Integrating High-Throughput Experiments with Machine Learning Models for Macromolecule-Based Nanomaterials—A Case Study in DNA-Stabilized Metal Clusters

While data science is revolutionizing the study and design of hard materials, the promise of these approaches for soft and polymeric materials is manifesting more slowly. A Materials Genome approach to polymeric materials is challenged in part by the greater degree to which environmental conditions affect the salient structures and properties of soft matter as compared to hard/crystalline matter. From an experimental point of view, the dynamic and sensitive nature of soft materials makes it more challenging to amass sufficiently large and well-controlled experimental data sets to enable machine learning and data mining approaches. This talk will present a case study in merging highly-controlled experimental synthesis and characterization with machine learning strategies for the study and design of macromolecule-scaffolded nanomaterials. Our research focuses on a fluorescent nanomaterial with “genomic” properties: DNA-stabilized silver nanoclusters (Ag_N-DNAs) with atomic sizes and fluorescence colors selected by DNA sequence. These tiny fluorophores are finding applications in sensing and bioimaging, but it remains complex to understand how DNA ligands sculpt silver clusters. To answer this question and enable informed design of these macromolecule-based nanoclusters, we have developed high-throughput experimental synthesis and optical characterization of Ag_N-DNAs in order to establish a well-controlled data library that connects the optical properties of thousands of Ag_N-DNAs to the sequences of their DNA templates. Through a combination of statistical sampling and machine learning classification, we can extract the features of DNA oligomers that are discriminative for the fluorescence colors of Ag_N-DNAs, providing fundamental insights into how these macromolecules select the atomic sizes and shapes of few-atom silver clusters. The approach we have developed can be expanded to numerous sequence-encoded materials and provides a roadmap for developing machine learning approaches that exploit experimental data libraries to understand and design soft and polymeric materials.<br/>This work is supported by NSF-CBET- 2025790.