Adaptive Memory and In Materia Reinforcement Learning Enabled by Flexoelectric-like Response from Ultrathin HfO₂

Reinforcement learning (RL) is a mathematical framework of neural learning by trial and error that revolutionized the field of artificial intelligence. However, until now, RL has been implemented in algorithms with the compatibly of traditional complementary metal-oxide-semiconductor-based von Neumann digital platforms, which thus limits performance in terms of latency, fault tolerance, and robustness. Here, we demonstrate that nanocolumnar (∼12 nm) HfO₂ structures can be used as building blocks to conduct the RL within the material by combining its stress-adjustable charge transport and memory functions. Specifically, HfO₂ nanostructures grown by the sputtering method exhibit self-assembled vertical nanocolumnar structures that generate voltage depending on the impact of stress under self-biased conditions. The observed results are attributed to the flexoelectric-like response of HfO₂. Further, multilevel current (>10^-3 A) modulation with touch and controlled suspension (∼10^-12 A) with a short electric pulse (100 ms) were demonstrated, yielding a proof-of-concept memory with an on/off ratio greater than 10⁹. Utilizing multipattern dynamic memory and tactile sensing, RL was implemented to successfully solve a maze game using an array of 6 × 4. This work could pave the way to do RL within materials for a variety of applications such as memory storage, neuromorphic sensors, smart robots, and human-machine interaction systems.