A Scientific Programmer
Deep Learning
Scientific Machine Learning
Computational Physics
Applied Mathematics
Through my research, I implemented a new technique to denoise images of the moon's most mysterious terrains—the Permanently Shadowed Regions (PSRs) at the lunar poles. These areas, located within deep,
ragged craters at the extreme north and south of the moon, have never seen sunlight, remaining dark for billions of years. The utterly lightless conditions make imaging these regions impossible, even
with very high spatial resolution cameras onboard the orbiters. The data gathered by the Orbiter High-Resolution Camera (OHRC) on ISRO's Chandrayaan-2 orbiter presented a unique
opportunity to illuminate the unknown using transformative algorithms, given its extremely high spatial resolution of 0.25 m/pixel. We developed a novel
computational framework leveraging Denoising Diffusion Probabilistic Models (DDPMs) and reference-less contrast adjustment to process the
CH-2 observations of the PSRs. Despite the complete absence of light, my approach enabled imaging of surface features within the shadows for
the first time through unsupervised reconstructions of those regions using simultaneous usage of pre-trained Spatial and Frequency Models.
The breakthrough results provided remarkable new details about the topography and composition of these previously hidden terrains, which
could contain troves of frozen water. Achieving almost State-Of-The-Art denoising performance of the moon's permanently shadowed craters demonstrates
the huge potential of AI in space exploration and planetary science. The insights unlocked by my technique will help guide future lunar missions, especially Chandrayaan-3,
while advancing our understanding of the moon's geology.
📄 Paper Draft (in progress) | 📮 IPSC 2023 Poster | 👉 Extra-Sample-II | 👉 Extra-Sample-I
My research focused on optimizing the thermal parameters of High Power Single Quantum Well (SQW) lasers to improve output lasing performance through precise modeling. SQW lasers are semiconductor devices that rely on complex interactions between temperature, current density, and other constrained parameters. However, finding the optimal configuration is incredibly challenging due to the multidimensional complexity. To tackle this, I devised a unique approach by reframing the multivariate maximization as a physics-based genetic programming problem. I achieved robust modeling of thermal properties and prediction of optimal designs to maximize the lasing output given a certain wavelength. The work demonstrated that combining constraint optimization, sequential programming, and evolutionary algorithms can provide pathways to effectively model even highly complex systems like SQW lasers. My specialized techniques resulted in accurate modeling of thermal dynamics within the quantum well structure. This research proves the power of merging computational optimization, machine learning, and physics-based techniques to advance discoveries even in convoluted multidimensional systems.
This project aimed to detect gravitational waves in noisy data from three Gravitational Wave Observatories (LIGO Hanford, LIGO Livingston, Virgo). Gravitational wave signals are extremely faint within the detectors' time-series output. To amplify the patterns, I implemented the constant Q transform and Butterworth bandpass filtering as preprocessing steps. The resulting spectrograms offered visual representations of the data's frequency composition over time. However, pinpointing the presence of gravitational waves remained challenging due to structural noise and interference. To address this, I used a CNN architecture that classified spectrogram inputs, identifying those containing gravitational wave signals versus noise. The model detected patterns within the frequency information that confirmed the presence of the GWs. Through training on labeled LIGO data, the network learned to isolate the signals with over 90% accuracy despite a low signal-to-noise ratio. This work enables the rapid identification of new events as LIGO continues to survey gravitational disturbances.
This project focused on improving short and long-term price prediction for cryptocurrency assets through an integrated machine learning approach. Modeling cryptocurrency value trends poses challenges due to high volatility and complex factors influencing prices. To address this, I implemented a hybrid model combining Dense 1D CNN modules with Recurrent Bayesian Neural Networks. The CNN extracted local features and patterns from time series data, while the RBNN identified longer-term correlations and uncertainties. I trained this integrated network on historical price data for BTC-USD pairs. Comparative analysis with established models like Local and Semi-Local Linear Trend demonstrated significant improvements in predicting both near-term fluctuations and overall price trajectories. The hybrid model achieved high directional accuracy across lengthy backtesting. Cryptoasset values have vast implications across financial, technological, and social domains.
This current project focuses on developing a deep learning framework for the robust identification of airplane condensation trails in images. Condensation trails are important indicators for aviation-related climate impact studies, yet differentiating them from other cloud formations remains challenging. To improve detection, I am implementing attended variational autoencoders to learn efficient latent space representations of both segmented and unsegmented condensation trail images. This allows for precise clustering based on visual features. Additionally, I am implementing a deep energy-based model utilizing Langevin sampling to optimize latent displacement across clusters. This specialized framework integrates reconstruction, segmentation, and sampling to achieve state-of-the-art performance in condensation trail recognition. The aim is to overcome limitations in existing atmospheric studies by leveraging AI to unlock new insights. The work will help drive progress in aviation emission analysis and climate modeling through the robust identification of airplane exhaust trails.
