This Small Business Innovation Research (SBIR) Phase I project advances the design of ADCs. The ML ADC decomposes an analog signal into multiple channels, each containing only a portion of the signal’s total information. Each channel is sampled separately, slower than the input signal’s Nyquist rate, producing aliased and complicated digital outputs. A neural network, trained to approximate the inverse transfer function of the analog front-end, maps those digital outputs to the standard Shannon-Nyquist signal representation. 

This project implements one half of a common communication device, a Serializer/Deserializer receiver, using a ML ADC in printed circuit board prototype form. This project will examine the ability of the physical prototype to capture complete information in gigahertz-range pulse amplitude modulated (PAM) signals. This project expects to replicate computer simulations indicating high accuracy of the ML ADC when driven by a jittery and phase-drifting clock running below the Nyquist rate. The project will produce neural network models trained on the receiver’s data, and report the error rates achieved by the models in decoding PAM-4 signals recorded by the receiver.

A Serializer/Deserializer coordinates the sending and receiving of multiple data streams across one or more data channels. As information travels quickly over a distance, it stops behaving as ones and zeros, and starts looking more like waves of intermediate values. There is a lot of complicated engineering that is done in traditional SerDes to adjust what is being measured and how the measurement is done, and these procedures are reaching the limits of how fast they can operate.

Using our ML ADC architecture, we take lots of partial measurements at the same time, and use fully digital neural networks to interpret those partial measurements.  This lets us run the measurement operations more slowly, because each one doesn't need to capture all the information. It also lets us simplify the measurements that need to be taken, because the digital neural networks can be trained to handle complex or overlapping measurements.

If our device recovers all the signal information that is sent into it, it must be fully measuring the information.  There is no way to recover wideband signal information without capturing it and interpreting it. We've simulated the circuit using computer modeling, but we want to show that a physical device behaves the same way. By measuring the received signal and comparing it to the transmitted signal, we can see the error rate.

Data transmission is limited, in part, by the precision of the clocks that tell the circuit when to measure the signal.  Precise clocks take a lot of energy.  If the prototype behaves like our computer models, then we don't need precise clocks.  This will let our communication device operate faster, with less power, than traditional designs.  Faster data speeds at lower power consumption are useful in many different settings.