Machine Learning in Cosmology (Münchmeyer)

Extracting information about fundamental physics from the massive amount of complicated data of current and upcoming cosmological surveys is a very difficult task. Machine learning gives us an entirely new toolkit to solve such problems and promises to enhance or replace traditional data analysis methods.

In Moritz Münchmeyer’s cosmology group we develop machine learning methods that are specifically designed for the physical problems we want to solve, and make use of the mathematical structure of the data.

Recent papers from our group in this domain: 2105.12024, 1905.05846

Topological data analysis coupled to Machine Learning (Shiu)

Topology provides a rigorous mathematical framework to identify interpretable patterns in noisy high dimensional data. Recent work of Gary Shiu’s group (2009. 14231), in the context of detecting and classifying phases of matter, further suggests that Topological Data Analysis (TDA) can be integrated into ML approaches to perform both unsupervised and supervised learning tasks. TDA reduces datasets into their summary statistics and yet reveals multi-scale shape information inaccessible by other techniques. The web-like features of the Large Scale Structure (LSS) of the universe make it a paragon for TDA methods. Dark matter tracers and voids have long been used by cosmologists as discerning probes of fundamental physics (e.g., neutrinos, dark energy). However, these observables can only be precisely defined and identified from noisy cosmological data using persistent topology. We have shown that TDA can provide a competitive detection of non-Gaussianity in the CMB (1712.08159) and LSS (2009.04819) data, and furthermore suggests new observables. The subsampling technique we developed to speed up TDA computation for cosmological data has wider applications, e.g., in neural science (2102.08623). The statistical tools we developed for TDA can be combined with other ML approaches to perform cosmological parameter inference of the LSS survey data.

TDA also finds applications in the high energy physics domain, e.g. in identification of classes of jets in a noisy collider environment. When combined with the low-dimensional latent representation from other ML approaches (e.g., variational autoencoder), we can obtain a combined unsupervised anomaly score or supervised event classification. As another example, we have uncovered structure and correlations in the string theory landscape using TDA (1812.06960). We further showed in 1907.10072 that stochastic optimization methods such as genetic algorithms can speed up the search for phenomenologically optimal models. This opens up the possibility to combine and contrast with reinforcement learning techniques in the study of the vast string theory landscape.

Machine Learning in High Energy Physics (CMS Group)

Machine learning techniques are used by the members of the CMS group led by Profs. Black, Bose, Dasu and Herndon for improving the sensitivity of new physics searches and increasing the accuracy of measurements.

Further machine learning techniques are being developed to more efficiently select events online using auto-encoders, using massively parallel processors used in the trigger systems. Because the event selection code is processed on GPUs and FPGAs ML techniques for usage on such resources is under investigation.

Machine Learning for the Physics of Materials (Gilbert)

Pupa Gilbert uses ML in collaboration with Markus Buehler at MIT to understand patterns in nanocrystal orientations (detected with Pupa’s new synchrotron methods) and fracture mechanics (detected at the atomic scale with MD simulations by Markus Buehler).

Machine Learning for Condensed Matter Physics (Vavilov)

A project in Maxim Vavilov’s group aims to develop a neural network that analyzes energy spectra of a superconducting device, such as the transom or fluxonium qubit and predicts microscopic parameters of the qubit. In the case of fluxonium, we analyze the spectrum as a function of the magnetic flux through the superconductor. Our research demonstrated that even a magnetic flux dispersion curve of a single 0-1 transition provides an accurate estimate for the three energy scales of the fluxonium: charging, inductive, and Josephson energies. The inclusion of higher energy transitions makes such predictions better. Our next steps are to analyze spectra of multiple fluxoniums and make predictions for optimized single and two-quits gates for the fluxonium systems.