M. Tamasi, A.J. Gormley
Plexymer, Inc.,
United States
Keywords: machine learning, robotics, materials design, AI, polymers
Summary:
The vast chemical design space of polymers; spanning millions of potential structures and compositions, presents both extraordinary opportunities and formidable challenges for materials innovation. Traditional approaches rely on either rational design guided by expert intuition or systematic high-throughput screening, yet both face fundamental limitations. Rational design demands extensive knowledge to anticipate complex structure-property relationships and often overlooks non-intuitive chemical combinations. In contrast, experimental screening alone generates large datasets but struggles to efficiently navigate exponentially expanding search space without intelligent guidance. This creates a critical bottleneck: the chemical diversity of synthetic polymers remains largely unexplored, while applications demand tailored macromolecules with increasingly sophisticated performance profiles. The convergence of laboratory automation, advanced analytics, and artificial intelligence now enables a fundamentally different paradigm: closed-loop experimental systems where real-world data directly powers predictive models that iteratively guide subsequent experiments. We present an autonomous discovery platform that transforms polymer development from a linear, human-limited process into an accelerated, self-improving cycle. Our approach centers on a reprogrammed Hamilton Microlab STARlet liquid handling robot executing automated PET-RAFT polymerization (photoinduced electron/energy transfer reversible addition–fragmentation chain-transfer). This system enables combinatorial synthesis across 96-well plates and glass vial arrays, producing 100+ distinct polymer formulations simultaneously with precise control over composition, molecular weight, and architecture. Crucially, this infrastructure enables the creation of rich experimental datasets that can fuel reinforcement learning algorithms trained directly on physical experimental outcomes. Here, we demonstrate this paradigm by iterating designs for polymer chaperones that preserve enzyme catalytic activity under denaturing conditions, achieving high levels (>80%) of active enzyme recovery after thermal stress. Each design-build-test-learn cycle generates experimental training data, enabling Bayesian Optimization strategies to be deployed for accelerating subsequent designs. This methodology extends naturally to functional materials where complex property targets demand exploration of vast chemical libraries. By integrating robotics, advanced polymer chemistry, and reinforcement learning, we aim to compress development while systematically expanding accessible chemical space. These advances position autonomous materials labs as a catalyst for the next generation of accelerated R&D.