Scalable Discovery of Fundamental Physical Laws
While traditional machine learning algorithms have excelled in classifying phenomena and uncovering empirical relationships, they have struggled to derive generalizable dynamical models directly from complex data. We have developed a scalable framework that addresses this gap, enabling the discovery of governing equations from high-dimensional, intricate datasets across a wide range of scientific disciplines. By combining novel advanced sparse regression techniques with an extensive library of candidate terms, we accurately recover the underlying dynamics of complex systems, including subtle effects critical to their behavior. This approach establishes a powerful, computationally efficient tool for uncovering fundamental laws in physics and beyond, driving innovation in data-driven scientific discovery.
Accelerating Time-Domain Simulations
The pace of accelerating simulation algorithms has slowed, partly due to the stagnation of improvements in traditional computer architectures and the limitations of single-thread performance. Machine learning offers the potential to drive the next major leap in performance. In black-hole astrophysics, one of the most significant computational bottlenecks is the calculation of ray tracing and radiative transfer, particularly when incorporating general relativistic (GR) effects. We are developing novel ML-driven algorithms that overcome these limitations, enabling faster and more efficient simulations without compromising accuracy. These advancements hold promise for transforming the landscape of time-domain simulations in astrophysics and beyond.
Advancing Black Hole Imaging with Novel Algorithms
We are developing groundbreaking machine learning algorithms for interferometric black hole imaging that emphasize interpretability and robustness. Using the Event Horizon Telescope (EHT) data, the PRIMO algorithm applies dictionary-learning principles, trained on a library of high-fidelity black hole simulations, to reconstruct images with unprecedented detail. This approach has sharpened the M87 black hole image, revealing a thinner, brighter ring. Additionally, the kernel-based KRISP algorithm offers robust, data-driven reconstructions, operating without prior training and excelling in sparse array configurations. By addressing challenges in sparse Fourier coverage and angular resolution, these advancements set a new standard for high-fidelity black hole imaging and deepen our understanding of these enigmatic objects.