that have to be determined, even high speed computers can take days to learn to classify these objects.

Figure 2. Thousands of binary spectra, each composed of intensities
at 512 colors, are combined at a few "hidden" neurons before being used
by the network to train itself in recognizing the key features that it will use to
classify the component stars.The extra feedback section is used to correct
for systematic errors at the extremes of the spectral images.

Because there are many neural network architectures to explore, I often have four or five computers working simultaneously day and night on forming new solutions. After about two years of continuous computations, I have started to arrive at the best solutions. An example of the results is shown in Figure 3. As expected, as the stars contribute less light to the system (the left hand side of the graph), the errors become larger.

Figure 3. The errors in determining the temperature classes for the hotter
star of a binary pair. Stars on the right are those that contribute most of the
light in a system, while those on the left, with larger errors, are minor contributors.

The primary difficulty in evaluating how well the ANN classification system works is that this kind of classification is very difficult for humans. These results from the synthetic binary spectra are much better than a human expert can do. How can I test the ANNs on real binary spectra when the ANNs will probably classify the spectra better than any other technique?

I hope to finish the research by the end of the year but, even so, it will have taken nearly three years to complete (not counting the years at the telescope to gather the single star spectra!). Each time I think I have finished the project, a new idea comes up which requires a few more months of computations to check out.

Of course, none of this would have been possible without the support of you, the Friends of MIRA. Many thanks for helping us develop this new tool for probing the well-guarded secrets of binary stars.