Here's a number that surprised me when I first started at Lawrence Livermore National Laboratory: more than half of the Large Hadron Collider's computing budget goes to simulation, not to analyzing actual collision data. Physicists need simulated particle showers to know what to look for in real ones, and generating those simulations with traditional software like GEANT4 can take several minutes per event.
Minutes. For a single simulated collision. The LHC produces about 600 million collisions per second.
I worked in the Physics Division at LLNL, where the mission was merging computing with physics to advance both. My project sat right at that intersection: could we train a generative adversarial network to produce synthetic particle physics images that were statistically indistinguishable from the real thing?
The cost problem
When physicists discovered the Higgs boson in 2012, the discovery required more simulated data than actual experimental data. That's not unusual. It's how particle physics works. You simulate billions of collision outcomes under different theoretical assumptions, then compare those predictions against what the detector actually measured. The mismatch (or match) between simulation and reality is how you find new physics.
Simulating particle showers is computationally brutal. When a particle hits the calorimeter layers of a detector, it cascades into a shower of secondary particles (photons, electrons, positrons) depositing energy across the detector in complex spatial patterns. GEANT4, the gold-standard simulation toolkit used at CERN and national labs worldwide, models this from first principles. It tracks every particle, every interaction, every energy deposit. Beautiful, exhaustive, and painfully slow.
It's getting worse. The high-luminosity upgrade of the LHC will produce an order of magnitude more data, requiring hundreds of billions of simulated collisions just to keep Monte Carlo uncertainty low enough for precision measurements. At current simulation speeds, the computing infrastructure can't keep up.
What GANs actually do here
A GAN is two neural networks locked in a game. The generator takes random noise and tries to produce fake calorimeter images. The discriminator looks at both real images (from GEANT4 simulations) and the generator's fakes, trying to tell them apart. They train together: the generator gets better at fooling the discriminator, the discriminator gets better at catching fakes, and through this adversarial process the generator learns to produce remarkably realistic synthetic data.
But particle physics isn't like generating faces or landscapes. You can't just produce something that looks right. It has to be right. The total energy deposited across calorimeter layers must be conserved (energy doesn't vanish or appear from nowhere). The spatial distributions have to match known physics. The sparsity patterns, where most pixels are zero because most of the detector isn't hit, need to be statistically accurate.
That's what made our research interesting. We weren't just asking "can the images fool a classifier?" We were asking "can these synthetic images fool physics?" Would analysis techniques applied to the synthetic data produce the same physical conclusions as analysis applied to GEANT4 data?
The work
I spent roughly two years in the Physics Division researching and iterating on GAN architectures tuned for calorimeter image generation. Training data came from high-fidelity GEANT4 simulations of electromagnetic showers, the expensive images we were trying to learn to produce cheaply.
The architecture needed modifications you wouldn't see in a standard image GAN. Calorimeter images are single-channel (each pixel is energy, not color). The dynamic range is extreme: some cells deposit kiloelectronvolts, others deposit nothing. The images have three layers corresponding to the physical layers of the calorimeter, each with different granularity. And the generator had to model the sequential physical relationship between these layers, since a shower develops top-to-bottom through the detector, meaning layer 2's pattern depends on layer 1.
We also added physics-informed loss functions to penalize energy non-conservation. In a standard GAN, the loss is purely adversarial: "did the discriminator catch you?" In our setup, there was an additional penalty for producing showers where the total reconstructed energy diverged from the nominal input energy. Domain knowledge baked into the training loop, enforcing physical constraints a general-purpose architecture would have no reason to learn.
The research was presented to the lab multiple times but was not published, which is common at national labs where only a fraction of projects go through the publication pipeline. The related public work in this space (CaloGAN from Lawrence Berkeley National Lab) demonstrated speedups of up to 100,000x over GEANT4, and our approach explored similar territory with different architectural choices.
What I took away
Working at a national laboratory recalibrates your sense of scale. The problems aren't "how do I make this API endpoint faster." They're "how do we generate a petabyte of synthetic data that's physically correct." The bar for correctness is different too. In a web application, a 1% error rate is usually fine. In particle physics, a 1% systematic bias can mean the difference between discovering new physics and missing it entirely.
I also came away with a clearer picture of where computing and science are converging. Some of the most important discoveries in physics now come from data, not from theory. The Standard Model is so complete that finding cracks in it requires examining vanishingly rare events, things that happen once in a billion collisions. You need enormous datasets, both real and simulated, just to have statistical power over these processes. Machine learning isn't a nice-to-have here. It's becoming part of the scientific infrastructure.