(4be) Prediction of Molecular Properties of Porphyrins Using Machine Learning with Database Screening Using Kernel Similarity

Porphyrin molecules play an important role in many biological and chemical processes such as solar energy conversion, photo-sensing, photomedicine, photocatalysis etc. Several of these applications involves proper tuning of the band gap and orbital energy levels of these molecules to obtain the desired results. The band gaps are usually tuned either by changing the central metal atom, or by changing the side groups, or by the combination of both. Computational studies have been extensively used to estimate these electronic properties before they are synthesized. However, because of the large size of these molecules, computational studies are usually expensive for large databases. In order to address this challenge, we devised a kernel similarity based machine learning method to predict the orbital energies of these molecules much more quickly and with reasonable accuracy. Most approaches to increase the prediction accuracy involves increasing the complexity of the representations. However, this will increase computational demand, and further slows down the process. Here, we used an efficient way to screen molecules in the database based on kernel similarity. This approach will reduce the dataset used for training based on the similarity of its structure with the target molecule, and thus reducing the computational resources required for training the model. This also enhances prediction accuracy because the algorithm only selects molecules based on its similarity. Among the various methods used for cut-off, we found that a similarity cut-off of around 0.9 was found to be the best in terms of accuracy.

Research Interests

My research primarily focuses on developing and applying computational tools, including machine learning, to study and develop advanced materials for applications such as energy storage, photocatalysis, organocatalysis, water treatment, and drug development. During my Ph.D., I modeled energy storage materials using reactive molecular dynamics (ReaxFF) and developed a novel Monte Carlo-based method for training reactive force fields. My post-doctoral work involved using dissipative particle dynamics (DPD) to study polycarbonate-based polyurethane networks. As an assistant professor, I continue to explore molecular simulations and have initiated research collaborations with experimentalists in the areas of organic synthesis, energy research, and drug design using various computational tools. Current projects include using machine learning to predict molecular properties, enhancing photocatalytic activity with graphitic carbon nitride (g-C_3N_4) 2D layers, studying organocatalysis mechanisms, and collaborating on biomolecular systems research. With nearly 30 research articles and around 500 citations, I have been steadily building my research profile.

Teaching Profile

With almost 8 years of experience teaching undergraduate chemical engineering students, I have taught nearly all courses in the field. I am passionate about interacting with students and enjoy the process of transferring knowledge to the next generation. I have developed strategies to make complicated topics easy and fun to learn, achieving more than a 90% approval rating from students. My teaching methodology is practical, ensuring strong teacher-student interaction during lectures. My techniques to improve student learnability include:

• Encouraging students to ask questions.
• Promoting peer discussions and learning.
• Conducting lab sessions to explain concepts in detail.
• Introducing the latest software (such as Python, ANSYS, MATLAB and ASPEN) and designing assignments to integrate these tools with theoretical learning.
• Focusing on average and below-average students to ensure their understanding.
• Meeting with struggling students outside class to discuss their issues and providing additional support.
• Implementing a peer review system where good students earn points for helping weaker students, thereby improving learning outcomes for both groups.