🔬 materials science

Mapping Order in Semicrystalline Polymers using Machine Learning of Nanobeam Electron Diffraction

This paper demonstrates that a machine learning model trained on synthetic data can rapidly and accurately detect polymer diffraction peaks in noisy nanobeam electron diffraction datasets, significantly outperforming conventional correlative algorithms to enable near-live visualization of structure-property relationships in semicrystalline organic mixed ionic electronic conductors.

Original authors: Nicholas Marchese, Arthur R. C. McCray, Yael Tsarfati, Karen Bustillo, Adam Marks, Alberto Salleo, Colin Ophus

Published 2026-07-21

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Nicholas Marchese, Arthur R. C. McCray, Yael Tsarfati, Karen Bustillo, Adam Marks, Alberto Salleo, Colin Ophus

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 understand how a city is built, but instead of looking at the streets from a helicopter, you are standing on the ground with a flashlight, trying to figure out the layout by looking at the shadows cast by the buildings. This is the challenge scientists face when studying a special class of materials called "organic mixed ionic electronic conductors" (OMIECs). These are stretchy, plastic-like materials that can conduct both electricity and ions (charged atoms), making them perfect for building brain-like computers, super-efficient electronics, and even bio-electronic devices that talk to our bodies.

To understand how well these materials work, scientists need to see their tiny internal structure, specifically how the long polymer chains are arranged. They use a powerful tool called a Transmission Electron Microscope (TEM), which shoots a beam of electrons at the material. When these electrons hit the material, they bounce off in specific patterns, creating a "diffraction pattern" that acts like a fingerprint of the material's structure. However, these materials are fragile; the electron beam can burn them, so scientists have to use very weak beams. This results in images that are incredibly noisy, like trying to hear a whisper in a hurricane. For years, the computer programs used to read these noisy fingerprints were like old, rigid templates: they could only find perfect, sharp shapes and often missed the messy, fuzzy, or broken patterns that real polymers make.

This paper introduces a clever new solution: a machine learning model trained to be a super-smart detective for these noisy fingerprints. Instead of using a rigid template, the researchers taught a computer program (specifically a "U-Net" architecture) to recognize the different types of patterns found in these polymers by showing it 500,000 fake, noisy images first. They created these fake images to look exactly like the real thing, complete with the three main types of structural signals: sharp pairs of spots (lamellar stacking), curved arcs (backbone periodicity), and fuzzy, spread-out blobs (π-π stacking).

The results show that this new AI detective is significantly better than the old methods. When tested on real data from a material called oxidized p(g3T2), the machine learning model found far more structural details than the traditional "correlative" algorithms. For instance, it detected lamellar peaks in 84.90% of the scanned spots, whereas the old method only found them in 23.57%. It also found 95.81% of the fuzzy π-π peaks compared to just 60.05% for the old method. Perhaps most impressively, the AI did this roughly 5.5 times faster, taking only 3 minutes and 20 seconds to process the data compared to over 18 minutes for the traditional method.

The paper also reveals that the old methods often missed the "faint" or "broken" patterns that are actually there, leading scientists to think the material was less organized than it really was. By using the new AI, the researchers were able to draw much more continuous and complete maps of the material's internal orientation, showing that the structure is often more connected and uniform than previously thought. While the model still struggles with extremely faint signals or very specific edge cases (like certain harmonic patterns), it represents a major step forward. It allows scientists to see the hidden order in these complex, messy materials with greater speed and accuracy, potentially opening the door to designing better materials for the electronics of the future.

Technical Summary: Mapping Order in Semicrystalline Polymers using Machine Learning of Nanobeam Electron Diffraction

Problem Statement
Organic mixed ionic electronic conductors (OMIECs) are promising materials for neuromorphic and bioelectronic applications, yet the relationship between their microstructure and performance properties, such as charge carrier mobility, remains poorly understood. While Transmission Electron Microscopy (TEM), specifically four-dimensional scanning TEM (4DSTEM), offers a powerful means to characterize nanoscale structural heterogeneity, analyzing semicrystalline polymer diffraction patterns presents significant challenges. Unlike highly crystalline materials, semicrystalline polymers exhibit three distinct diffraction signal morphologies: sharp lamellar stacking reflections near the central beam, sharp but often asymmetric backbone periodicity arcs, and diffuse $\pi$ - $\pi$ stacking intensity distributed over the amorphous halo.

Conventional correlative template matching algorithms (e.g., in the py4DSTEM toolkit) struggle with this complexity. These methods require manual tuning of multiple hyperparameters for each peak class, often leading to a trade-off where optimizing detection for one peak type increases error rates for others. Furthermore, polymer samples are highly beam-sensitive, necessitating low-dose imaging that results in diffraction patterns with very low signal-to-noise ratios. This leads to Type I errors (false positives from noise) and Type II errors (missed weak peaks), hindering accurate orientation mapping and microstructural analysis.

Methodology
To address these limitations, the authors developed a machine learning (ML) model based on a U-Net architecture to automate the detection of Bragg peaks in noisy 4DSTEM datasets.

Synthetic Data Generation: Since ground-truth peak positions and intensities are unknown in experimental data, the team generated 500,000 synthetic training images. These simulations included ground-truth sets of $\pi$ - $\pi$ , lamellar, and backbone peaks, along with amorphous halos and noise. The generation process utilized "recipes" with Poisson statistics to vary peak intensity, radial distance, annular spread, and asymmetry, ensuring the model was robust to diverse morphological variations.
Model Architecture: The U-Net model takes a 256 $\times$ 256 diffraction pattern as input and outputs two 256 $\times$ 256 channels: one for peak position (modeled as sharp 2D Gaussian distributions) and one for peak intensity (modeled as broad circles). The architecture consists of four layers with three convolutions per layer, starting with 32 filters and a 3 $\times$ 3 kernel size.
Training and Inference: The model was trained for 200 epochs, with weights selected at epoch 191 where training and validation losses were lowest. During inference, a Gaussian filter suppresses noise in the position channel, followed by a local-maximum criterion to identify candidate peaks. Coordinates are refined to subpixel accuracy via quadratic fitting, and intensities are sampled from the intensity channel at these refined coordinates.
Parameter Efficiency: The method requires only a single user-specified detection threshold, a significant reduction compared to the 15 hyperparameters often required for manual tuning in correlative methods.

Key Results
The model was validated using separate synthetic datasets and experimental 4DSTEM data from an oxidized p(g3T2) sample.

Performance on Synthetic Data: The model's F1 score (harmonic mean of precision and recall) improved monotonically with peak intensity and radial distance for lamellar and backbone peaks. Detection thresholds were established at approximately 50 counts for lamellar peaks and 60–80 counts for backbone peaks. For $\pi$ - $\pi$ peaks, detection performance was dependent on annular spread, with thresholds ranging from 150 to 800 counts. The model demonstrated robustness even when lamellar peaks partially overlapped the central beam.
Experimental Comparison: When applied to an oxidized p(g3T2) dataset, the ML model significantly outperformed the conventional correlative algorithm:
- Lamellar Peaks: Detected 50,579 peaks (84.90% of probe positions) versus 7,529 peaks (23.57%) by the correlative method.
- Backbone Peaks: Detected 49,490 peaks (78.94%) versus 23,006 peaks (65.62%).
- $\pi$ - $\pi$ Peaks: Detected 76,815 peaks (95.81%) versus 25,569 peaks (60.05%).
- Overall Coverage: The ML model achieved peak detection in 99.65% of total probe positions, compared to 85.16% for the correlative algorithm.
Processing Speed: The ML model processed the dataset in 3 minutes and 20 seconds on a GPU, approximately 5.5 times faster than the 18 minutes and 21 seconds required by the correlative algorithm.
Orientation Mapping: Orientation maps generated from ML-detected peaks revealed nearly continuous crystalline domains, particularly for lamellar stacking, whereas correlation-based maps appeared patchy. This suggests that previous "loss" of order was an artifact of insufficient detection sensitivity rather than poor crystallinity.

Significance and Claims
The paper claims that this ML approach provides a more accurate, robust, and user-friendly method for analyzing semicrystalline polymer microstructures. By eliminating the need for extensive hyperparameter tuning and improving detection of faint and morphologically distinct peaks, the method enables more complete extraction of information from 4DSTEM datasets.

The authors highlight that the computational efficiency of the model makes near-real-time visualization of 4DSTEM experiments feasible, allowing operators to adjust experimental conditions during acquisition. While the model was tested on specific OMIEC systems (oxidized and reduced p(g3T2), and reduced PB2T-TEG), the authors suggest the approach may be viable for other weakly scattering materials, though retraining or tuning might be necessary for significantly different material classes. The study concludes that future improvements could involve expanding training datasets to include edge cases like peak harmonics and more complex peak configurations to further enhance robustness.

Technical Summary: Mapping Order in Semicrystalline Polymers using Machine Learning of Nanobeam Electron Diffraction

More like this