Prof. Cara Battersby, Department of Physics, University of Connecticut

The Milky Way Laboratory

Galaxy centers are the hubs of activity that drive galaxy evolution, from supermassive black holes to feedback from dense stellar clusters. While the bulk of our Milky Way Galaxy is a prime example of present epoch “normal” star formation, our galaxy’s center has gas properties that are more reminiscent of star formation during its cosmic peak. In our research group, the Milky Way Laboratory, we capitalize on both the “normal” and “extreme” star formation in our own cosmic backyard in order to resolve the interplay of physical processes in detail. In this talk, I will discuss efforts to measure how stars gain their mass and how the star formation process may vary across the Galaxy. In our galaxy’s central molecular zone, the process of star formation is complicated by constant gas inflow, high levels of turbulence, and more. I will present both simulations and observations toward this region that aim to understand the role of the gas inflow, the 3-D geometry of the region, properties of the gas, and incipient star formation.

Dr. Guochao Sun, Northwestern University

Unveiling the Physics of Galaxy Formation and its Large-Scale Effects at Cosmic Dawn

Cosmic Dawn, loosely defined here to be the first billion years of cosmic time, is an ever-intriguing era that witnessed the formation of the first generations of galaxies. Toward the end of it there was also the last major phase transition of our Universe, the epoch of reionization (EoR), which is believed to be driven by the hydrogen-ionizing background emerged from the early galaxies formed. In this talk, I will explain how Cosmic Dawn becomes a real exciting epoch for unveiling the physics of galaxy formation thanks to the James Webb Space Telescope (JWST), as well as several forthcoming facilities such as SPHEREx, Roman Space Telescope, Square Kilometer Array, and LiteBIRD focusing on the large-scale effects. I will discuss the theoretical landscape galaxy formation at Cosmic Dawn informed by new JWST observations, with a particular focus on the phenomenon of bursty star formation. I will introduce methods and ideas to shed light on different aspects of early galaxy formation, including the star formation history, stellar feedback, outflows, and the ionizing output, using both individual galaxies and their effects on the large-scale structure and cosmic background radiations. With a few case studies, I will demonstrate how to harness the power of the aforementioned facilities and their synergies for these purposes.

Everyone is welcome to attend (undergraduate and graduate students, staff, and faculty at the physics department). This big event is part of our efforts to foster a welcoming work environment and solidify our physics community. 

Monica Vidaurri, Stanford

Title and abstract TBA

Prof. Mingda Li, Nuclear Science and Engineering, MIT

Exploring Potential Roles of Machine Learning in Quantum Materials Research

In recent years, machine learning has achieved great success in chemistry and materials science, but quantum materials face unique challenges. These include the scarcity of data (volume challenge), high dimensionality and computational costs (complexity challenge), elusive experimental signatures (experimental challenge), and unreliable ground truth (validation challenge).

In this Physics Colloquium, we present our recent efforts to support the study of quantum materials with machine learning. For scenarios with high data volumes, such as density-functional-theory (DFT) level studies with weak correlation, machine learning can predict lower-dimensional properties. We introduce a convolutional neural network classifier predicting band topology class based on X-ray absorption (XAS) signals [1]. This approach can also be applied to experimental data, demonstrated by an autoencoder-based protocol to study the magnetic proximity effect with polarized neutron reflectometry, improving fitting resolution [2].

For lower data volumes due to higher computational costs, incorporating symmetry into neural networks can reduce data volume needs. Using the O(3) Euclidean neural network, we predict phonon density-of-states [3], dielectric functions [4], and quantum weight [5] directly from crystal structures. Machine learning without data can also be performed by using differential equations as constraints [5].

For high output dimensions and low input data volumes, such as phonon dispersion relations, we introduce additional approaches like virtual nodes in a graph neural network [6], showing improved efficiency compared to machine-learning potential without losing accuracy.

To address unreliable ground truth, we use machine learning to distinguish Majorana zero modes in scanning tunneling spectroscopy for topological quantum computation [7]. For cases like quantum spin liquids, where experimental signatures are unclear and computational costs are high, we generate materials with potential geometrical frustration. Our latest work, SCIGEN, produces eight million materials belonging to Archimedean lattices, with over 50% passing DFT stability checks after pre-screening [8].

Despite progress, applying machine learning to quantum materials is still in its infancy. We reflect on the out-of-distribution problem, aiming to generate genuine surprises and new features rather than merely recognizing patterns. Additionally, we must address accuracy limitations in many machine learning approaches, especially with complex quantum systems and phase diagram studies.

[1] “Machine learning spectral indicators of topology,” Advanced Materials 34, 202204113 (2022).

[2] “Elucidating proximity magnetism through polarized neutron reflectometry and machine learning,” Applied Physics Review 9, 011421 (2022).

[3] “Direct prediction of phonon density of states with Euclidean neural networks,” Advanced Science 8, 2004214 (2021).

[4] “Ensemble-Embedding Graph Neural Network for Direct Prediction of Optical Spectra from Crystal Structure,” arXiv:2406.16654.

[5] “Panoramic mapping of phonon transport from ultrafast electron diffraction and machine learning,” Advanced Materials 35, 2206997 (2023).

[6] “Virtual Node Graph Neural Network for Full Phonon Prediction,” Nature Computational Science 4, 522 (2024).

[7] “Machine Learning Detection of Majorana Zero Modes from Zero Bias Peak Measurements,” Matter 7, 2507 (2024).

[8] “Structural Constraint Integration in Generative Model for Discovery of Quantum Material Candidates,” arXiv:2407.04557.