Efficient discovery and property optimization of quantum materials via reinforcement-inspired active learning: a case...

High-throughput discovery of functional MXenes requires efficient strategies to identify promising candidates within vast compositional spaces. In this work, we develop a reinforcement-inspired active learning framework to accelerate the search for MXenes with desirable electronic properties. Using a curated dataset of 4448 MXenes, including 946 semiconducting compositions with PBE0 hybrid-functional band gaps, we demonstrate two complementary tasks: (i) identifying semiconductors within a targeted bandgap window (1.65-3.35 eV) corresponding approximately to the visible region relevant for photovoltaic applications, and (ii) identifying the maximum-bandgap material within the semiconducting subset as a materials optimization problem. These tasks reflect two core objectives in high-throughput materials discovery. Task (i) corresponds to practical screening for device-compatible bandgaps (e.g., photovoltaics and optoelectronics), while task (ii) captures optimization-driven discovery, where the goal is to identify the best-performing semiconductor in a large library using as few expensive evaluations as possible. For the first task, we initialize with only 1% of the data as labeled input and apply a suite of exploration-exploitation policies, including ε-greedy, uncertainty sampling, entropy, upper confidence bound (UCB), Thompson sampling, and diversity-augmented strategies. Remarkably, with a budget of only 500 queries, the framework retrieves 130 of the 150 possible target MXenes, substantially outperforming random selection. For the second task, we restrict to 946 semiconductors and employ batch-mode UCB active learning to efficiently identify the material with the maximum bandgap. Starting with 2% labeled data and examining only 60 candidates in total, the algorithm recovers the top-bandgap MXene within six iterations across multiple batch sizes. The proposed framework provides a generalizable approach for sequential materials discovery and optimization while substantially reducing the number of expensive hybrid-functional DFT evaluations required for large-scale materials screening.