HOME
RESEARCH
EDGE OF THE WILD
Photon-Induced Surface Reactions
The interstellar medium is populated with large amounts of micron-sized dust grains composed of silicaceous and carbonaceous material. In cold sources, the low temperature causes gaseous compounds to accrete onto the grain surfaces, forming rich layers of mostly water, carbon monoxide, carbon dioxide, and trace amounts of organic molecules. During cloud collapse, these grains and icy mantles warm up, allowing efficient diffusion for increased reactivity and desorption back into the gas. For many years, this explained the increased complexity observed in warm sources, but recent detections have shown a large amount of molecular complexity on cold clouds pre-collapse. To illustrate this phenomenon, I used a large chemical kinetic solver, Nautilus, to add an experimentally observed surface reaction induced by radiating the surface with ultraviolet photons. The photons cause the photodesorption of small molecules such as molecular oxygen and carbon dioxide and can produce oxygen atoms in the electronically excited 1D state. These atoms possess increased reactivity compared to their ground 3P state, allowing immediate reactions with neighbors on or in the ice mantle without prohibiting reactive barriers. By including such reactions with common hydrocarbons (methane, ethane, ethylene, and acetylene), it produces oxygen-bearing organics that can then reactively desorb from the surface and into the gas for detection.
Reaction Prediction with Deep Learning Techniques
Traditional chemical modeling for astrochemical purposes involves utilizing experimental results, quantum chemical calculations, or chemical intuition. While using quantum chemistry and realistic experimental data is preferred, both require specialized knowledge and large amounts of time to complete. Therefore, chemical intuition has become a standard in adding new reactions in chemical networks, typically based on similar reactions in the data. This method should involve a certain level of experience to make knowledgeable additions and can be inherently biased depending on the purpose of the proposed reactions. And with an increasing need for more significant additions to networks due to more molecules being detected, these additions are a significant time crunch. With the emergence of powerful machine learning tools, we are instead trying to use AI to make these predictions for us. The task of reaction prediction is helpful in the context of organic chemistry with the idea of finding novel drug synthesis pathways and, therefore, should be able to show similar performance in the context of astrochemical synthetic pathways. We currently utilize the Transformer architecture to complete this task, similar to machine translation, where the model is given a pair of reactants and asked to predict the different product channels.