Imagine trying to capture a clear picture of a super-fast zooming cloud of electrons. This "cloud" doesn't just have a standard shape; each particle inside is moving in different directions at different speeds. Capturing all this rich information—exactly where the particles are and where they're going—is called Phase Space Reconstruction.
Normally, scientists use simple mathematical assumptions (like a perfect bell curve) to estimate what this cloud looks like. But real electron beams can have messy, irregular shapes with long trails and complex internal structures. Simple math misses all those important, intricate details!
Instead of guessing with simple equations, I implemented an advanced AI called a Particle Transformer to reconstruct the true, exact shape of the beam from a series of 2D images.
We pass the electron beam through a set of four magnets called a Quadrupole. These magnets squash and stretch the beam in different directions (vertically and horizontally). By changing the strength of the squish, the beam casts different cross-sectional "shadows" (images) onto a glowing screen further down the line.
Our AI starts with a random blob of particles and simulates squishing them exactly like our real-world magnets did. It compares its simulated shadows against the real shadows we captured at the SwissFEL facility. Using a process called Differentiable Simulation, the AI slowly sculpts the random blob until its shadows perfectly match all the real ones—giving us the true, complex 4D picture of the beam!
For experts reviewing this work, here is a deeper mathematical context mapping the physical system directly to the differentiable pipeline:
Standard phase space diagnostics typically extract beam parameters like the transverse emittance \(\epsilon_{x,y}\) and Twiss parameters (\(\alpha, \beta, \gamma\)) assuming perfectly finite Gaussian constraints. However, realistic bunches exhibit non-linear correlations and significant tail deviations. We establish a framework capable of recovering the complete 4D transverse phase space covariance matrix \( \Sigma \) by varying the quadrupole gradient. Sweeping the magnetic field strength captures an expansive Phase Advance (\(\mu\)) spectrum between \(0 \rightarrow \pi\), generating the linearly independent projections necessary for true non-parametric tomographic reconstruction across the drift section limit.
The core Neural Architecture exploits a Multi-Head Self-Attention Particle Transformer. The model computes non-Gaussian macroscopic \(\mathbb{R}^4\) coordinate vectors transformed from initial simple independent normal noise inputs (\(X \sim \mathcal{N}(0, I)\)) via parameterized network weights \( \theta \). Instead of statistical moment comparisons, parameter optimization depends strictly on the measured pixel intensity footprint. The model trains by minimizing the Mean Image Difference loss:
Here, \( R_n \) is the experimental reference image pixel array and \( Q_n \) is our density estimation formulated via Differentiable Splatting. This splatting kernel mathematically approximates the non-differentiable rigid striking map by leveraging spatial Gaussian blurs. This preserves continuous gradients, smoothly mapping the empirical image error directly back to the transformer weights sequentially without ever breaking the underlying thick-lens optics lattice assumptions!
Stay updated for full overview and understanding accelerators.