From the windswept plateaus of the Atacama Desert to the farthest reaches of the universe, high-resolution spectroscopy offers a unique window into cosmic evolution. John K. Webb, a spectroscopist and astrophysicist based at the Institute of Astronomy in Cambridge, UK, is at the forefront of this endeavor. His research focuses on analyzing the faint light from distant quasars – extremely luminous, ancient objects whose spectra carry the imprints of intervening galaxies.
In this interview, Webb shares how he and his colleagues are using spectroscopy and AI-powered tools to unravel the chemical and physical history of the early universe – and, potentially, detect signs of life on faraway Earth-like planets.
Figure 1. The European Southern Observatory VLT.
Please could you give us an introduction to you and your work as a spectroscopist in the astrophysics field?
One of my research activities concerns the highest possible resolution spectroscopy of distant quasars. Although intrinsically very luminous, their cosmological distances mean they appear quite faint to an Earthling. Obtaining sufficiently high-quality data involves using the world’s largest optical telescopes such as the VLT in the Atacama, Chile (Fig. 1), with its ESPRESSO spectrographs, providing a resolution of R=145,000.
Because of their enormous distances, the long sightline between the Earth and a high redshift quasar unavoidably intersects gaseous halos around intervening galaxies. High resolution spectroscopy reveals absorption lines imposed on the light from the background quasar by this gas. The finite speed of light thus provides a unique view into the past, allowing us to study the physics of the universe at early times. By studying many such absorption systems at different cosmological distances, one can build up a detailed picture of the chemical and physical evolution of the universe.
Figure 2. Synthetic quasar spectrum. The wavelength is arbitrary since λobs=λrest(1+ζ), where ζ is the quasar emission redshift. The strongest emission feature near the centre is Lyman-ɑ plus NV. The rightmost features are SiIV and CIV (note the astrophysics notation: NV means quadruply ionised nitrogen, SiIV means triply ionized silicon, etc.). The plethora of absorption lines shortwards of green are the Lyman-ɑ forest, arising in extended gaseous regions around early galaxies and in the Cosmic Web. The strongest absorption feature is a damped Lyman-ɑ line. Longwards of Lyman-ɑ emission, where the Lyman-ɑ forest ceases to be present, heavier element lines are seen (they are also present in the forest, but blended with hydrogen).
What are some of the main challenges involved in analyzing distant quasars?
A sightline through even a single galaxy halo may encounter a large number of “cloudlets”; the complex distribution of the gas seen around galaxies forms a correspondingly complex absorption pattern. We have no a priori knowledge of the number of absorption components that may be seen. This makes theoretical modelling of the clouds complicated; we do not know how many components to fit. Moreover, since many elements are present, in different ionization stages, an extended gaseous galaxy halo can create many absorption components, which can blend together.
Historically, in order to build a statistically acceptable model, and hence derive parameter estimates (and their uncertainties) elucidating the physics of the gas, an individual analyst would painstakingly construct a complex model, component by component, species by species. That process is detailed and extremely time-consuming. A single quasar spectrum has very often formed the basis for a PhD thesis.
What led you to develop AI-VPFIT, and what specific challenges in spectral analysis does it address?
The tasks described above evidently cry out for automation. However, because of the complexity of the problem this has only been recently achieved, with the development of a new approach based on a hybrid system involving AI, statistical information criteria, and non-linear least squares. The system is called AI-VPFIT and has been described in detail in a suite of peer-reviewed papers published in astrophysics literature.
A key driver stems from the fact (mentioned above) that whilst it is easy to identify the chemical species, the number of absorption components making up a given complex is unknown. Since each absorption component may be randomly placed in wavelength (because of random motions or redshifts), subtle blends may occur. These are extremely difficult to identify unambiguously; two different (expert) analysts might produce almost identical models (producing almost identical normalized residuals, and almost identical values of chi-squared), and yet such models might possess quite different numbers of free parameters.
Figure 3 shows a pair of models, one derived using AI, requiring no human decision making, the other from a human interactive modeller. As can be seen, whilst the models are virtually indistinguishable, the number of free parameters and absorption components is significantly different in these two models. In this example, the human interactive modeller has over-fitted the data, and AI-VPFIT finds a far simpler kinematic structure, the number of absorption components is determined in a reproducible way using the SpIC information criterion. Figure 4 illustrates a different section from the same absorption complex, where Mg 2796Å is saturated, but where other species are not, such that modelling the combined dataset can find a good model for the kinematic structure.
Figure 3. Illustrating two different models fitted to the same spectral data. The red curve (AI-VPFIT) required no human decision making. The blue curve was fitted interactively by an expert. The modelling was carried out simultaneously modelling several transitions in several species, only one of which is shown here. The tick marks show absorption component placement, and hence illustrate model non-uniqueness. Model non-uniqueness increases as blending becomes worse. Parameter estimates for these two models obviously vary considerably, yet the models are virtually indistinguishable.
In other words, model non-uniqueness plays an important role in modelling uncertainties. Non-uniqueness can, in some cases, dominate the overall uncertainties. Therefore the important considerations or motivations in developing AI-VPFIT were (i) the requirement of a fully automated replacement that could produce a theoretical model at least as well as a human analyst, and (ii) a reproducible method to decide on the appropriate number of absorption components. For the former, we adopted a hybrid approach formed from the amalgamation of a deterministic (a non-linear least-squares method) and a heuristic technique (a genetic algorithm). This approach is referred to as a “memetic algorithm” by some authors.
AI-VPFIT calculations are complex (as can be seen from the sidebar below). Depending on the complexity of the model, a high performance computer might be required. Because quasar absorption spectroscopy often ends up requiring several hundred free parameters, we generally carry out such calculations using the OzSTAR supercomputing facility in Australia. Parallel programming and/or porting to GPUs could potentially speed calculations significantly, but this has not been attempted so far. Nevertheless, the AI-VPFIT performance dramatically out-performs a human in several respects. Simpler problems with fewer free parameters can be done on a laptop.
SpIC and Span
To create a reproducible method to decide on the appropriate number of absorption components (see point (ii) above), we implement an Information Criterion (IC). The general form of an IC is IC=Χ2 + P where Χ2 is defined in the usual way and P is a penalty term that increases as a function of the number of free parameters introduced into the fit. This is a crucial part of the procedure because the IC minimises as a function of the number of free parameters, enabling AI-VPFIT to decide on an optimal model without human input. The penalty term P depends on which IC is applied. The most commonly used are the Bayesian Information Criterion (BIC) and the Akaike Information Criterion (AIC). It is well known that the BIC has a tendency to underfit the data (too few parameters) and an AIC tends to overfit (too many parameters).
Hence, we invented a new IC, designed specifically for spectroscopy that optimizes the number of free parameters, and also introduces a more appropriate parameter weighting scheme (taking into account that not all parameters produce an equal sensitivity to a variation in Χ2). The new IC is called the Spectroscopic Information Criterion, or SpIC.
Finally, the local non-linear least-squares procedure implemented within the AI-VPFIT system, carried out within each generation of the Genetic calculation, is itself a hybrid of the well-known Gauss-Newton and Levenberg-Marquardt methods. This new hybrid method was designed to produce a slightly more efficient calculation (requiring less iterations).
Figure 4. Illustrating two different species in the same absorption system. MgII is mostly saturated, so cannot provide reliable component structure. However, FeII 2344 (and other species, not shown) is unsaturated, enabling AI-VPFIT to find a multi-component model. Normalised residuals are shown above each transition.
You mention the system is specific to astrophysical data – could it be adapted for other spectroscopic domains or applications?
AI-VPFIT was specifically developed for high resolution absorption line spectroscopy, the core model being the Voigt profile. But AI-VPFIT could certainly be adapted to other scientific problems. Its value arises when (a) the number of model parameters is large, such that the task for a human is onerous, and (b) when the number of model components is unknown and is best decided using an IC. These are the two circumstances under which AI-VPFIT can save an enormous amount of human time and also produce more reliable results. The core function need not be Voigt or any other absorption model, and could instead be Gaussian emission or some other function. One could imagine far-reaching applications, well removed from the original astrophysical context, perhaps in medicine, chemistry, biology, and so forth.
How do you see the role of AI evolving in spectroscopy, and what are the potential limitations we should be aware of?
I see AI playing an increasing role in spectroscopy, the extent depending on the user requirement. One very interesting possibility that AI-VPFIT opens up is the option of what we call “Monte Carlo AI”. In our own work, we discovered situations where Χ2 - parameter space can be complex, exhibiting multiple minima for a parameter of interest. It is very important the user can be made aware of such occurrences. Since AI-VPFIT requires no human input, it becomes possible to run AI-VPFIT multiple times, each time using different random number seeds associated with trial absorption component placement. This means that AI-VPFIT can emulate the range of possibilities created by independent expert human modelers. It is important to identify this kind of model non-uniqueness when it occurs (it does not always, in our own astrophysics application). This evidently represents a limitation of a human analyst, not AI.
Do you have any thoughts regarding our current ability to detect molecules on Earth-like planets using spectroscopy?
Many years ago (in 2001) I wrote a paper entitled “Could We Detect Molecular Oxygen in the Atmosphere of a Transiting Extra-Solar Planet?” This subject had developed enormously since that paper, but life elsewhere has still not yet been detected. And yet we may be fortunate enough to see even more important discoveries soon – especially given recent developments in instrumentation and analytic techniques, such as AI-VPFIT. But such a detection is complicated (for ground-based telescopes) by the presence of oxygen absorption bands arising in Earth’s atmosphere, which dwarf the tiny signal from the extrasolar planet. The saving grace is the extrasolar planet’s velocity with respect to Earth, providing a small offset between the terrestrial and extrasolar lines. However, this has not yet been achieved. Recent interest has been heightened by tentative suggestion of absorption from dimethyl disulfide (DMDS), a potential biosignature gas, using the James Webb Space Telescope, which evidently avoids confusing terrestrial features. In such cases, where unknown species may be present and the correct model is not known a priori, AI-VPFIT offers a promising approach.