Development of an FPGA-Based Low-Field Pulsed NMR System with AI-Driven Data Processing for Point-of-Care Applications

Abstract

High-field NMR devices provide high spectral resolution, but are usually bulky and expensive. Low-field NMR offers a cost-effective alternative suitable for point-ofcare applications, though at the expense of spectral resolution. The aim of the project is to design a low cost, FPGAbased low-field NMR device that uses artificial intelligence for spectral analysis of blood. The system utilizes a Red Pitaya STEM 125-14 platform and comprises four components: a probe and magnet assembly, analog excitation and receiving chains, an FPGA-based digital processing chain, and a convolutional neural net-work (CNN) for compound identification. RF pulses were generated via a Direct Digital Synthesis (DDS) module to be fed into the probe with a 5 W power amplifier. The receiving chain consisted of two cascaded low-noise amplifiers intended to amplify weak return signals from the probe. Quadrature detection was implemented on the FPGA to extract the in-phase (I) and quadrature (Q) components. The CNN was trained on simulated NMR spectra of twelve compounds across multiple frequency ranges. The excitation chain successfully produced 21 MHz pulses, with the amplifier delivering approximately 94 Vpp to the probe. The receiving chain amplified signals as low as 115 μV while preserving signal integrity. Testing the quadrature detection chain confirmed the expected 90° phase relationship between I and Q components, with data rate reduced from 125 MS/s to 0.976 MS/s. The CNN achieved accurate compound classification on 21.29 MHz spectra, demonstrating the feasibility of AIassisted spectral analysis at low field strengths.