Physics-Aware Super-Resolution of CMS Jet Images Using Transformer - Diffusion Architectures

At the CMS detector, jet images are stored at reduced resolution, discarding ~75% of total deposited energy. Standard super-resolution fails here because it optimises for visual similarity, not physical validity — a model without energy constraints converges to a response of 0.52, recovering only half the missing energy. As part of the ML4Sci/CMS test task, I developed a complete physics-aware conditional GAN pipeline (github.com/rajveer43/cms-superres-reconstruction) that maps 3×64×64 LR jet images to 3×125×125 HR reconstructions using a residual CNN generator, PatchGAN discriminator with spectral normalisation, and a three-component hybrid loss (LSGAN + L1 + energy response). The best configuration achieves raw-space L1 of 0.00698 vs 0.00846 for bicubic (17.5% improvement), PSNR 14.32 vs 14.13 dB, and energy response 1.0098 — within 1% of simulation ground truth. This GSoC project extends that validated baseline by implementing and benchmarking SwinIR-style transformer SR adapted for η-φ calorimeter coordinates, and a conditional DDPM/flow-matching diffusion variant as a stretch goal. Every architecture is evaluated on the same physics-constrained suite: energy response must remain within [0.99, 1.01] or the model is rejected regardless of pixel quality. Deliverables: reproducible GAN baseline with full metric suite, trained transformer SR model with benchmark comparison table, mid-project GAN vs transformer report, final benchmark across all architectures, technical report, and organized codebase for CMS collaboration use.