A Structurally Robust Framework for Intelligent Graphene Thermometry via Few-Shot Transfer Learning and Algorithm-Hardware Co-Design.

While ultrathin graphene sensors hold promise for flexible electronics, their applicability faces a challenging arising from the unavoidable nanoscale stochasticity and structural variability of compliant substrates. Features like intrinsic grain boundaries and microscopic wrinkles inevitably create pronounced device-to-device heterogeneity, rendering traditional batch calibration unreliable for high-precision applications. Rather than eliminating these inherent physical imperfections, we introduce a variability-resilient sensing framework built on an algorithm-hardware co-design. By employing a few-shot transfer learning architecture with a frozen-backbone neural network, our system effectively learns the universal physics of graphene carrier scattering from a source array and can rapidly adapt to the unique electrical footprint of new, uncalibrated devices using less than 1% of conventional calibration data (R2 > 0.99). This framework not only enables high-precision graphene-based flexible thermometry for the first time (±0.2°C), but also supports real-time temperature inference across wide-range robotic perception in harsh environments up to 200°C. Furthermore, we demonstrate that graphene's high intrinsic sensitivity remains functionally robust despite structural discrepancies, as a 1D convolutional neural network decodes complex physiological waveforms directly from raw signals with 94.95% accuracy. Collectively, this sensing-as-inference paradigm transforms graphene into a viable, calibration-efficient material for intelligent wearable thermometry.