Accessing Creativity and Institutional Knowledge of Polymer Chemists via Expert-in-the-Loop AI—Case of Acrylic Polymer Design
With ever increasing demand of the modern world for the materials with advanced properties, we are facing the pressure to accelerate the discovery through a paradigmatic shift. One of the major practical bottlenecks – actionability bottleneck - is majority of the computationally generated hypotheses fail to transition to the experimental phase. In this case, the target properties are expected to meet project requirements, but the candidates fail to satisfy broader project-specific constraints. The most problematic constraints represent institutional knowledge, broadly defined. Institutional knowledge is not systematically captured in publications, but rather carried by subject matter experts (SMEs). This explains fundamental difficulty of capturing institutional knowledge, introducing it into the training phase of statistical modeling, and propagating into the hypothesis generation phase. This contribution addresses resolution of the actionability bottleneck via capture and instantiation of SME’s creativity and institutional knowledge in the form of AI models. This is accomplished using a two-tiered expert-in-the-loop (EITL) approach [1,2]. Specifically, we considered monomer design for acrylic polymers. The first tier of EITL AI included training a discriminator that was reproducing SME’s decision-making in the adjudication of the candidate monomers. The candidate monomers were generated using recently developed system for molecular inverse-design . The SME – a polymer chemist responsible for the polymerization experiments – accepted or rejected the candidates as they answered the question: “Is it practical to synthesize this monomer?” The question, as discussed with the SME, implied subjective assessment of the synthetic accessibility, expected behavior during polymerization, and novelty. The second tier of EITL AI leveraged the adjudication results and the trained discriminator to train a generative model producing monomer candidates that were likely to be accepted by the SME. We trained 2 sequential neural network models and 3 generative adversarial networks (GANs); generative models were refined via separate adjudication of their output until acceptable performance was reached. At this point the generative form of EITL AI combined implicit models of SME’s knowledge and creativity as both these components informed the adjudication processes. The trained generative component of EITL AI was queried by the SME to identify candidates for the experimental phase. The generated structures were evaluated on the basis of novelty and synthetic complexity. Selected structures were additionally assessed based on the following criteria: - the ease of synthesis or commercial availability for the monomer precursor - the number of transformations, their difficulty, commercial availability and price of starting materials and reagents - the ease of polymerizable handle installation on the monomer precursor - the ability of the obtained monomer to be polymerized in a free radical polymerization process Five monomers were selected to produce acrylic polymers. Performance metrics of the AI models, experimental procedures, and polymer characterization results will be discussed along with existing challenges and potential extensions.