Quantum Computing and Loophole-Robust Certification

Our research examines how quantum computers can demonstrate genuine advantages over classical methods and how such claims can be tested against hidden assumptions. We develop practical ways to ask not only whether a quantum device outperforms a classical benchmark, but also whether that advantage would survive if the classical model were given additional information about the test itself. This approach extends the logic of loophole-free Bell tests to quantum computing, distributed information processing, and quantum machine learning.

We also study the preparation nonstationarity loophole (https://arxiv.org/pdf/2601.08290), which arises when the physical state prepared by an experiment changes over time. If data collected at different times are assumed to come from the same source when they do not, apparent quantum correlations or violations can be exaggerated or misinterpreted. By identifying and quantifying such effects, we aim to make experimental claims in quantum science more reliable, reproducible, and resistant to alternative classical explanations.

Machine Learning for Materials and Human Health

Our machine-learning research combines physical insight, experimental data, and uncertainty-aware modeling to address problems in materials discovery, thermal management, and personalized health. A central goal is to build models that are not only accurate, but also aware of when a new sample differs from the data used for training.

For solid-state electrolytes, we are developing a distribution-aware workflow that couples machine-learned interatomic potentials with Phonopy calculations. This approach enables rapid prediction of crystal vibrations and phonon-derived descriptors across large materials datasets, while identifying compounds that lie outside the model’s reliable domain. The resulting framework is designed to uncover relationships between lattice dynamics, structure, and lithium-ion conductivity and to accelerate the screening of candidate solid electrolytes.

We also use physics-informed machine learning to study heat transport. By combining infrared thermography with physically motivated image features, our models estimate the thermal conductivity of polymer-composite thermal-interface materials and help connect spatial heat-flow patterns with material performance. (American Chemical Society Publications)

In digital health, we integrate continuous glucose monitoring, smartwatch-derived heart rate, activity and sleep, and real-time self-reports to study personalized signatures of happiness and stress. Rather than seeking a single model for everyone, this work emphasizes within-person learning and shows that physiological–affective relationships are heterogeneous: stress can be more predictable than happiness for some individuals, but wearable and glucose signals do not provide a universal affect predictor. (ResearchGate)

Across these projects, we use machine learning as a scientific tool for discovering interpretable relationships, quantifying uncertainty, and guiding experiments rather than as a purely predictive black box.