Voltage-to-temperature calibration of the High Altitude THz Solar telescope acquisition system

This paper presents the experimental characterization and voltage-to-temperature calibration of the High Altitude Terahertz Solar (HATS) photometer's acquisition system. The HATS telescope, installed at the Observatorio Astronómico Félix Aguilar (OAFA) in the Argentinean Andes, uses a Golay cell detector to capture incoming THz radiation modulated by a 20 Hz fork chopper. The study aims to establish a precise correlation between the detector's output voltage and the source's brightness temperature, which is necessary to produce scientifically interpretable solar data.

The experimental setup used a Tydex GC-1P Golay cell, an Omega BB-4A blackbody calibrator, a Thorlabs fork chopper with MC1000A control unit, and a Tektronix MSO2014 digital oscilloscope. The blackbody calibrator temperature was varied from 100°C to 375°C (373.15 K to 648.15 K) in 50°C intervals, with each temperature reading lasting 200 ms and 15-minute intervals between readings for stabilization. A total of 23 signals were generated, each containing 125,000 samples. The blackbody calibration temperature range was chosen to match the linear dynamic range expected at the sensor input under real observational conditions, since the effective temperature variation reaching the primary detector during solar observations corresponds to an equivalent brightness temperature deflection of only a few hundred Kelvin (around 500 K).

Two signal recovery methodologies were compared: 1) sinusoidal curve fitting (the Sine Method), where signals were fitted to V(t) = U sin(ωt + ϕ) using the curve fit function from scipy, and 2) windowed FFT combined with six windowing functions (rectangular, Hamming, Hann, Bartlett, Blackman, and Flat Top), where the amplitude at 20 Hz in the frequency domain was extracted. Correction factors were applied to each window function to avoid signal amplitude losses in the frequency domain (rectangular: 1.0, Hamming: 1.9, Hann: 2.0, Bartlett: 2.0, Blackman: 2.4, Flat-top: 4.7).

For each blackbody temperature, the output voltage amplitude U was obtained, and a linear regression U = aT + b was fitted, where a is the gain [mV/K] and b is the offset [mV]. The results showed high linearity in the system's response over the input source's temperature range. The gain values obtained with the windowed FFT were equal within 1-σ, with the Hamming window achieving the highest precision, yielding a Root Mean Square Error (RMSE) of 15.48 and a calibration temperature-to-voltage factor of 10.32 ± 0.13 mV/K. The Bartlett window presented the highest error (RMSE 16.02). The choice of windowing function had only a minor effect on the calibration factors, which ranged from 10.32 to 10.43 mV/K, all within the experimental uncertainties.

The weighted means for the windowed FFT method were a = (10.38 ± 0.02) [mV/K] and b = −(3630 ± 29) [mV]. The Sine Method yielded a = (9.88 ± 0.18) [mV/K] and b = −(3366 ± 95) [mV], which differ from the weighted means of the windowed FFT method by 3σ. A final overall weighted mean yielded a = (10.37 ± 0.02) [mV/K] and b = −(3607 ± 28) [mV]. Converting b to temperature yields T off = −348 [K], where the negative sign indicates this temperature is subtracted from the signal, and its absolute value can be interpreted as the sensor's internal noise.

The authors conclude that both methodologies are highly effective for system calibration, showing strong linearity within the range of 373.15 K to 648.15 K. The Hamming window provides slightly superior accuracy, but the minimal differences between the curves confirm that the choice of window function does not significantly impact the final calibration results, provided appropriate correction factors are applied. Beyond solar photometry, the signal modulation and windowing concepts established here are highly applicable to other fields requiring high-sensitivity thermal detection, such as industrial pyrometry for high-temperature manufacturing, environmental monitoring of atmospheric water vapor, and the development of medical imaging systems based on Terahertz radiation for non-invasive tissue analysis.

Improvements to AI Systems:

1. Adaptive Signal-Processing Pipeline for Chopped Detector Data

• Implement a hybrid AI module that automatically selects between sinusoidal curve fitting and windowed FFT based on real-time noise characteristics (e.g., SNR, harmonic distortion). The module would use the paper’s correction factors (Hamming: 1.9, Flat-top: 4.7) as priors, but learn to adjust them dynamically for non-stationary signals.

• Capability: An AI-driven calibration tool that processes raw detector voltages from any chopped THz or IR sensor, outputting brightness temperature with uncertainty bounds—without manual parameter tuning.

2. Physics-Informed Neural Network for Nonlinearity Compensation

• Train a small neural network to map raw voltage amplitudes to true temperatures, using the paper’s linear calibration (10.37 mV/K, offset −3607 mV) as the baseline. The network would learn residual nonlinearities beyond 373–648 K, extrapolating safely using the physical constraint that gain is positive and offset corresponds to internal noise.

• Capability: A self-correcting sensor model that maintains accuracy even when the blackbody source drifts or the detector ages, reducing recalibration frequency.

3. Uncertainty-Aware Window Selection via Meta-Learning

• Build a meta-learner that, given a short signal snippet (e.g., 200 ms), predicts which windowing function (among the six tested) will yield the lowest RMSE for that specific noise profile. Train it on synthetic data mimicking the paper’s 23-signal dataset, using the reported RMSE values (Hamming: 15.48, Bartlett: 16.02) as labels.

• Capability: An AI that autonomously picks the optimal spectral analysis method for any new thermal detection system, minimizing calibration error without human intervention.

4. Cross-Domain Transfer for Terahertz-Based Medical Imaging

• Use the calibration factors and linearity range to pre-train a convolutional or recurrent AI model for THz tissue analysis. The model would take modulated THz signals (as in the HATS setup) and output temperature maps, leveraging the known gain/offset to normalize inputs before feature extraction.

• Capability: A portable AI system for non-invasive burn-depth assessment or tumor margin detection, calibrated to absolute temperature rather than relative intensity, improving diagnostic consistency across devices.

5. Real-Time Anomaly Detection in Atmospheric Monitoring

• Develop an AI that monitors the HATS-like photometer’s output voltage stream, using the calibrated temperature-to-voltage relationship to flag sudden deviations (e.g., >3σ from expected solar brightness) as potential cloud cover, instrument malfunction, or water vapor spikes. The system would use the paper’s weighted mean offset (−3607 mV) as a baseline for internal noise.

• Capability: An early-warning system for environmental sensors that distinguishes genuine atmospheric events from detector drift, with false-alarm rates tuned using the reported RMSE values.

6. Automated Calibration for Industrial Pyrometry

• Create an AI controller that replicates the paper’s 15-minute stabilization protocol but optimizes it: it predicts when thermal equilibrium is reached using the detector’s voltage derivative, then triggers a measurement. It would use the paper’s gain (10.37 mV/K) to convert raw readings to temperature in real time, with a confidence interval derived from the 1-σ uncertainties.

• Capability: A self-calibrating pyrometer for high-temperature manufacturing (e.g., steel or glass), reducing calibration time by up to 50% while maintaining the paper’s ±0.13 mV/K precision.

The THz range has been under-explored for solar astronomy, mainly due to technological limitations. Only recently, a few new telescopes, such as the High Altitude Terahertz Solar (HATS) photometer, have been monitoring the solar activity in this range of the spectrum. This paper presents the experimental characterization and voltage-to-temperature calibration of the HATS acquisition system. HATS uses a Golay cell for capturing the incoming radiation, modulated by a 20 Hz fork chopper. The amplitude of the signal is then obtained at predetermined time intervals by applying a windowing function and an FFT to the signal. To characterize this acquisition system, a blackbody calibrator, with temperatures varying from 100 to 500 C (373.15 K to 648.15 K), was used as a THz source, and two signal recovery methodologies were compared: sinusoidal curve fitting and FFT combined with six windowing functions (rectangular, Hamming, Hann, Barlett, Blackman, and Flat Top). Our quantitative results demonstrate high linearity in the system's response over the input source's temperature range, with the Hamming window achieving the highest precision, yielding a Root Mean Square Error (RMSE) of 15.48 and a calibration temperature-to-voltage factor of 10.32 +- 0.13 mV/K. In contrast, the Bartlett window presented the highest error (RMSE 16.02). The choice of windowing function had only a minor effect on the calibration factors, which ranged from 10.32 to 10.43 mV/K, all within the experimental uncertainties. Beyond solar photometry, the signal modulation and windowing concepts established here are highly applicable to other fields requiring high-sensitivity thermal detection, such as industrial pyrometry for high-temperature manufacturing, environmental monitoring of atmospheric water vapor, and the development of medical imaging systems based on Terahertz radiation for non-invasive tissue analysis.