🔬 materials science

An End-to-End Differentiable Forward Model for High-Energy Diffraction Microscopy

This paper introduces the first end-to-end differentiable forward model for High-Energy Diffraction Microscopy (HEDM) that covers far-field, near-field, and point-focused geometries, enabling gradient-based joint parameter refinement and uncertainty quantification while achieving pixel-exact agreement with established simulators.

Original authors: Hemant Sharma, Nina Andrejevic, Simon Zhang, Mathew Cherukara

Published 2026-08-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Hemant Sharma, Nina Andrejevic, Simon Zhang, Mathew Cherukara

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 trying to figure out the exact shape, position, and internal stress of millions of tiny, invisible crystals inside a piece of metal, just by looking at the pattern of light they scatter when hit by a super-bright X-ray beam. This is the job of High-Energy Diffraction Microscopy (HEDM). Think of it like a cosmic game of "Where's Waldo?" but instead of a cartoon character, you are hunting for the precise location and orientation of individual grains in a metal sample, using the shadows and spots they cast on a giant detector screen. For decades, scientists have used powerful computers to simulate how these crystals should scatter light, and then compared that simulation to the real data to guess the answer. But there was a catch: the computer code used to make these simulations was like a black box. You could put numbers in and get a picture out, but if you wanted to tweak the numbers to make the picture match better, the computer couldn't tell you which way to turn the knobs. It had to guess blindly, trying one setting after another, which was slow and often got stuck in dead ends.

Now, imagine if that black box suddenly became transparent and could whisper, "Hey, if you turn this knob just a tiny bit to the left, the picture gets 10% clearer." That is exactly what this new paper achieves. The researchers have built the first "end-to-end differentiable" forward model for HEDM. In plain English, they rewrote the simulation software so that it can calculate not just the answer, but also the exact "slope" or direction needed to improve that answer instantly. They tested this new software against the old, trusted "gold standard" simulators used by scientists for years, and found that it matches them perfectly, down to the individual pixel on the detector screen. This means scientists can now use a smart, math-based "gradient" to simultaneously adjust the position of the crystals, their internal strain, and even the geometry of the camera itself, all in one smooth, fast motion. It turns a slow, blind guessing game into a guided tour where the computer knows exactly how to find the best solution.

The Paper's Core Discovery

The paper presents a new software framework called midas-diffract that acts as a universal translator for High-Energy Diffraction Microscopy. The authors created a single, unified computer program written in PyTorch (a popular tool for artificial intelligence) that can simulate three different ways of looking at crystals: Far-Field (where the detector is far away), Near-Field (where it's close), and Point-Focused (where the beam is tiny).

The main breakthrough is that this new model is differentiable. In the world of math and computer science, this means the software can calculate how a tiny change in an input (like moving a crystal slightly) changes the output (the pattern on the screen). Because of this, the software can use "gradient-based" methods to refine its guesses. Instead of guessing and checking, it can slide down a mathematical hill to find the lowest point—the perfect fit—much faster and more accurately than before.

What They Proved and What They Rejected

The authors explicitly reject the idea that you need to use "learned surrogates" (neural networks trained on data) to get differentiability. They argue that while neural networks are fast, they are approximations that can introduce errors and don't follow the strict laws of physics. Instead, they proved that you can keep the exact, rigorous physics of the old C-language code (which has been battle-tested for 20 years) and still make it differentiable.

They demonstrated this by comparing their new PyTorch model against the established "MIDAS" reference simulators. The results were stunningly precise:

Far-Field: They matched 162 out of 162 spots with pixel-level accuracy.
Near-Field: They matched 2,304 out of 2,304 pixels exactly, even when the detector was tilted.
Point-Focused: They matched 1,088 out of 1,096 spots, with the few misses happening only in extremely tricky mathematical edge cases near the center of the beam.

The paper confirms that their new model is pixel-exact with the old physics, meaning it doesn't sacrifice accuracy for speed. It is not just a "good enough" approximation; it is a mathematically identical twin of the trusted code, just with the added superpower of knowing its own derivatives.

Real-World Tests and Limits

To show it works in the real world, the team ran three major tests:

The "Perturb-and-Recover" Test: They took a real dataset of 214 grains from a titanium sample, scrambled the starting positions by 1.5 degrees, and let their new software fix it. It recovered 100% of the grains, finding the correct orientation and position with a precision of about 6 nanometers (that's smaller than a virus!).
The "Joint Refinement" Test: They showed that the software can fix the camera's geometry (like its tilt and distance) at the same time it fixes the crystal positions. In a synthetic test with four detector panels, it corrected the camera's rotation axis to within 26 microradians, a level of precision that was previously impossible to achieve in a single step.
The "Basin of Convergence" Test: They checked how far off a starting guess could be before the software got lost. They found that as long as the starting guess was within about 10 to 12 degrees of the truth, the software would successfully find the correct answer. This is great news because standard tools usually get you within 5 degrees, meaning this new tool is ready for immediate use.

However, the paper also sets clear boundaries. The software is not a magic wand that works from any starting point; if you start more than 15 degrees off, it might get stuck in a local trap. Also, while it is incredibly fast on a computer with a powerful graphics card (GPU), it still relies on the assumption that each crystal grain acts like a single point of light. It doesn't yet model the complex 3D shape of the beam hitting a large grain, though the authors say this is a natural next step.

Why This Matters

This paper doesn't just offer a faster calculator; it opens the door to a whole new class of experiments. Because the model is differentiable, scientists can now:

Combine Physics and Data: They can add rules about how crystals should behave (like conservation of mass) directly into the math, forcing the solution to be physically realistic.
Track Changes Over Time: They can watch crystals change shape and stress in real-time during heating or loading, linking the experiment directly to computer simulations of how materials break.
Quantify Uncertainty: They can calculate exactly how sure they are about their measurements, rather than just guessing.

In short, the authors have built a bridge between the rigid, proven physics of the past and the flexible, powerful optimization tools of the future. They haven't just improved the speed; they've changed the rules of the game, allowing scientists to solve problems that were previously too complex to tackle.

Technical Summary: An End-to-End Differentiable Forward Model for High-Energy Diffraction Microscopy

Problem Statement
High-Energy Diffraction Microscopy (HEDM) is a primary technique for the three-dimensional, non-destructive characterization of polycrystalline metals at synchrotron sources. It recovers grain-level quantities—crystallographic orientation, lattice strain, and position—from rotating-crystal X-ray diffraction patterns across three geometries: far-field (FF), near-field (NF), and point-focused (pf-HEDM).

Despite its widespread use, existing HEDM forward models (implemented in C/C++ in codes like MIDAS, HEXRD, and ImageD11) are not differentiable with respect to their inputs. This structural limitation prevents the use of gradient-based joint optimization, physics-informed regularization, and Bayesian uncertainty quantification. While learned surrogates (e.g., BraggNN) offer differentiability, they introduce uncharacterized approximation errors and lack pixel-exact agreement with established physical conventions. The open problem addressed by this work is whether HEDM can be made end-to-end differentiable while preserving pixel-exact agreement with the validated C physics engines used in production.

Methodology
The authors present midas-diffract, an end-to-end differentiable forward model implemented in PyTorch that covers FF, NF, and pf-HEDM geometries within a single unified nn.Module class.

Unified Architecture: The model accepts grain states (Euler angles, positions, lattice parameters/strain tensors) and instrument states (detector geometry, beam parameters) to simulate detector signals. It branches into mode-specific output heads: angular coordinates for FF/pf-HEDM and Gaussian-splatted detector volumes for NF-HEDM.
Differentiability Preservation: To convert production C idioms into a differentiable form, the authors replaced non-differentiable operations with autograd-compatible primitives:
- Branching: Replaced if statements with torch.where to ensure gradients flow across decision boundaries (e.g., degenerate cases in the omega quadratic solver).
- Trigonometry: Clamped arguments to inverse trigonometric functions (acos, asin) to avoid unbounded derivatives at domain boundaries.
- Orthogonalization: Replaced SVD-based SO(3) projection with a Newton–Schulz iteration, which is differentiable even at repeated singular values.
- Validity Handling: Enforced spot validity via binary masks multiplied into loss terms, ensuring invalid spots contribute zero gradient.
Intensity Model: The model treats grains/voxels as point sources with CIF-based relative intensities, matching the conventions of MIDAS and HEXRD to ensure pixel-exact agreement with reference simulators.

Key Contributions

First End-to-End Differentiable HEDM Simulator: A unified PyTorch implementation covering all three HEDM modalities (FF, NF, pf) that exposes gradients through the full diffraction physics.
Pixel-Exact Agreement: The model achieves floating-point precision agreement with the established MIDAS reference simulators (ForwardSimulationCompressed and simulateNF):
- FF: 162/162 spots matched with sub- $10^{-3}$ pixel residuals.
- NF: 2304/2304 pixels matched identically, even with detector tilts up to $1^\circ$ .
- pf-HEDM: 1088/1096 spots matched (99.27%), with mismatches occurring only at known numerical branch-selection boundaries near the beam axis.
Open-Source Release: The framework is released as the midas-diffract package (v0.1.0) on PyPI, accompanied by reproducibility scripts.

Results
The framework was validated through three primary demonstrations:

Joint Refinement in NF-HEDM: Achieved sub-voxel precision recovery of orientation, strain, and position simultaneously. In noise-free conditions, position was recovered to $\sim6$ nm; under resolution-matched noise, to $\sim66$ nm.
Real-Data Round-Trip (FF-HEDM): Applied to a 214-grain $\alpha$ -Ti dataset from the 1-ID-E beamline. The differentiable model matched the production MIDAS fit residuals (within 0.3%) and achieved 100% grain recovery from a $1.5^\circ$ initial perturbation. The basin of convergence on real data was found to be $\sim10^\circ$ , consistent with synthetic data.
Multi-Panel Geometry Calibration: Demonstrated joint refinement of per-detector geometry (including beam-axis rotations and global wedge angles) and per-grain states on a synthetic four-panel setup. The model recovered panel rotations to $\sim10$ $\mu$ rad and global wedge angles to $\sim26$ $\mu$ rad, outperforming sequential refinement strategies.
Performance: Benchmarks show the GPU (NVIDIA RTX A6000) provides a $\sim3\times$ speedup over CPU for batch sizes of 100–1000 grains, enabling sub-second feedback for typical beamtime reconstructions.

Significance and Claims
The paper claims that this work closes a two-decade gap in HEDM methodology by enabling gradient-based inference without sacrificing physical fidelity. The authors state that this differentiable substrate is the foundational requirement for a new class of capabilities that were previously foreclosed, including:

Sub-voxel grain mixtures: Soft assignment of voxels to multiple orientations.
Physics-informed regularization: Enforcing crystal symmetry and microstructure priors directly within the loss function.
Bayesian Uncertainty Quantification: Enabling Hamiltonian Monte Carlo or variational inference via exact likelihood gradients.
Temporal 4D-HEDM: Joint reconstruction across time frames with temporal smoothness.
Crystal-Plasticity Coupling: Direct comparison of experimental grain maps with differentiable finite-element simulations (e.g., JAX-FEM) to tune material parameters.

The authors emphasize that this is not a replacement for existing tuned pipelines where they already work, but a necessary prerequisite for extending HEDM into these more complex, coupled, and probabilistic regimes. The work establishes that HEDM can be differentiable while maintaining the pixel-exact agreement required for production-level scientific analysis.

The Paper's Core Discovery

What They Proved and What They Rejected

Real-World Tests and Limits

Why This Matters

Technical Summary: An End-to-End Differentiable Forward Model for High-Energy Diffraction Microscopy

More like this