Physical reservoir computing leverages the intrinsic history-dependence and nonlinearity of hardware to encode spatiotemporal signals directly at the sensor level, enabling low-latency processing of dynamic inputs. Encoding fidelity depends on the separability of multi-state outputs, yet in practice it is often hampered by empirically chosen, suboptimal operating conditions. Here, we apply Bayesian optimization to improve the encoding performance of solution-processed Al₂O₃/In₂O₃ thin-film transistors. By exploring a five-dimensional pulse-parameter input space and using the normalized degree of separation for output state distinguishability, we demonstrate high-fidelity 6-bit temporal encoding corresponding to 64 output states. We further show that a model based on simpler 4-bit data can effectively guide optimization for more complex 6-bit tasks, substantially reducing experimental effort. Using a six-frame moving-car image sequence as a benchmark, we find that the optimized 6-bit pulse conditions significantly enhance encoding accuracy, with 4-bit derived parameters performing comparably in terms of pixel errors. Shapley Additive Explanations (SHAP) analysis further reveals that gate-pulse amplitude and drain voltage are the dominant contributors to output state separation. This work establishes a data-driven strategy for identifying optimal operating conditions in reservoir devices and outlines a framework that can be transferred to diverse material platforms and physical reservoir implementations.
Meza-Arroyo et al. (Tue,) studied this question.