Monolithically Integrated Flexible Thermoelectric Encoder: Toward Artificial Intelligence-Enhanced Intent Recognition

Achieving intelligent human-computer interaction requires interfaces that are both highly integrated and reliable, capable of capturing subtle physiological and behavioral information from users. Flexible thermoelectric devices offer distinct advantages for low-power interactive applications, yet progress in system-level integration remains limited. Here, we report a monolithic integration strategy that combines thermopile arrays with analog-to-digital conversion circuits. By embedding the electrodes of flexible thermoelectric devices directly into the circuit layout, a fully integrated thermoelectric encoder is fabricated on a single polyimide substrate, consisting of four functional units. At an ambient temperature of 20 °C and a hot-side temperature of 35 °C, the device exhibits an open-circuit voltage of 21.2 mV and withstands over one million cycles of pressing and bending with less than 5% change in internal resistance. Using this encoder, three interaction states—hard press (∼1 N), light touch (∼0.3 N), and no contact—are distinguished, enabling a simple yet complete coding scheme that covers all 26 letters. To ensure robust operation across varying environmental conditions, a lightweight Tri-stream Spatiotemporal Fusion Network is introduced, with only 29292 parameters and a size of 114.4 kB. By integrating depthwise separable convolutions, bidirectional long short-term memory, and dynamic channel attention, this model achieves letter recognition accuracy exceeding 98%. Finally, the thermoelectric encoder is combined with an AI agent to demonstrate an intelligent interaction system for patient monitoring, supporting both daily assistance and emergency response.