Terrence Oscar-Okpala

Terrence Oscar-Okpala

Biography

Terence Oscar-Okpala is an undergraduate at Bethune-Cookman University studyingChemistry and Mathematics. He is driven by a thirst for knowledge in the field of physical sciences and a desire to understand the world around him. At his home institution, he engaged in research with his mentor, Dr. Vishwa Trivedi, focusing on investigating the rhodops insignaling pathway by characterizing the structure of Anabaena Sensory Rhodopsin and its transducer protein. At The Scripps Research Institute, under the mentorship of Dr. Ashok Deniz, he investigated the effect of charge and valency on phase separation in apolyuridine-polylysine system to understand protein-RNA interactions. In the Kulik lab at theMassachusetts Institute of Technology, he investigated the effect of local electric fields on the selectivity and reactivity of alpha-ketoglutarate-dependent non-heme iron enzymes. Throughthese experiences, he has developed a passion for utilizing computational techniques and the concepts of physical chemistry to understand how molecules form in outer space.

Abstract

Leveraging Optical Simulations and Machine Learning to Emulate Real-World Interferometers

Gravitational wave (GW) observatories, such as Advanced LIGO, rely on sensitivity and precision for interferometric measurements. However, optical misalignments and noise limit the performance and sensitivity of these detectors, which rely on linear control schemes and constant human intervention to maintain optimal correction. AI-assisted sensing and control using reinforcement learning (RL) has emerged as an alternative, with success when applied to GW detectors. An experimental setup, TestbedAI, at the MIT LIGO laboratory, facilitates the testing of automated and intelligent optical control problems in interferometry experiments. However, existing RL algorithms are sample inefficient, slowing down the training process. This project developed simulators that reproduce observed light fields within the TestbedAI Mach-Zehnder Interferometer. These simulations predict interference patterns by changing actuator signal parameters within the engine. A reduced-order model neural network will be trained using the output states and reward generated from the simulator and the current RL model to optimize the prediction of interference patterns for a wide range of actuator signal changes. The neural network will be integrated into the RL algorithm and fine-tuned on experimental outputs to compensate for external noise, such as ghost beams and actuator hysteresis, to bridge the simulation-to-reality gap between simulated and real-world interferometers.