More information
Description
The idea of the intelligent sensor design was proposed long ago, however, there were no tools and methods allowing for reliable on-chip implementation. Utilizing deep sub-micron technologies like CMOS 28 nm allows for dense on-chip logic synthetizing and therefore gives an opportunity for artificial neural network (ANN) placing as close to the signal as possible. Even though ANNs are mostly used for much bigger problems in biology and physics where they are big, complex and require enormous computation power, the PI claims that for simple signal processing like pattern recognition of low level signals they are small enough to fit into an IC. X-ray detectors are on the continuous race for better energy resolution and higher speed. Scientific groups around the world like CERN, PSI, Fermilab, SPring-8 are reaching limits of the theoretical values, but still scientists using those detectors want to have more precise numbers. The traditional way used by most groups is to overcomplicate the circuit schematics by adding more and more circuits which are mostly used for compensation of integrated circuits imperfections in various stages of the signal processing path. In contrast to this approach, the PI is proposing to use an artificial neural network to compensate for moderate parameters of the sensor quality and moderate front-end signal processing parameters.
Summary of project results
The project was oriented toward the design and fabrication of the pixel readout integrated circuit (IC) with intelligent X-ray photons detection. The intelligence was realized by the on-chip implementation of an artificial neural network (ANN). The network, located as close to the source of the photon event as possible, enables precise reconstruction of photon energy, assuming limited Analog-to-Digital Converter (ADC) resolution and moderate parameters of the IC front-end.
As a result of the project, the "IntelPixel" IC was designed as a matrix of 8 x 8 pixels, each 200 um side size. The IC was fabricated in a deep submicron 28 nm technology. The architecture of every single pixel consisted of the following blocks: Charge-Sensitive Amplifier (CSA), ADC, and ANN. In the presented type of detector, a photon event is visible at the input of the pixel as a current pulse, which is then converted to a voltage pulse by the CSA block. It is the amplitude of this voltage pulse that carries the information about the energy of the photon event. After the CSA block, the pulse is quantized by the 6-bit ADC that feeds the ANN block with the output samples. The ANN is trained beforehand to withdraw photon energy from the input data with a precision exceeding the one that could be obtained solely by the ADC.
The IntelPixel is the first IC with in-pixel ANN implementation in the field of nuclear science, proving it is possible and justifiable to implement such a solution in particle detectors. Moreover, even after its fabrication, the presented solution can still be retrained for different purposes than precise photon energy measurement. For example, the ANN can be retrained to recognize a type of incoming particle, or so-called pile-up event, in which two photons arrive so close to each other that they are hardly recognizable separately. Consequently, the designed detector presents an alternative solution starting a new branch in detectors development.