Density functional theory (DFT) calculations yield stable material structures and their properties, yet are extremely computationally intensive. In this paper, we show that equivariant graph neural networks can model the electron density distribution in a wide range of materials, amounting to a 27% speedup in DFT calculations. This work could be applied for faster discovery of materials used in solar cells, batteries, and other electronics.
Winner of Best Paper Award, REMIA Workshop, MICCAI 2022 Limited labeled data in the medical domain poses a challenge to training deep learning models for medical diagnostics tasks. In this paper, we address this lack of labeled data with a radiology report generation pretraining task. Not only does such a task learn chest X-ray representations for efficient downstream transfer, but it also provides interpretability to practicioners through textual output.
Small teams in the field (search and rescue, reconnaisance, etc.) can benefit from the situational awareness and data streams provided by small Uncrewed Aerial Systems (sUAS), yet the operation of such vehicles draws personell away from mission-critical tasks. In this paper, we develop modular autonomous capabilities for sUAS systems beyond what is available in off-the-shelf systems, with the goal of reducing burden on operators. We demonstrated a computer vision and radar-guided aerial chase of a rogue sUAS.
Radiology reports are an underused resource for training deep
learning algorithms for automated medical diagnostics. Leveraging paired chest x-rays and corresponding text reports, we present a
novel method for training models that extracts more useful information
from medical images to improve image analysis tasks.
Colorimetric sensing materials exhibit color-change in the presence of
environmental stimuli, acting as low-SWAP detectors for hazardous
chemicals. The addition of a finely tuned Fabry-Perot cavity to the
surface of chemically-active sensing materials is quantitatively shown
to increase color change, as perceived by the human eye, and a model
for optimization of structure is presented.