Autonomous materials discovery through a tabula rasa active learning framework

Developing high-dimensional multi-element alloy electrocatalysts is crucial for clean energy, yet the discovery process is fundamentally hindered by the combinatorial explosion of the vast chemical space. Traditional trial-and-error methodologies are prohibitively slow, while current active learning approaches heavily rely on human-engineered features and frequently fall into the exploitation trap of localized optima. Here, we show an autonomous materials discovery framework operating from a tabula rasa state by coupling a descriptor-free surrogate model (PC-BAN) with a neuroscience-inspired acquisition function (LCBDS) that enforces surprise-driven exploration. Within a design space of 6074 quaternary combinations, our framework autonomously identified the absolute global optimum catalyst exhibiting an exceptional overpotential of 206 mV using only 220 queries to an experimental oracle, substantially outperforming traditional baselines and human-guided searches. By mathematically formalizing curiosity without prior chemical biases, this work lays a scalable, descriptor-free algorithmic foundation for the next generation of prospective autonomous laboratories.