Kareem Hegazy
Kareem Hegazy
Postdoctoral Researcher · UC Berkeley · Lawrence Berkeley National Lab · ICSI
I am a postdoctoral researcher at UC Berkeley, Lawrence Berkeley National Lab, and ICSI, working with Michael Mahoney and Benjamin Erichson at the intersection of Physics, Chemistry, and AI. I received my Physics PhD from Stanford in 2023 under Phil Bucksbaum and Ryan Coffee, where I studied excited-state quantum molecular dynamics through ultrafast gas-phase diffraction. In summer 2019 I was an AI Research Fellow at Google X on the Blueshift team. I previously earned an Honors B.Sc. in Physics and Math from the University of Michigan, where I researched the Higgs Boson at CERN under Bing Zhou.
I am broadly interested in scientific machine learning (SciML) apply ML to time series dynamics — PDE modeling, control, inverse problems, and LLMs for time series.
Selected Publications
Neural Equilibria for Long-Term Prediction of Nonlinear Conservation Laws
Submitted: JCP arXiv
NeurDE — a 'neural twin' of Boltzmann-BGK kinetic theory that pairs exact lattice-Boltzmann transport with a learned equilibrium map — delivers data-efficient, stable long-horizon forecasts of shock propagation and complex compressible dynamics.
Recency Biased Causal Attention for Time-series Forecasting
Introduces a recency-biased causal attention mechanism that reweights Transformer attention with a smooth heavy-tailed decay — strengthens temporally local dependencies and achieves competitive or superior performance on time-series forecasting benchmarks.
The False Promise of Zero-Shot Super-Resolution in Machine-Learned Operators
Machine-learned operators cannot actually do zero-shot super-resolution — they're brittle and susceptible to aliasing off the training grid. A simple, data-driven multi-resolution training protocol restores accurate cross-resolution inference.
Foundation Models for Discovery and Exploration in Chemical Space
Submitted: Nature arXiv
A foundation-model approach to discovery and exploration in chemical space — unified molecular representations that generalize across diverse downstream tasks in chemistry and materials discovery.
Conformer-Specific Photochemistry Imaged in Real Space and Time
Directly images conformer-specific electrocyclic ring-opening of α-phellandrene with MeV ultrafast electron diffraction, confirming Woodward–Hoffmann rules in real space and time — a new tool for resolving conformer-dependent reaction dynamics.
The Photochemical Ring-Opening of 1,3-Cyclohexadiene Imaged by Ultrafast Electron Diffraction
Femtosecond, sub-ångström imaging of 1,3-cyclohexadiene's photochemical ring-opening with MeV ultrafast electron diffraction — a textbook electrocyclic reaction and model for photobiological reactions such as vitamin D synthesis.
Imaging CF3I Conical Intersection and Photodissociation Dynamics with Ultrafast Electron Diffraction
MeV ultrafast electron diffraction images the photodissociation of CF3I through a conical intersection — the first real-space mapping of a coherent nuclear wave packet bifurcating onto two potential energy surfaces as it passes through the intersection.
Applying Bayesian Inference and Deterministic Anisotropy to Retrieve the Molecular Structure from Gas-Phase Diffraction Experiments
Novel Bayesian statistical framework that transforms gas-phase diffraction into 3D molecular images — 100× resolution improvement, no simulation bottleneck.
Tracking Dissociation Pathways of Nitrobenzene via MeV Ultrafast Electron Diffraction
Resolving photoinduced dissociation pathways of nitrobenzene using MeV ultrafast electron diffraction.
Please see my scholar page for all my up-to-date publications.
Selected Blogs
Blog posts coming soon.