Tilman Hartwig

Assistant Professor at the University of Tokyo

Since April 2023, I work as Chief Data Scientist in the Application Lab for AI and Big Data for the German Environment Agency.


When did the first stars form? How can we use metal-poor stars as time machine to the early Universe? Can artificial intelligence help us to discover new science? Why are the first supermassive black holes so heavy? And how can we make research more inclusive? These are some of the major questions that I try to answer together with my colleagues. You can find more about my research under Publications.

My invited talk at the SAZERAC conference on "The First Stars"

Student Research Projects

Our groups offers several interesting research projects for undergraduate and graduate students. The University of Tokyo also offers two fellowships for international students in physics and astronomy: GSGC program, MEXT Fellowship. We are very happy to assist you with the application. Please contact me for further details.

Recent Articles

Unveiling the contribution of Pop III stars in primeval galaxies at redshift > 6


To facilitate and improve the analysis of high-redshift galaxies with JWST, we have estimated the relative contribution of the first stars to such primeval galaxies. Our semi-analytical model A-SLOTH allows us to simulate a wide range of halo masses, redshifts, and explore the cosmic variance, which was not possible with previous simulations.

For example, this figure shows the relative contribution of Pop III stars to the bolometric luminosity of galaxies of different halo masses (colours) at different redshifts. We should aim for low-mass galaxies at high redshift to find Pop III-dominated stellar populations.

Neural Networks Fail to Learn Periodic Functions and How to Fix It


Deep neural networks (DNN) have two problems: they can not generalise well beyond the interval on which they were trained, and they fail at modelling periodic functions. We derive the extrapolation theorem that shows that the activation function dictates the limit of the extrapolation behaviour of a DNN. We propose a new activation function, x+sin(x)2, and show that it provides a similar learning performance as previously used activation functions, such as ReLU and Tanh, on standard tasks. We also demonstrate the advantage of this new implementation on periodic problems, such as financial and meteorological data.

We fit a superposition of two periodic signals (blue) with a simple recurrent neural network. The prediction (orange) with our new activation function (bottom panel) is better compared to other activation functions.

Implications of inhomogeneous metal mixing for stellar archaeology


We investigate how metals from the first supernovae mix with the previously pristine ISM. For this purpose, we define the metallicity shift “dZ” that quantifies the difference in metallicity of the star-forming gas and the average metallicity of a galaxy. If the Pop III supernova and the second-generation star formation occur in the same galaxy (internal enrichment, blue points), the metals mix fairly homogeneous with the hydrogen. However, if the supernova occurs in a neighbouring halo (external enrichment, red points), the metallicity of the star-forming gas can be much lower than the average metallicity. We discuss under what circumstances this effect can have observational consequences and propose a new way to implement inhomogeneous metal mixing in a semi-analytical model.

Scatter plot between average metallicity of all gas vs. the metallicity difference between dense gas and average gas.

Formation of carbon-enhanced metal-poor stars as a consequence of inhomogeneous metal mixing


We present a novel scenario for the formation of carbon-enhanced metal-poor (CEMP) stars. Carbon enhancement at low stellar metallicities is usually considered a consequence of faint or other exotic supernovae. An analytical estimate of cooling times in low-metallicity gas demonstrates a natural bias, which favours the formation of CEMP stars as a consequence of inhomogeneous metal mixing: carbon-rich gas has a shorter cooling time and can form stars prior to a potential nearby pocket of carbon-normal gas, in which star formation is then suppressed due to energetic photons from the carbon-enhanced protostars.

Illustration of how a C-normal SN ([C/Fe] < 0.7) can trigger the formation of CEMP stars ([C/Fe] > 0.7). The crucial quantities are the separation of the two clumps, d, and their carbonicity difference, ∆[C/H].

Descendants of the first stars: the distinct chemical signature of second-generation stars

http://arxiv.org/abs/1801.05044, http://arxiv.org/abs/1810.04713

We have developed a new diagnostic to distinguish mono- from multi-enriched second-generation stars based on their chemical fingerprint. These results will help to classify spectroscopic observations of extremely metal-poor stars and thereby allow us to infer the masses and multiplicities of the first stars.

Probability of mono-enrichment for magnesium over carbon vs. the metallicity. There are regions of the parameter space in our model with a probability of almost 100% for finding second-generation stars that formed from gas that was enriched by only one previous supernova.

Probability distributions for mono-enrichment for seven recently observed extremely metal-poor stars from the TOPoS survey. The black stars mark the probability for the observationally derived abundances and the blue contours illustrates the probability density for values within 0.5dex of this derived abundance. Only one out of seven stars is more likely to be mono-enriched.

Gravitational wave signals from the first massive black hole seeds


We investigate the gravitational wave signature from supermassive stars that form as binaries at high redshift. In our optimistic model we find up to 0.6 detections per year that can be clearly attributed to such progenitors. Also a non-detection of these binary black hole mergers at z>15 over the lifetime of LISA can provide upper limits on the abundance and binary fraction of supermassive stars.

Rates for the merging of black hole binary from supermassive stellar binaries as a function of redshift. Only our optimistic scenario with Jc = 30J21 and binarity of 100% can produce a population of BHB mergers at z>15 that are clearly distinguishable from other channels of binary black hole formation.

Active Galactic Nuclei outflows in galaxy discs


Galactic outflows, driven by active galactic nuclei (AGN), play a crucial role in galaxy formation and in the self-regulated growth of supermassive black holes. With a novel 2D analytical model for AGN-driven outflows in a gaseous disc we demonstrate the main improvements, compared to existing 1D solutions. We find significant differences for the outflow dynamics and wind efficiency (see Figure). The recovery time of gas in the disc plane is remarkably short, of the order 1Myr. This indicates that AGN-driven winds cannot suppress BH growth for long. Published in MNRAS, 2018, 476, 2288.

Time evolution of the position of the wind shock front in polar coordinates for our fiducial model. Top: 2D solution with the galactic disc in the horizontal plane. Bottom: 1D spherical solution. The concentric grey rings indicate the radius in log(r/pc) from 10pc to 100kpc. Whereas in the 1D model all the gas is ejected out of the halo, only 10% of the galactic gas is ejected perpendicular to the disc plane in our improved 2D model.

Gravitational Waves from the Remnants of the First Stars


Initiated by the groundbreaking observation of the first gravitational wave signal in 2015, we studied the detectability of the first stars by the merger of their remnant black holes. We developed a cosmologically representative dark matter merger tree and coupled it to a model for stellar binary evolution. The overall contribution from the first stars is small, but individual events can still be detected with aLIGO and depending on the initial mass function of the first stars, we will see several black hole mergers in the next years that originate from the first stars. Turning this result around, we can use the upcoming detections of gravitational waves to learn more about the masses and binarity of the first stars. Our catalogues of Pop III binaries are publicly available here. Published in MNRAS, 2016, 460, L74

Expected number of BH-BH merger detections per year as a function of the total binary mass for the current aLIGO sensitivity (top) and final design sensitivity (bottom). The mass range of GW150914 is indicated by the grey area. With sufficient detections around Mtot ~300Msun, we could discriminate different Pop III IMFs based on their GW fingerprint.

© 2022 Tilman Hartwig