Abstract
Fully-printed temperature sensor arrays-based on a flexible substrate and featuring a high spatial-temperature resolution-are immensely advantageous across a host of disciplines. These range from healthcare, quality and environmental monitoring to emerging technologies, such as artificial skins in soft robotics. Other noteworthy applications extend to the fields of power electronics and microelectronics, particularly thermal management for multi-core processor chips. However, the scope of temperature sensors is currently hindered by costly and complex manufacturing processes. Meanwhile, printed versions are rife with challenges pertaining to array size and sensor density. In this paper, we present a passive matrix sensor design consisting of two separate silver electrodes that sandwich one layer of sensing material, composed of poly(3,4-ethylenedioxythiophene):polystyrene sulfonate (PEDOT:PSS). This results in appreciably high sensor densities of 100 sensor pixels per cm[Formula: see text] for spatial-temperature readings, while a small array size is maintained. Thus, a major impediment to the expansive application of these sensors is efficiently resolved. To realize fast and accurate interpretation of the sensor data, a neural network (NN) is trained and employed for temperature predictions. This successfully accounts for potential crosstalk between adjacent sensors. The spatial-temperature resolution is investigated with a specially-printed silver micro-heater structure. Ultimately, a fairly high spatial temperature prediction accuracy of 1.22 °C is attained.
Introduction
High-resolution thermal mapping is critical for emerging technologies like "artificial skins" for soft robotics, wearable healthcare monitors, and managing the heat of high-performance multi-core processor chips. However, traditional silicon-based sensors are often expensive and rigid, limiting their integration into flexible or large-area systems. While printed electronics offer a flexible, low-cost alternative, they face major hurdles in scaling sensor density and managing electrical crosstalk in dense "passive matrix" grids.
Methods and Design
The researchers utilized a passive matrix design consisting of three primary layers screen-printed onto a flexible PET (polyethylene terephthalate) substrate. This "sandwich" architecture simplifies fabrication while maximizing density:
- Bottom Layer: A series of parallel silver electrodes.
- Sensing Layer: A layer of PEDOT:PSS, a conductive polymer whose resistance changes predictably with temperature, acting as a thermistor.
- Top Layer: A series of parallel silver electrodes oriented at 90° to the bottom layer.
The intersections of these electrodes create individual sensor pixels. By using this vertical stack, the team achieved densities of 100 pixels/cm², far surpassing previous printed efforts that relied on more complex lateral structures.
Machine Learning for Signal Correction
To translate raw resistance data into accurate temperatures, the team employed a software-hardware co-design approach. Instead of attempting to eliminate every manufacturing variation in the hardware, they used a four-layer fully-connected Neural Network (NN). The NN takes raw resistance values along with the pixel's row and column coordinates as inputs. This allows the model to "learn" the specific quirks of each pixel and mathematically cancel out the electrical crosstalk inherent in passive matrix grids.
Experimental Results
The system's performance was evaluated using custom datasets. Training data was collected by placing the sensor array on a Peltier element and sweeping temperatures from 20 °C to 90 °C. To test spatial mapping, the authors printed custom silver micro-heaters in specific shapes to inject heat into the array, verifying the results with an infrared camera.
Key Results: * Accuracy: The system achieved an average temperature prediction error of 1.22 °C with a thermal resolution of 0.02 °C. * Speed: A single pixel can be read in 90 ms, with the entire 100-pixel array processed in approximately 27 seconds. * Stability: The sensors showed excellent temporal stability, maintaining accuracy even after a 12-hour sleep period.
Limitations and Future Work
While promising, the researchers identified several areas for improvement. PEDOT:PSS is sensitive to humidity, suggesting a need for encapsulation in future iterations. Additionally, while the 27-second scan time is sufficient for many applications, faster readout electronics would be required to monitor rapid thermal transients in high-speed microelectronics. Expanding the dynamic range beyond the current 20 °C to 90 °C window also remains a goal for industrial applications.
Summary updated