This paper describes a temperature sensor based on the thermal diffusivity of silicon. Its digital output is insensitive to both process spread and packaging stress and is a near-linear function of absolute temperature. The sensor's accuracy is mainly limited by lithographic resolution, and so benefits from Moore's Law. A sensor fabricated in a 0.18µm CMOS process exhibits an untrimmed device-to-device inaccuracy of ±0.2°C (3σ) from -55°C to 125°C. This is significantly better than that of similar sensors fabricated in a 0.7µm CMOS process [1,2].The AC thermal characteristics of silicon are determined by its thermal diffusivity, D. For lightly doped IC-grade silicon, D is insensitive to process spread and has a well-defined temperature dependence [3]. D can be determined by measuring the phase response of an electrothermal filter (ETF), which consists of a heater and a relative temperature sensor located in the same substrate. Fig. 17.4.1 shows a CMOS implementation, in which the heater is an n + -diffusion resistor and the sensor is a p + /aluminum thermopile. AC heat diffusing from the heater creates temperature fluctuations at the thermopile, which are low-pass filtered by the substrate's thermal inertia. The phase shift of the thermopile's output (φ ETF ), with respect to a harmonic heater drive signal (f drive ), is determined by D and by s, the spacing between the heater and the thermopile. For a constant f drive , φ ETF is a near-linear function of absolute temperature [1].The spread of φ ETF is mainly determined by spread in the position of the diffusion/aluminum contacts, since this defines the effective value of s ( Fig. 17.4.1). This, in turn, is determined by the lithographic accuracy of the process. In a 0.7µm CMOS process, with s~ 23µm, this resulted in a temperature inaccuracy of about ±0.7°C (3σ) [2]. In this work, greater accuracy was achieved by improving the interface electronics and by leveraging the more accurate lithography of a 0.18µm process.As shown in Fig. 17.4.2, φ ETF is digitized by a 1 st -order phase-domain ΔΣ modulator (PDΔΣM). The ETF's output signal, V ETF , is first converted to a current by a g m -stage, and then multiplied by a copy of f drive with a relative phase shift of φ fb . The multiplier implements a phase-differencing node, since the DC component of the resulting current, I demod , is proportional to cos(φ ETF -φ fb ). This current is then applied to the loop filter: an active integrator. Based on its output, a comparator selects φ fb from one of two digitally generated phase references (φ 0 and φ 1 ), which straddle the temperature range of interest.When driven by a 42kHz 2.5mW square wave, the amplitude of the ETF's output is rather small (~400µV pp ). Since the thermopile's resistance (R tp = 20kΩ) produces thermal noise, measurement bandwidth must be traded off against temperature-sensing resolution. The PDΔΣM's sinc decimation filter can be used to define a conversion rate of 0.16Hz, which corresponds to a temperature sensing resolution of 0.02°C rms . At a samp...

No abstract