Machine learning-based data analysis for applications in diffraction, scattering, and porous materials

Dnr:

NAISS 2024/5-125

Type:

NAISS Medium Compute

Principal Investigator:

Magnus Röding

Affiliation:

Chalmers tekniska högskola

Start Date:

2024-03-27

End Date:

2025-04-01

Primary Classification:

10106: Probability Theory and Statistics

Secondary Classification:

10304: Condensed Matter Physics

Tertiary Classification:

10105: Computational Mathematics

Webpage:

This proposal is made in connection to the three projects below. There are three research directions which this application concerns; they are already in the pipeline but additional resources are required.
First, we are concerned with analysis for position averaged convergent beam electron diffraction (PACBED) which is a relatively new technique for characterizing materials, in particular 2D materials. We are developing methods for analysis of PACBED data for predicting materials parameters such as phase (crystal structure), thickness (number of atomic layers), tilt (relative to the imaging plane), and stress and strain (compression/expansion in different directions due to mechanical load). The projects concern both regression and classification approaches, both utilizing convolutional neural networks, both conventional CNNs and ResNet type deep CNNs. In one case, we are also developing an online learning-based method to account for very large amounts of data. We have spent considerable time modelling the experiment in terms of experimental design parameters, noise characteristics etc to facilitate accurate and robust prediction methods. We are investigating predictions for many different experimental/simulation parameters to optimize the method, resulting in heavy computations both from a data generation and CNN training standpoint.
Second, we continue earlier work by the PI (Prifling et al, 2021, DOI:10.3389/fmats.2021.786502) concerning prediction of mass transport properties, i.e. effective diffusivity and fluid permeability, using analytical equations as well as ANNs and CNNs. The current extension is about a kind of interpretable AI models that are monotonic and separable functions represented as a combination of separate ANNs. The monotonicity and separability is obtained by construction, using approaches from Runje & Shankaranarayana (2023). This class of models provides better understanding and interpretability than general ANN models that approximate arbitrary functions. The downside is a more complex implementation and more numerical instability during training.
Third, we are working on machine learning-based models for analysis of small angle X-ray scattering (SAXS) experiments. SAXS is a technique complementary to imaging for characterization of materials and molecules at small spatial scales. The field is dominated by fairly simple analytical models, but we are working on complex spatial models for nanoparticles with non-standard geometries. The models as well as the simulated experimental signals is computationally heavy and involve Monte Carlo simulations. This makes parameter inference in the model computationally heavy as well. A machine learning-inspired shortcut that has turned out fruitful is to train a neural network to predict the signal. In this manner, we have demonstrated a 200x-300x speed-up for analysis of data from one particular system. We intend to finish the investigation for this system as well as developing similar approaches for accelerating data analysis for other systems.
The project involves both experimental specialists from the Eva Olsson group at Physics at Chalmers, and AI expertise from the Mathematical Sciences and Electrical Engineering departments. To the extent possible, code and data will be published open access on Zenodo, ensuring optimal spread of the methods.