Impactful Weather is a research project aimed and downsampling the resolution of coarse precipitation data utilizing ML concepts.
Precipitation impacts society on a very granular scale, often changing with just 100 meters of distance. However, attaining precipitation data (which is useful for forecasting tropical cyclones, monitoring humidity, monitoring flood and dought conditions, and other important fields) can be a challenge, especially at such a detailed level. Modern statistical and dynamic downscalling methods are either to coarse or too computationally expensive to use practically. However, Machine Learning methods can accurately downscale coarse data, while simultaneously providing rapid compute times.
Concepts Used: Deep Learning, CNNs, Data Augmentation, MultiThreading
Technologies Used: Python, Tensorflow, Keras, Torch, AWS
The Catalyst Bubble Detection project aims to use ML to detect bubbles in high-throughput microscope images of catalyst surfaces.
Detecting bubbles from a catalyst surface to determine reaction rate allows for the determination of highly reactive surfaces. Statistical methods, like Hough Circle Transformations, prove to be inaccurate, and manually labeling data takes far too long. Deep Learning architectures, like Faster-RCNN and Mask-RCNN can be used to speed up inference while boosting accuracy.
Concepts Used: Deep Learning, RCNNs, Heavy Data Augmentation, Hyper-Parameter Optimization
Technologies Used: Python, PyTorch, TIMM, ACLF Theta