AI Challenge 2025 Winners
MathWorks has announced the winners of the 2025 AI Challenge. Congratulations and thanks to all the participating students.
Residual block with grouped convolutions and Squeeze-and-Excite (SE) channel attention; grouped convolutions and SE attention drive efficient, accurate feature learning from multichannel jet images
1st Place
Adit Shah
SVKM's Dwarkadas J. Sanghvi College of Engineering, India
Challenge project completed: Top Quark Detection with Deep Learning and Big Data
In high-energy physics, particle collisions leave “sprays” of energy called jets. Adit’s solution turns raw jet data into multichannel images and uses a custom CNN, designed with the Deep Network Designer app from Deep Learning Toolbox, to distinguish top-quark events from the background. The approach pairs thoughtful preprocessing with aggregated residual (ResNeXt-style) blocks and Squeeze-and-Excite attention so the model focuses on real physics patterns, not noise, achieving over 90% test accuracy. It also outlines a workflow for HDL code generation and FPGA deployment, bridging lab research to real-time systems.
This work highlights Adit’s innovative thinking and an end-to-end engineering mindset, reframing a physics problem as an image task, selecting modern model architectures, and planning for hardware deployment. Congratulations to Adit on an outstanding first-place achievement!
Real-world ISM-band capture with the model correctly segmenting Bluetooth and ignoring noise
2nd Place
Giacomo Aragnetti, Nicola Gallucci, Matteo Malagrino
Politecnico di Milano
Challenge project completed: Classify RF Signals Using AI
Giacomo, Nicola and Matteo tackled the crowded 2.4 GHz band by turning STFT spectrograms into pixel-level maps of Wi-Fi, Bluetooth, ZigBee, and SmartBAN activity. Their complete pipeline trains attention-gated U-Net and DeepLabv3+ (with ResNet-18/50 encoders) on a synthetic, channel-aware data set with class-priority masks, then validates on real airwaves using two synchronized ADALM-Pluto SDRs to capture and stitch 80 MHz of spectrum. Built in MATLAB with Computer Vision Toolbox and Parallel Computing Toolbox, and using Deep Learning Toolbox and Signal Processing Toolbox workflows, the solution is ready for practical spectrum monitoring and interference mitigation.
This work highlights Nicola and Matteo’s innovative, end-to-end engineering, from careful data set design and model choice to decisive real-world validation. Congratulations to Nicola and Matteo on a fantastic achievement!
True versus predicted velocity magnitude and vectors panel (instantly shows that the physics-informed model matches the flow)
3rd Place
Soham Gupta
Indian Institute of Technology, Ropar
Challenge project completed: Fluid Flow Simulation Using Physics-Informed Neural Networks
Soham’s project builds a physics-informed neural network in MATLAB to solve 2-D incompressible flow around a cylinder. Instead of learning from labeled examples, the network is trained to obey the governing equations and boundary conditions, so its predictions stay physically consistent. It produces the full flow field—velocity (u, v) and pressure (p)—and even estimates the fluid’s viscosity. A simple two-step training recipe and well-organized scripts make the work easy to reproduce, with clear visuals comparing predicted and true velocity, pressure, and flow rotation (vorticity).
This work highlights Soham’s blend of scientific rigor and practical engineering, turning core fluid-mechanics principles into a reliable learning system that others can run and extend. Congratulations to Soham on an elegant and insightful achievement!
