In this proposal we describe a technical approach to detect never-before-seen chemicals with existing chemical detection equipment. First, we will generate never-before-seen, synthetically viable, toxic chemicals. To do this we will use open-source software that enables synthetically viable chemicals to be generated as SMILES strings from a seed chemical, e.g., VX, RDX, etc. For each novel chemical, we will estimate toxicity and convert the SMILES string to a 3D molecular structure using open-source software. Next, we will use quantum chemistry software to perform a geometry optimization and calculate the theoretical infrared (IR) spectrum for each chemical. Concurrent with the computation of theoretical IR spectra, we will obtain experimental IR spectra for TICs, explosives, CWAs, etc. from online resources. For all chemicals with experimental spectra, we will compute theoretical IR spectra. Using the dataset of experimental and theoretical spectra we will train a machine learning algorithm (MLA) to map experimental to theoretical IR spectra. We anticipate the MLA will learn to remove the effects of complex chemical phenomena such as solvation, anharmonicity, etc., as it develops a mapping from the experimental spectra to the simpler theoretical spectra. During deployment the MLA will process experimental spectra to identify known and never-before-seen chemicals. For Phase I, we focus on IR spectra, but note the approach should generalize well to other types of chemical spectra. If successful, the proposed technology would find many applications within the Government, industry, and academia, for example in drug discovery, materials research, and remote sensing.
