Accelerated Science Intern, Representation Learning for Molecules and Materials

Toyota Research Institute•Los Altos, CA

10d•$45 - $65•Hybrid

About The Position

Toyota Research Institute (TRI) is seeking a researcher or engineer to study how machine learning models represent molecules and materials, and whether different models discover the same underlying physical structure. The intern will collaborate with TRI researchers to investigate how training data, learning objectives, and model architectures shape learned representations. The work combines mathematical reasoning and computational experiments to examine when models arrive at shared representations, what information those representations capture, and how they relate to the physics of the systems being studied. The position is a paid, 12-week, hybrid internship. The intern will contribute well-documented code and help translate research results into internal reports, publications, and/or patent disclosures as appropriate. TRI’s Accelerated Materials Design and Discovery program aims to accelerate the development of emissions-free mobility by combining computational materials modeling, experimental data, artificial intelligence, and automation to advance materials research.

Requirements

Currently enrolled in a doctoral program in materials science, computer science, statistics, applied math, machine learning, physics, or a related discipline.
Strong foundations in machine learning, linear algebra, probability, and optimization, with demonstrated experience training models on real datasets.
Proficiency in Python and modern machine learning tooling, such as PyTorch, experiment tracking, and reproducibility best practices.
Ability to work collaboratively across disciplines and translate research ideas into working code.

Nice To Haves

Familiarity with representation learning or scientific machine learning.
Experience studying model convergence, learning dynamics, symmetries, or identifiability.

Responsibilities

Implement and evaluate machine learning models for molecules and materials using reproducible training and analysis workflows.
Design controlled experiments to study how learned representations vary across training runs, model architectures, and learning objectives.
Investigate the geometry and information content of learned representations and their relationship to physical quantities.
Develop evaluation methods to distinguish meaningful shared structure from similarities introduced by model parameterization or analysis choices.
Contribute high-quality, well-documented code to an internal codebase and help translate results into internal reports, publications, and/or patent disclosures as appropriate.