Continuous Perception for Immersive Interaction and Computation in Molecular Sciences

Chemistry aims to understand the structure and reactions of molecules, which involve phenomena occurring at microscopic scales. However, scientists perceive the world at macroscopic scales, making it difficult to study complex molecular objects. Graphical representations, such as structural formulas, were developed to bridge this gap and aid in understanding. The advent of Quantum Mechanics further increased the complexity of the representation of microscopic objects. This dichotomy between conceptual representation and predictive quantification forms the foundation of Chemistry, now further explored with the rise of Artificial Intelligence. Recent advancements in computational sciences, increased computational power, and developments in Machine-Learning (ML) raise questions about the traditional scientific method. Computational scientists, who have relied on approximations based on fundamental rules, now face the possibility of accurately simulating nature without strictly adhering to its laws. This shift challenges the association between progress in understanding a phenomenon and the ability to predict it. Deep learning models can not only make predictions but also create new data. While these techniques find applications in fields like Natural Language Processing, they suffer from limitations and lack true intelligence or awareness of physical laws. The thesis aims to create mathematical descriptors for atom types, bond types, and angle types in ML procedures, ensuring the retention of their chemical meaning. The goal is to make quantitative predictions while interpreting changes in descriptors as chemical changes. To achieve this, the thesis develops a software called Proxima for Molecular Perception, which automatically perceives features from molecules. Proxima treats strongly coupled electrons as covalent bonds and lone pairs, while delocalized electrons are modeled using a Tight-Binding model. The resulting Molecular Graph captures the weak interactions between these units. Overall, this thesis explores the intersection of computational chemistry and Machine-Learning to enhance our understanding and predictive capabilities in the field of Chemistry by building the so-called Virtual Laboratory, a virtual environment with automatic access to structural databases to test chemical ideas on the fly (pre-processing) and explore the output of computational software (post-processing).