Low-probability events detection using unsupervised multi-prototype clustering for single-molecule electronics

Artificial intelligence for science (AI4S) has emerged as a new horizon in state-of-the-art scientific research, and single-molecule electronics could be considered an ideal prototype in AI4S due to the opportunities in correlating high-throughput and high-quality data with clear physical mechanisms. Towards using artificial intelligence for single-molecule electronics (AI4SME), the unsupervised extraction of low-probability events from the massive experimental data becomes the key step, which has emerged for accurate detection of different configurations and even structural changes in single-molecule junctions. However, the present algorithms suffer from the “uniform effect”, in which the majority events are erroneously allocated to minority ones, resulting in a relatively equal spread of cluster sizes and hindering the investigations for charge transport mechanisms with subtle and complex behaviors in single-molecule electronics. In this work, we propose a new multi-prototype clustering technique for precisely discriminating molecular events during the break junction process, especially those occurring with a probability below 10%, and further precisely extract the product species at the onset of the electric field-driven single-molecule keto-enol reaction with a probability as low as 1.5%. Our work tackles the long-term bottleneck of uniform effect for the precise detection of low-probability single-molecule events.