Abstract
<jats:title>Abstract</jats:title> <jats:p> Controlling specific neuronal dynamics with electrical stimulation is critical for therapeutic neuromodulation, yet deriving optimal control policies remains challenging due to the complex and non-stationary nature of biological neuronal networks. While reinforcement learning (RL) offers a powerful closed-loop control framework, its reliance on prolonged stimulus-driven exploration is difficult to reconcile with the physiological limits of living tissue. Here, we demonstrate an <jats:italic>in silico</jats:italic> -to- <jats:italic>in vitro</jats:italic> transfer strategy that achieves efficient state-dependent control of network bursting in cultured neurons. The transferred policy outperforms heuristic controls, while maintaining constrained stimulation usage. Concurrent calcium imaging reveals the mechanistic basis of the learned policy: the agent optimizes stimulation spatially and temporally, exploiting local network topology and intrinsic physiological temporal dynamics. These results establish <jats:italic>in vitro</jats:italic> brain-on-chip cultures as a tractable stepping stone for RL-based neuromodulation and demonstrate that effective control policies can be derived in biophysically calibrated digital twins and transferred directly to living networks. </jats:p>