🔬 atomic physics

The BiP-PRISM algorithm for fast and scalable core-loss STEM-EELS simulations

The paper introduces the BiP-PRISM algorithm, which utilizes beam partitioning and natural-neighbor interpolation to significantly reduce the computational and memory costs of quantitative STEM-EELS simulations, enabling high-accuracy, full-resolution elemental mapping on consumer-grade GPUs.

Original authors: Philipp Pelz

Published 2026-07-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Philipp Pelz

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to take a high-resolution photograph of a tiny, intricate crystal structure using a super-powerful electron microscope. To understand exactly what elements are in the crystal (like iron, oxygen, or platinum), scientists use a technique called STEM-EELS. However, simulating how electrons bounce through this crystal to create a map is like trying to solve a massive, multi-layered puzzle. It's so computationally heavy that it usually requires supercomputers and can take days or weeks to run, often crashing standard computers because it needs too much memory.

This paper introduces a new algorithm called BiP-PRISM (Beam Partitioning - PRISM) that acts like a "smart shortcut" to solve this puzzle much faster and with less memory, all while keeping the picture incredibly accurate.

Here is how it works, using simple analogies:

1. The Problem: The "Re-doing Everything" Trap

In the old way of doing this simulation (Conventional Multislice), the computer acts like a very diligent but slow worker. For every single spot on the crystal it wants to scan, it has to:

Calculate how the electron beam enters the crystal.
Calculate how it bounces through every single layer of the crystal.
Calculate how it hits the detector.
Then it has to repeat this entire process for the next spot, and the next, and the next.

If you have a large map with thousands of spots, the computer is doing the exact same heavy lifting over and over again. It's like a delivery driver who, for every house on a street, drives from the warehouse to the house, drops off a package, drives back to the warehouse, and then drives to the next house. It's incredibly inefficient.

2. The First Shortcut: The "Master Blueprint" (PRISM)

Previous improvements (called PRISM) realized that the way the electron beam enters the crystal doesn't actually change based on where the beam is pointing; only the phase (a timing shift) changes.

The Analogy: Instead of driving from the warehouse every time, the driver pre-calculates a "Master Blueprint" of how the road looks from the warehouse to the middle of the street. They just apply a small timing shift to this blueprint for each house. This saved a lot of time, but the computer still had to do the heavy lifting for the exit side of the journey (how the signal gets to the detector) for every single spot.

3. The New Innovation: "Beam Partitioning" (BiP-PRISM)

The authors of this paper took the shortcut a step further. They realized they could compress both the entry blueprint and the exit blueprint.

The "Sparse Set" of Parents: Imagine you have a massive choir of 1,000 singers (the full electron beam). Instead of recording every single singer's voice, you pick a small group of 20 "Parent" singers who represent the whole choir.
The "Natural-Neighbor" Reconstruction: When you need to know what the full choir sounds like at a specific spot, you don't record them again. You just use a smart mathematical formula (Natural-Neighbor Interpolation) to blend the voices of those 20 Parents to perfectly recreate the sound of the 1,000 singers.
The Result: The computer only needs to do the heavy calculations for the 20 Parents, not the 1,000 singers. This is done for both the entry (probe) and the exit (detector) sides of the simulation.

4. The "Local Window" Trick

The paper proves a crucial mathematical fact: You only need to be accurate right where the atom is.

The Analogy: If you are trying to hear a specific conversation in a noisy room, you only need to focus your ears on the two people talking. You don't need to perfectly reconstruct the sound of the entire room.
Because the electron interaction happens in a tiny, localized spot (the "ionized atom"), the algorithm only needs to reconstruct the beam perfectly in that tiny window. It can be "fuzzy" everywhere else without affecting the final result. This allows the computer to use even fewer "Parent" beams, saving massive amounts of memory.

5. The Results: From "Impossible" to "Instant"

The paper demonstrates that this method works wonders:

Memory Savings: For a complex simulation of an iron-platinum nanoparticle (23,000 atoms), the old method required 12.7 GB of memory, which would crash many standard graphics cards. The new method did it in just 2.6 GB. It turned an "Out of Memory" error into a successful run.
Speed: It significantly reduced the time needed to calculate these maps.
Accuracy: Despite using these shortcuts, the resulting maps are nearly identical to the "perfect" (but slow) simulations. The paper shows a 99.9% match in the patterns and corrects the intensity so the numbers are scientifically accurate.
Versatility: The authors successfully simulated a complex interface between two different materials (LaAlO3 and SrTiO3) showing five different chemical elements simultaneously, and even created "momentum-resolved" maps (qEELS), all on consumer-grade computer hardware.

Summary

Think of BiP-PRISM as a smart compression tool for electron microscopy simulations. Instead of calculating every single detail for every single spot on a map, it calculates a few "master examples" and mathematically blends them together to create the full picture. It proves that you only need to be perfect in the tiny spot where the atom is, allowing the rest of the calculation to be simplified. This makes high-end, atomic-level chemical mapping possible on standard computers that previously couldn't handle the load.

Technical Summary: The BiP-PRISM Algorithm for Fast and Scalable Core-Loss STEM-EELS Simulations

Problem Statement
Quantitative interpretation of atomic-resolution Scanning Transmission Electron Microscopy - Electron Energy Loss Spectroscopy (STEM-EELS) requires dynamical simulations of the electron probe before and after core-loss transitions. Conventional transition-potential multislice algorithms are computationally prohibitive for large, heterogeneous fields of view because they repeat elastic propagation for every probe position and inelastic event, scaling poorly with simulation size. While the PRISM (Plane-wave Reciprocal-space Interpolated Scattering Matrix) algorithm accelerated these simulations by reusing a probe-forming scattering matrix ( $S_1$ ) across scan positions, it still faces significant memory and computational bottlenecks, particularly regarding the exit-side propagation and the storage of large scattering matrices required for detector-integrated elemental maps.

Methodology
The authors introduce the BiP-PRISM (Beam Partitioned PRISM) algorithm, which extends the PRISM framework by partitioning both the probe-forming scattering matrix ( $S_1$ ) and the detector-propagating scattering matrix ( $S_2$ ).

Beam Partitioning: Instead of propagating every aperture beam, the algorithm projects the scattering matrices onto a sparse set of "parent beams" (selected via a hex-ring subsample of the aperture). The full aperture is reconstructed locally at the ionized atom using natural-neighbor interpolation.
Dual-Matrix Partitioning:
- $S_1$ (Probe-forming): Computed on parent beams and reconstructed at the ionization plane.
- $S_2$ (Detector-side): Representing the exit propagation, this matrix is constructed via adjoint multislice back-propagation from the exit surface to the ionization plane. It is also partitioned onto a sparse set of parent detector beams.
Locality Theorem: The paper establishes a locality result demonstrating that the total simulation error is governed entirely by the on-atom reconstruction error. Because the transition potential ( $H_{n0}$ ) is sharply localized around the ionized atom, errors in the probe or detector reconstruction outside this small window do not affect the final intensity. This allows the matrices to be reconstructed only on a small "crop window" around each atom, rather than the full grid.
Magnitude-Preserving Reconstruction: To address amplitude loss caused by the decoherence of complex natural-neighbor interpolation, the authors introduce a magnitude-preserving reconstruction technique. This restores the correct absolute intensity scale, which is critical for quantitative mapping, without introducing approximation errors in the full-parent limit.
Focal Back-Propagation: For the probe-forming matrix ( $S_1$ ), the algorithm back-propagates parent columns to a scattering centroid plane before interpolation to minimize phase spread, further reducing reconstruction error.

Key Contributions

Double Partitioning: The first application of beam partitioning to both the probe-forming ( $S_1$ ) and detector ( $S_2$ ) matrices in the context of EELS.
Memory Reduction: By partitioning $S_2$ , the algorithm eliminates the need to store the full detector matrix, which is often the dominant memory consumer. This enables simulations on consumer-grade GPUs that would otherwise exceed memory limits.
Removal of Per-Scan Exit Propagation: For detector-integrated elemental maps, the algorithm completely removes the per-scan exit wave propagation loop, replacing it with a single matrix-vector product over the scan positions.
Unified Output Modes: The framework unifies the treatment of elemental maps, momentum-resolved qEELS, and full 4D cubes. While elemental maps and qEELS benefit from double partitioning ( $S_1 + S_2$ ), the full 4D cube utilizes only $S_1$ partitioning, as the full detector basis is required.
Validity Regime Characterization: The authors define the algorithm's validity regime, showing that error is primarily driven by specimen thickness (due to channeling) and is robust to defocus and edge depth variations.

Results
The authors benchmarked BiP-PRISM against conventional multislice (ground truth) and exact dual-matrix PRISM (exact reference):

Memory Efficiency: In a simulation of a 23,000-atom FePt nanoparticle, the double-partitioned algorithm reduced peak GPU memory from 12.7 GB (exact dual-matrix) to 2.6 GB (a ~5x reduction). At full resolution (19,082 grid), the exact method exceeded 48 GB of GPU RAM (OOM), while BiP-PRISM ran successfully at 12.6 GB.
Accuracy: At the 9,362 grid resolution, the double-partitioned map achieved a Pearson correlation of 0.999 with the exact reference and a relative L2 error of 9.3%. The magnitude-preserving reconstruction was shown to be essential for recovering absolute intensity, reducing errors from up to a factor of two (without correction) to within ~10%.
Scalability: The algorithm demonstrated linear scaling with specimen thickness and near-linear scaling with the number of scan positions, contrasting with the quadratic scaling of conventional multislice.
Multimodal Demonstration: The method successfully simulated a simultaneous five-edge (Sr-L, La-M, Ti-L, Al-K, O-K) atomic-resolution map of a LaAlO $_3$ /SrTiO $_3$ interface, resolving sublattices and chemical steps with a peak memory usage of only 2.4 GB.

Significance and Claims
The paper claims that BiP-PRISM is primarily an accelerator for memory feasibility, enabling full-resolution, large-field-of-view STEM-EELS simulations on consumer-grade hardware that were previously impossible due to memory constraints. While it offers significant speedups for the matrix construction phase, the overall speedup is Amdahl-limited by the exit propagation required for non-integrated outputs (like the 4D cube).

The authors emphasize that the algorithm maintains high quantitative accuracy (both pattern fidelity and absolute intensity) when the number of parent beams is tuned to satisfy the on-atom reconstruction error metric. The method is presented as a practical tool for enabling first-principles transition potential simulations and momentum-resolved EELS on standard GPUs, bridging the gap between theoretical simulation capabilities and experimental data scales. The paper concludes that the approximation is controllable and converges to exact PRISM as the number of parent beams approaches the total beam count.