Subseasonal timescales, spanning approximately 2 weeks to 2 months, provide actionable lead times for various sectors such as energy and water management. While skillful forecasts on these timescales are societally important, subseasonal forecasting remains a difficult task. Previous research has shown that specific states of the climate system can lead to enhanced subseasonal predictability (e.g. state-dependent predictability). However, the presence of biases in Earth system models can affect the representation of these states and their subsequent impact. Here, we present a machine learning framework to identify state-dependent biases in Earth system models. In particular, we use historical simulations from the Community Earth System Model version 2 (CESM2) to investigate the utility of explainable neural networks in tandem with transfer learning to identify tropical state-dependent biases that are relevant for midlatitude subseasonal predictability.