Portrait of Omar Alterkait

Omar Alterkait

PhD Candidate in Physics, Tufts University

IAIFI Junior Investigator · DOE SCGSR Fellow at SLAC

I build foundation models and differentiable simulators for particle physics, turning raw detector signals into 3D pictures of nature's rarest interactions.

Multi-Modal Foundation Models Differentiable Simulation Sparse 3D Learning ML for Science
CV

About

I work at the intersection of machine learning and neutrino physics. My research centers on sparse 3D inverse problems: reconstructing particle interactions from raw, sensor-level data in detectors like liquid argon time projection chambers and large optical detectors. To do that, I develop GPU-accelerated differentiable simulators in JAX, achieving 100× to 10,000× speedups over traditional Monte Carlo, and self-supervised multi-modal foundation models trained on datasets of more than 10 million events. Throughout, I care about what happens at the sensor level: signal processing on raw waveforms, fusing multiple readout modalities, and robustness to the domain shift between simulation and real detector data.

I am a PhD candidate at Tufts University advised by Taritree Wongjirad, an IAIFI Junior Investigator at MIT, and currently a DOE SCGSR Fellow at SLAC working with Kazuhiro Terao.

Research

JAXTPC: Differentiable simulation and foundation-model data for LArTPCs

with Kazuhiro Terao (SLAC)

A GPU-accelerated, differentiable liquid argon TPC simulator in JAX, modeling the full detector response chain: charge recombination, electron drift, diffusion, wire and pixel readout, electronics, noise, and digitization, for arbitrary multi-volume geometries from SBND to DUNE ND-LAr scale, with gradients flowing through every physics parameter. It powers production of a 10-million-event dataset pairing raw sensor waveforms with 3D ground truth, 30× larger than existing sets and capturing charge and light from the same events, as the substrate for multi-modal foundation models of particle reconstruction.

GitHub

LUCiD: End-to-end differentiable optical detector simulation

with César Jesús-Valls (CERN)

The first end-to-end differentiable simulator for optical particle detectors, unifying simulation, calibration, and tracking in a single JAX framework. Physics-informed SIREN networks stand in for GEANT4 photon generation, and gradients flow through ray tracing, scattering, and sensor response. LUCiD calibrates some twenty global optical parameters together with per-sensor efficiency maps across tens of thousands of PMTs, approaching the Cramér-Rao bound, and matches or surpasses conventional non-differentiable methods in accuracy and speed on water Cherenkov geometries from WCTE to Hyper-Kamiokande scale.

GitHub arXiv:2602.24129

Sparse Euclidean equivariant networks for particle reconstruction

with Tess Smidt (MIT)

Combining sparse convolutions with E(3) equivariant architectures so networks respect rotation, translation, and reflection symmetry by construction: no data augmentation and better sample efficiency through geometric inductive biases, targeted at particle reconstruction in the MicroBooNE detector.

News

Publications

These are highlighted works. For the complete and up-to-date list, see my Google Scholar profile.

Talks & Presentations