Machine Learning & Data-Driven Methods

The machine-learning program in the lab combines modern data-driven methods with the physical models, systems theory, and information-theoretic tools developed in the group’s other areas. The emphasis is on methods that carry guarantees — sample-complexity bounds, physical constraints as inductive bias, and rigorous evaluation against systems-theoretic baselines — rather than purely empirical modeling.

Active threads include:

• Physics-augmented deep learning for classifying single-molecule force-spectroscopy data, connecting deep networks to the mechanical models used in the systems-biology program.
• Machine-learning approaches to topology change estimation in power distribution networks — a bridge between the energy and foundations programs.
• Reinforcement-learning-based vulnerability analysis of electricity markets under cyber-topology attack, contributing to the cyber-physical-security subthread in the energy area.
• Information-theoretically optimal sample complexity of learning dynamical DAGs — foundational bounds on how much data is needed to identify network structure.
• Non-intrusive load monitoring using hybrid classification-regression methods, scalable to high-frequency residential data.
• Support-vector-machine methods for preventing cascading failures in microgrids.

The area works most productively in combination with the others: a paper is often cross-tagged with energy or systems-biology, reflecting that the ML methods are developed to answer a substantive question in an applied domain, not for their own sake.