WorksPhysics and Astronomy
Eigenstructure Analysis of Bloch Wave and Multislice Formulations for Dynamical Scattering in Transmission Electron Microscopy
Abstract
We investigate the eigenstructure of matrix formulations used for modeling scattering processes within materials in transmission electron microscopy. Dynamical scattering is crucial for describing the interaction between an electron wave and the material under investigation. Unlike the Bloch wave formulation, which defines the transmission function via the scattering matrix, the traditional multislice method is lacking a pure transmission function due to the entanglement of electron waves with the propagation function. To address this, we reformulate the multislice method into a matrix framework, which we refer to as the transmission matrix. This allows a direct comparison to the scattering matrix derived from Bloch waves in terms of their eigenstructures. Through theory, we demonstrate their equivalence with eigenvectors related by a two-dimensional Fourier matrix, given that the eigenvalue angles differ by modulo (integer ). We numerically verify our findings as well as demonstrate the application of the eigenstructure for the estimation of the mean inner potential.
Introduction
Modeling the three-dimensional atomic structure of a material is essential for studying its physical characteristics, like atomic electrostatic potential[1, 2, 3]. Information about the structure can significantly aid in the process of structure retrieval from experimental data such as diffraction patterns, when analyzing the material in transmission electron microscopes[4]. Determining the crystalline structure of materials from diffraction experiments is inseparable from the theory of multiple scattering of electrons, developed by Bethe[5]. Since then, numerous formulations have been developed to model multiple scattering phenomena in crystals, including the multislice method[6, 7, 8, 9] and the scattering matrix obtained from Bloch waves[10, 11]. Both models are frequently used for model-based reconstruction in transmission electron microscopy (TEM)[12, 13, 14, 15, 16, 17, 18, 19, 20].
Formulating the scattering matrix from the Bloch wave method incorporates the superposition of many periodic waves within a crystal to solve the Schrödinger equation. Therefore, the relation between electrostatic potential and Bloch waves within the crystal can be represented by an eigenvalue problem. Consequently, the scattering matrix is described by its eigenvalue decomposition, with the eigenvalues encoding parameters such as the specimen’s thickness, and the eigenvectors providing the Fourier coefficients of Bloch waves. On the other hand, the multislice method approximates the total electrostatic potential of the crystal by subdividing it into many thin slices of projected potentials. Then, the incident electron wave interacts with all slices consecutively, with the exit wave of one slice being used as the input wave for the following. A visualization of the Bloch wave and multislice approaches is given in Figure 1. An important benefit of the multislice method is that the eigenvalue decomposition can be avoided, facilitating a more efficient computation of the propagation through the crystal.
Figure 1. Comparison of the Bloch wave and the multislice approach. a, visualization of a SrTiO crystal. b, projected potential representation of the SrTiO crystal in a used in the multislice method. c, the construction of the respective scattering matrix S and, d, transmission matrix . e, f, diffraction patterns generated from matrices S and using plane wave illumination are similar, yet do not explicitly characterize the differences in eigenstructure of matrices in c and d.
The relationship between Bloch waves and the multislice method is discussed, for instance, in[21]. In several publications[22, 23, 24, 25], both models are numerically compared. Apart from the computation time, these analyses are mainly focused on the error in the intensity of the specimen exit wave functions for both methods, as this is the quantity that is finally obtained on a TEM detector.
While such comparisons are valid to capture the performance, accuracy and consistency of outputs produced by both approaches, the underlying model is treated rather as a black box system and similarity is measured only implicitly. More importantly, measuring similarity solely based on intensity captures the projection information, which does not provide insight into the spatial structure of the electrostatic potentials, neither by the Bloch wave nor the multislice approach. In our study, we therefore go a step further and, rather than relying on the projected intensity alone, directly evaluate the 3D electrostatic potential.
To compare the models, we study how both methods transform the incidence wave. Using Bloch waves, we have a direct expression in terms of the Fourier coefficients; such an expression can not be directly achieved for the multislice method. In the original formulation of the multislice method, the possibility of defining a pure transmission function, akin to the scattering matrix in the Bloch wave approach, is precluded due to the inherent entanglement between the exit wave across slices and Fresnel propagation. As a consequence, a direct comparison of the eigenstructure of the scattering matrix with the multislice method is made impossible. To circumvent this limitation, we reformulate the multislice model using matrix products and achieve a crucial separation between the input wave and the matrix that represents the transmission function of the crystal structure, which we call transmission matrix. This reformulation enables us to model the transmission function in a manner directly comparable to the scattering matrix by using eigenstructure analysis. Rather than the conventional approach of comparing resulting intensities, we present a direct mathematical comparison of the transmission matrix with the scattering matrix. Furthermore, we investigate whether the similarity between both matrices implies that also the underlying atomic structure is the same by evaluating diffraction patterns simulated by both methods. In addition to that, we leverage the determinant (the product of eigenvalues) of the transmission matrix to estimate the mean inner potential (MIP), an important physical property of the inner structure of a crystal often analyzed in TEM and particularly electron holography experiments[26, 27].
At first, we briefly introduce the scattering and transmission matrices:
Transmission Matrix.
Our formulation of the transmission matrix is derived from the multislice method. Instead of calculating an analytical solution, such as in the Bloch wave approach, the transmission function is built step-by-step using infinitely thin ‘sub-specimens’ (slices) represented by matrices. The (real-space) transmission matrix of a crystal with slices is given by
where is the diagonal matrix representing the potential of each slice, and is the Fresnel propagator in free space. The elements of both potential and Fresnel propagator matrices depend on interaction constants, spatial frequencies, and tilt angles of the incoming wave. As scattering is typically considered in terms of the Fourier coefficients, so-called reciprocal space, we transform equation (1) as
where and denote the two-dimensional Fourier and inverse Fourier transforms, respectively. The outlined transformation models the diffraction pattern in the back focal plane of a TEM.
Scattering Matrix.
The scattering matrix S describes the interaction of an input wave function with a crystal, and it can be used to determine a crystal’s transmission properties. The derivation of the scattering matrix in TEM is based on the theory of electron scattering in crystals, namely, by incorporating Bloch waves to solve the Schrödinger equation. Expanding Bloch waves into Fourier series leads to the scattering matrix of the form
where C is a unitary matrix whose columns are the eigenvectors of S, representing the Bloch wave Fourier coefficients, and is a diagonal matrix containing the corresponding eigenvalues. These eigenvalues of modulus one encode phase shifts introduced by scattering in the specimen.
Results
In this section, we discuss the structural similarity between transmission matrix and the scattering matrix S, which we investigated by leveraging their eigenstructures, i.e., their eigenvalues and eigenvectors. We discuss how this equivalency provides insights into the underlying physical processes of dynamical scattering and the accuracy of the approximations resulting from each method.
We conduct numerical simulations to generate the atomic potentials as well as scattering and transmission matrices for three different ideal crystals (GaAs, SrTiO, Au), directly from the multislice and Bloch wave models. Both matrices are then compared by analyzing the distribution of the eigenvalues in terms of plane-wave diffraction patterns, which we generate using and , the estimated projected potentials as well as MIPs.
We propose a novel approach for the estimation of the MIP facilitating the product of the eigenvalues (i.e. determinant) of the transmission matrix, which we then validate by computing the MIP of a MoS crystal from an experimental dataset obtained by four-dimensional scanning transmission electron microscopy (4D-STEM).
Characterizing Similarity.
Comparing the amplitudes of both exit waves after interaction with either the scattering or the transmission matrix is straightforward. However, it does not provide the means to investigate the (projected) potentials of a crystal solely from diffraction patterns. We, instead, propose to directly investigate the structural similarity between scattering and transmission matrix in terms of their eigenstructures, as presented in Figure 1c, Figure 1d.
Since the scattering matrix is directly defined by its eigendecomposition in equation (3), we perform the eigendecomposition of the transmission matrix , yielding the matrix of eigenvectors and the diagonal matrix containing eigenvalues. The equivalence between the scattering and the transmission matrix holds as long as the following conditions are satisfied:
Eigenvectors of scattering matrix and eigenvectors of matrix are equal to the relation and is eigenvector of real space transmission matrix ;
As the transmission matrix is unitary, its eigenvalues also have unit modulus. Thus, the eigenvalues and coincide if the respective angles differ by up to modulo for integers .
Given that and are not equal, the difference will affect the eigenvalues and eigenvectors, with the latter being less robust against errors2. Therefore, the first step of our simulation study is to compute the matrices and and their eigen decomposition for GaAs, SrTiO and Au for varying specimen thickness. We depict the Frobenius norm errors between the scattering and the transmission matrix, eigenvalues and eigenvectors in Figure 2. A very similar error of around is shown for and . Additionally, the error between transmission and scattering matrix is dominated by the error of their eigenvectors, emphasizing the sensitivity of eigenvectors to error. While the corresponding eigenvalues differ with an error of around , the error between eigenvectors remains constant at within the analyzed thickness range.
Figure 2. Frobenius norm as a function of crystal thickness, calculated using the sorted index of eigenvalues and eigenvectors. a, Frobenius norm between eigenvalues (transmission matrix) and (scattering matrix). b, Frobenius norm between eigenvectors (transmission matrix) and (scattering matrix). c, Frobenius norm between transmission matrix and scattering matrix . Results for GaAs, SrTiO and Au are plotted.
Evaluating the phases of the eigenvalues, as presented in Figure 3, shows a very similar behavior for all three datasets when comparing the eigenvalues sorted by index. While for the scattering matrix the phase increase is perfectly linear, the phase obtained using the transmission matrix shows small deviations of up to 64 mrad, which appear to be independent of index and matrix size. Overall, the eigenvalues of and are in good agreement across the analyzed range of spatial frequencies.
Figure 3. The alignment of the eigenvalue angles. Since the total of eigenvalues are the same for both transmission and scattering matrices, both axis represent the value of eigenvalue angles in radian. The x- and y-coordinates are the eigenvalue angles of the scattering and transmission matrix, respectively. a, GaAs with a thickness of 169.6 Å and a standard deviation (SD) of 0.037 rad. b, SrTiO with a thickness of 78.1 Å and a SD = 0.034 rad. c, Au with a thickness of 102 Å and a SD = 0.035 rad.
Diffraction Patterns.
The Bragg diffraction patterns of GaAs, SrTiO and Au are evaluated as a function of specimen thickness. To do so, the product of scattering or, respectively, transmission matrix with the illuminating plane wave is performed. The resulting amplitude of the computed diffraction patterns of all three specimens is shown in Figure 4.
!Plane wave diffraction patterns along the [001] zone axis for GaAs, SrTiO₃, and Au calculated using the transmission and scattering matrices bbox=0.10,0.08,0.88,0.56
Figure 4. Plane wave diffraction patterns along the [1] zone axis calculated using the transmission matrix (top row) and the scattering matrix (bottom row). a, d, GaAs with a simulated thickness of 175.2523 Å. b, e, SrTiO₃ with a simulated thickness of 126.48 Å. c, f, Au with a simulated thickness of 121.055 Å. The amplitude of the diffraction spots is given in arbitrary units. Crystalline materials are simulated with 2 × 2 unit cells with the parameters in Table II.
Comparing the results obtained with the scattering matrix to those using the transmission matrix, we find a very good agreement in peak amplitude as well as pattern structure for each specimen, indicating that any inherent differences are only marginal. To go into more detail, we evaluated the propagation of several diffracted beams with increasing crystal thickness, as shown in Figure 5. When propagating through the crystal, the transmitted and the diffracted beams exhibit alternating amplitudes depicting the Pendellösung effect. Our computed amplitudes using the scattering and transmission matrix are in good agreement, confirming our previous evaluation of the diffraction patterns. However, we observe slight differences, e.g. a small shift of maxima and minima in Figure 5a, Figure 5b, which is better visible for waves with larger oscillations.
Figure 5. Bragg beam amplitude as a function of specimen thickness, calculated using the transmission matrix (top row) and the scattering matrix (bottom row). a, d, GaAs. b, e, SrTiO₃. c, f, Au.
Projected Potential.
In the multislice method, the projected potentials are directly present in the real-space transmission matrix , as can be seen in equation (1). Therefore, for the Bloch wave method, we approximate , up to global phase factor, from the scattering matrix and obtain the projected potentials. Figure 6a-Figure 6c shows the unit cells of the simulated crystals exhibiting their atomic structure. The same structure can be observed in the projected potentials, which are depicted in Figure 7. For all crystals, the calculated projected potentials retrieved using the transmission matrix in Figure 7a-Figure 7c are structurally very similar to the potentials retrieved with the scattering matrix in Figure 7d-Figure 7f. The actual values show minor deviations, as can be seen in the difference plots in Figure 7g-Figure 7i.
Figure 6. Unit cell representations in [1] projection and perspective view of the materials used to generate the scattering and the transmission matrix. a, gallium arsenide (GaAs)[29]. b, strontium titanate (SrTiO)[30]. c, gold (Au)[29]. d, molybdenum disulfide (2H-MoS)[31] used in the 4D-STEM experiment.
Figure 7. Projected potential of 2 × 2 unit cells in [1] zone axis of, a, d, g, GaAs with a thickness of 169.6 Å (30 unit cells), b, e, h, SrTiO with a thickness of 78.1 Å (20 unit cells), and, c, f, i, Au with a thickness of 102 Å (25 unit cells). The top row is calculated with the transmission matrix and the middle row with the scattering matrix. Their absolute difference is shown in the bottom row.
As an experimental example, we use the transmission matrix to calculate the projected potential from a reconstruction of a 4D-STEM dataset of MoS. For the reconstruction, we leverage the sparse matrix decomposition algorithm[19]. The pipeline used for processing the experimental data is summarized in Figure 8. The resulting projected potential is shown in Figure 8d, where the hexagonal atomic arrangement of MoS can be seen. Experimental parameters are listed in Table III.
Figure 8. Workflow established for the evaluation of experimental diffraction data. a, 4D-STEM diffraction patterns acquired experimentally at every position of an electron transparent specimen rastered line by line using a convergent electron probe. b, transmission matrix reconstructed from the experimental data. c, estimation of the mean inner potential. d, projected potential of the specimen exit wave of a MoS sample. This output is presented as is; no aberration or focus correction was applied in the reconstruction. An atomic model of MoS is superimposed for reference.
| Parameters | Data |
| Acceleration voltage (keV) | 200 |
| Convergence semi-angle (mrad) | 32 |
| Scan dimension (pixel) | 128 × 128 |
| Diffraction pattern dimension (pixel) | 256 × 256 |
| Scan rotation | 176° |
| Real space pixel size (Å) | 0.26 |
Table III. Experimental 4D-STEM data acquisition parameters for MoS
Determinant and Mean Inner Potential.
The MIP reflects the interaction between the beam electrons and the material. It is closely linked to the inner structure of a crystal and is calculated by the average of the Coulomb potential over a unit cell,
Here, is the volume of the crystal unit cell[5] and is the total potential.
The MIP has been studied both theoretically and experimentally and was estimated for different materials in numerous works[32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46]. We estimate the MIP by employing the mathematical structure of the real space transmission matrix , as in equation (1). The determinant of , and consequently of , is given by
where, is the electron wavelength, is the total integral over reciprocal space, is the specimen thickness, and is the interaction constant. From this equation, we can extract the total potential that, in combination with the discrete grid of
Computational Complexity.
The apparent strength of the multislice over the Bloch wave method is the computational complexity. To compute the scattering matrix via equation (3), eigenvalue decomposition has to be performed, requiring operations. For the transmission matrix, we perform the sequential matrix product of Fresnel matrices and diagonal potential matrices . The products with and require and, respectively, operations. In the first case, we use that is diagonal, and in the second, we employ the fast Fourier transform. Repeating these steps for slices leads to a total complexity of operations.
pixels of known distance (step size), leads to the MIP estimate
with for an integer resulting from the periodicity of the complex exponent. In addition, is the total number of pixels, and the unit cell parameter , which is synonymous with the specimen thickness. Since equation (5) is prone to periodic artifacts, we follow an alternative route based on phase unwrapping for estimating the MIP . Our estimation is directly performed by calculating the total potential of each slice, i.e. the phase of the object of each slice , using the unwrapping technique.
The estimation of the MIP for simulated and experimental datasets is given in Table I. For comparison, results from density functional theory (DFT) calculations as well as experimental data reported in other studies is provided. By considering the determinant up to modulo for integers of , we were able to estimate the MIP using both simulated and experimental diffraction data in very good agreement with previously reported values.
Discussion
In this study, we discussed a new perspective for analyzing and understanding the relationship between the multislice and Bloch wave methods. Our presented reformulation of the multislice method into a matrix framework, which we call transmission matrix, not only enables leveraging eigenstructure analysis to measure the similarity to the scattering matrix but also allows the estimation of the MIP from diffraction data.
By expressing the multislice method as the transmission matrix, we have demonstrated that its structure parallels that of the scattering matrix. Firstly, we demonstrate mathematically that the eigenvalues of both matrices are equal up to an additive multiple of . Secondly, we prove that the eigenvectors are interconnected through the two-dimensional Fourier matrix, with the eigenvector of the transmission matrix representing the Bloch wave itself in real space.
Our matrix-based approach allows a consistent approximation of the projected potential. Moreover, our numerical evaluations of plane-wave diffraction, eigenvalues, as well as projected and MIP support these theoretical results, confirming the robustness and accuracy of the theoretical derivation of eigenstructures for both the transmission and scattering matrix.
Furthermore, the structure of the transmission matrix enables the utilization of the determinant, i.e., the product of all eigenvalues, to estimate the MIP of crystalline materials, making it a convenient tool for the evaluation of practical experiments, particularly due to its lower computational demands relative to the scattering matrix of the Bloch wave approach.
To put our results in Table I into perspective, our multislice simulations were conducted using ideal crystals without considering surface relaxations of the terminating layer at the specimen/vacuum interface, that in reality would be expected. Our estimated MIPs therefore do not account for any specific surface contributions. It is thus noteworthy that, nevertheless, our results compare well to experimentally determined values of real-world samples (), and to results of DFT calculations () previously reported. Only the estimation of the MIP of SrTiO remains a challenge which we attribute to its strong ionic character. We estimate that our presented method for determining the MIP to perform equally well or better with any experimental 4D-STEM data as well as with simulated 4D-STEM data from relaxed structures.
Table 1. Estimates of the mean inner potential from the determinant of the transmission matrix using equation (5) and mean inner potential by applying phase unwrap of each estimated slice before calculating MIP in comparison to literature values from calculated () and experimental () data. The value of , the periodicity from the complex exponent, is chosen such that is closer to . Values for and are labeled with (sim.) for simulated and (exp.) for experimental data.
Methods
Notation.
Vectors are written in bold small-cap letters and matrices are written in bold big-cap letters for a complex field and for a real field. Matrices can also be written as its element , where , . The set of integers is written as . For both matrices and vectors the notation is used to represent the element-wise or Hadamard product. The is used to represent conjugate transpose. A matrix is called unitary if , where is the identity matrix. For a matrix , the Frobenius norm is given by
Scattering Matrix.
The derivation of the scattering matrix in TEM is tied to the theory of dynamical electron scattering in crystals. We start with Klein-Gordon equation to describe the relativistic scattering of electrons in the crystal
where is the electrostatic potential at three-dimensional real space coordinates , is the relativistic electron mass, is the Planck constant, is the elementary charge, and is the relativistic wave number. Intuitively, the goal is to find proper wave functions that fulfill the Klein-Gordon equation.
As discussed in[21] (eq. 5.43), the strategy to solve the partial differential equation (6) is by involving the expansion in the Fourier space, also called reciprocal space. Firstly, the potential is expanded into its Fourier series,
where is the lattice potential in reciprocal space. Secondly, the Bloch wave is incorporated, given as
where are wave and reciprocal space vectors, respectively. Variable is the Bloch wave coefficient in reciprocal space. By rearranging this equation and writing the wave functions and potential in reciprocal space coordinates, we obtain the eigenvalue equation, the solution of which is the scattering matrix
For a complete derivation, we refer the interested readers to the literature, for instance[21, 47, 48]. Here is the structure matrix representing the electrostatic potential of the underlying crystal given by
where the diagonal and off-diagonal elements represent the excitation error and, respectively, potential in reciprocal space, and and are the reciprocal lattice vectors. This relation between the crystal and the scattering matrix is illustrated in Figure 1a, Figure 1c. Due to the nature of matrix exponentiation, both and share eigenvectors ,
Furthermore, the eigenvalues of are given as
where , are the eigenvalues of . Combined, this yields equation (3).
Transmission Matrix.
Instead of directly deriving an analytical solution as is done with the Bloch wave approach, the transmission function can be constructed using multiple infinitely thin (sub-)specimens, i.e., slices, separated by free space. This is the approach we take using the multislice method.
More precisely, let denote the potential energy of the three-dimensional crystal. The projection of onto the - plane at is expressed by . Then, the projected potentials are arranged as diagonal matrices of equation (1), each representing the -th slice. Its entries are given by
with the interaction constant for relativistic correction. For modeling the scattering process, we also account for the separation of the slices and by distance as depicted in Figure 1b. This results in a free space propagation described by the Fresnel matrix ,
where is a diagonal matrix. Its elements are
with
defined by the spatial frequencies in reciprocal space and two-dimensional tilt angles describing the direction of the incoming wave.
Then, the scattering of the incident beam within the crystal in the multislice model is obtained by consecutive multiplication with slice and Fresnel propagators . Combining all multiplications together yields the real space transmission matrix and its reciprocal space counterpart as in equations (1) and (2). For the detailed derivation, see[19].
As highlighted in[6, 21], the multislice method serves as an alternative to the Bloch wave formulation, making the transmission matrix an approximate solution to the Schrödinger equation (6). The main advantage that this reformulation offers is that the atomic structure of a crystalline object can now be represented independently by the transmission matrix , without being influenced by the electron probe, which is not the case in the classical multislice model.
Characterizing Similarity.
In the following, we compare the transmission matrix in equation (2) and scattering matrix in equation (3) in terms of their eigenstructures.
As is defined via eigenvalue decomposition in equation (3), we first establish the eigendecomposition for . By construction, we have
where matrices and are both diagonal with unit modulus entries, and are therefore unitary. Fourier transforms are also unitary and, consequently, is unitary being the product of unitary matrices. Hence, this also implies that is normal matrix, i.e., it satisfies . As a result, admits eigenvalue decomposition with a unitary matrix , which columns are the eigenvectors of , and with a diagonal matrix with eigenvalues of on its diagonal. Since is unitary, the eigenvalues are of unit modulus and can be written in terms of a complex exponential, , where for .
To quantify any differences, the scattering and the transmission matrix are compared in terms of the Frobenius norm. This norm acts as a distance metric, enabling the assessment of the accuracy and consistency between both matrices. For simplicity, we assume that the eigenvalues and eigenvectors of and are sorted in the order of increasing angles and . Then, using triangle inequality, we split the error into the eigenvector and eigenvalue errors as
We further transform the first term as
where in the second line we used the triangle inequality again and in the last line we used that Frobenius norm of the product is bounded by the product of Frobenius and spectral norms. The spectral norm is the magnitude of the largest eigenvalue and for unitary and it is equal to one. Moreover, since both and are unitary, we get , and
Consequently, the upper bound for the first term simplifies to
which is the error between the eigenvector matrices. Similarly, we bound the second term as the error on the eigenvalues,
Combined, it yields
We can further expand the eigenvalue error in terms of angles,
The equality (c) is derived from the fact that the eigenvalue of matrix is for and that matrix is a unitary matrix with modulus eigenvalues of the form for .
Determinant and Mean Inner Potential.
Apart from characterizing similarity, we can also estimate the mean inner potential (MIP) using the transmission and scattering matrices. As defined in equation (4), it requires the total potential that is estimated using the determinant of . More precisely, we compute as
The equality holds because and the determinant of a matrix product is equal to the product of the determinants of the individual factors. Using that the determinant of diagonal matrix is the product of the elements, the determinant of products
where we used notation
Returning to equation (11), the logarithm of the determinant of is
where is the total thickness of the crystal and with integer results from periodicity of complex logarithm.
The sum of potentials is linked to the total potential
by the Riemann sum approximation. Following equation C.20 in Kirkland, 1989[49], the inner integral is precisely . Hence, we can write
where and are the height, width and thickness of the unit cell in Ångstrom. With the unit cell volume in cubic Ångstrom, the MIP is
By combining equations (12) and (14), we obtain
and rearranging yields the estimate in equation (5).
Mean Inner Potential and Unwrapping.
An alternative way to compute the MIP is by directly using the slices . From equation (7), we observe that the phases of are wrapped periodically due to the complex exponent. Thus, can be recovered from by phase unwrapping[50] up to an additive constant ,
In crystallographic applications, the constant can be further estimated by assuming that the most prevalent part of the observed object is vacuum, which does not cause scattering. Hence, the minimal value of the potential is almost zero, leading to
and
We can then use the above values to compute the total inner potential via equation (13).
Transmission Matrix Reconstruction.
In this section, we describe the approach used to estimate the transmission matrix from experimental 4D-STEM data of a MoS crystal. Let denote the collection of diffraction patterns obtained by illuminating the crystal with focused electron beams for a total scanned specimen positions. Flattening the dataset as and combining the probes as allows us to rewrite the reconstruction as a matrix estimation problem
subject to
where the constraint is precisely equation (1). To solve equation (15), we employ the sparse matrix decomposition algorithm[19] which alternates between minimizing the objective for arbitrary and projecting it to satisfy equation (1). The reconstruction is performed with 3 slices with a total thickness of approximately 4.6722 nm. Since we have a total of scan pixels with a distance between adjacent scan positions of 0.026 nm/pixel, the scan area results to nm. These parameters are used to calculate the MIP of MoS form experimental data given in Table I as well as the projected potential in Fig. 8d.
Data availability
The data used to produce the results presented in this paper are available at Zenodo: https://zenodo.org/records/14646184. In the following, we give the parameters of both simulated and experimental datasets.
Simulated Datasets.
Three crystalline materials are used: gallium arsenide (GaAs), a binary compound semiconductor, strontium titanate (SrTiO), a ternary oxide, and gold (Au), a single element crystal.
These materials differ in terms of total number and type (atomic number Z) of elements, and cubic unit cell dimension. Representations of the unit cells are shown in Figure 6. The parameters used to generate the simulated datasets are summarized in Table II, outlining crucial characteristics such as unit cell dimensions, pixel size, and Fresnel distance for each material. These parameters serve as input for the simulation of the electrostatic potentials and the generation of the structure matrices.
| Parameters | GaAs (F43m) | SrTiO (Pm3m) | Au (Fm3m) |
| Unit cell dimensions (a, b, c) (Å) | (5.6533, 5.6533, 5.6533) | (3.905, 3.905, 3.905) | (4.08, 4.08, 4.08) |
| Unit cell angles (, , ) (°) | (90, 90, 90) | (90, 90, 90) | (90, 90, 90) |
| Pixel size (Å) | (0.1739, 0.1739) | (0.1201, 0.1201) | (0.1255, 0.1255) |
| Fresnel distance (Å) | (1.413325, 1.413325) | (1.9525, 1.9525) | (2.04, 2.04) |
| e acceleration voltage (keV) | 300 | 300 | 300 |
Table II. Simulation parameters used to generate scattering and transmission matrix with both the Bloch wave and the multislice method, respectively. Structure data for GaAs[29], SrTiO[30] and Au[29] was used. In the multislice data, each slice comprises the projected potential of all atoms within a single atomic layer. The Fresnel distance is the distance which the projected potential of a single slice is required to be propagated and corresponds to the lattice spacing along [1].
Our implementation of the multislice and the Bloch wave method, as well as the parameters used to generate the transmission and the scattering matrices are based on the MATLAB code by Durham et al., 2022[25].
Experimental Dataset.
For the experimental process, 2H-MoS sheets were exfoliated from a bulk crystal using a poly-dimethylsiloxane elastomer film placed on a glass slide and subsequently transferred onto a holey silicon nitride membrane for TEM analysis. The experimental data was collected using a probe-corrected Hitachi HF5000 field emission microscope in scanning TEM mode (STEM), operating at an acceleration voltage of 200 keV with a beam current of approximately 7.4 pA. The intensities of the diffraction patterns were recorded with a Medipix3 Merlin4EM camera, with a resolution of pixels. The distance between adjacent scan points was set to 26 pm in both the horizontal (x) and vertical (y) directions for the acquisition of an array of diffraction patterns. Each diffraction pattern was acquired with 6-bit dynamic range within 0.5 ms.
Code availability
The source code to reproduce the result in this paper is available at https://github.com/bangunayra/eigenstructure_bloch_ms and https://github.com/bangunayra/inverse_multislice.
Acknowledgements
We would like to thank Prof. K.-M. Caspary for making possible the 4D-STEM experiment at ER-C-1, Forschungszentrum Jülich.
Author contributions statement
A.B. conceptualized the study, wrote the original draft of the paper, implemented the Python code for generating and eigenstructure analysis of transmission and scattering matrix, as well as for simulating diffraction data. O.M. developed theoretical bound of differences between transmission and scattering matrix, as well as contributed to the mean inner potential (MIP) formulation for both logarithmic and unwrap version. B.M. conducted the 4D-STEM experiment and contributed to the code review, data analysis and interpretation. All authors discussed the results, wrote and reviewed the manuscript.
Supplementary Information
We describe the original multislice formulation, representation of the Fourier transform as a matrix, as well as additional numerical experiments.
Multislice Method.
We present the classical formulation of the multislice method and briefly present the derivation of an alternative formulation based on matrices. The complete derivation is given in Bangun et. al., 2022[19].
Suppose we have a matrix that represents the –th slice of the electrostatic potential of the specimen for a thin slice, similar to equation (7).
In fact, for each combination axis we collect the potential and construct it into matrix for a total slice number of . The interaction between the first slice and the electron wave for focused beam illumination at a specific scanning point is given by
The exit wave is used as the source for the next slice by propagation using the Fresnel propagator ,
The Fresnel propagator is given in Fourier space for a specific distance between slices. We define the Fourier space of the Fresnel propagator matrix . Its elements are given in equation (9). Hence, the Fresnel propagation can be written as
As a result, the intensity recorded at a specific scanning point in the microscope’s detector plane is given as
where the absolute value is applied to each element of the matrix.
It can be seen that, a separation of the entanglement between the electron wave and the object is difficult. Hence, an exact representation of the transmission function of the object, independent of the propagated exit wave starting from the probe illumination, is not available. As discussed in Bangun et. al., 2022[19], the reformulation by leveraging the property of element-wise or Hadamard matrix product, namely
where the function is used to convert a matrix into a vector and is the function to construct a diagonal matrix from a vector. The vectorized intensity at scanning point can be written as
Using this formulation, we can disentangle the illuminating electron wave and the object in order to get a pure transmission function
where the formulation of Fresnel matrix is derived in equation (8).
Matrix Representation of Fourier Transforms.
We present a matrix representation of Fourier transforms. The one-dimensional Fourier transform can be written as a matrix-vector product. Here, the term one-dimensional Fourier matrix represents the discrete implementation of the Fourier basis, i.e., when we sample the Fourier basis and store it as a matrix . Hence, the one-dimensional Fourier transform of vector can be written as
where the one-dimensional Fourier transform is given by[51] (eq. 5.44),
Here, for represents the sampling points at evenly spaced real space coordinates, and for is the sample in the Fourier space. Along the same line, the two-dimensional Fourier matrix can be constructed using the Kronecker product between two of the one-dimensional Fourier matrices . The two-dimensional Fourier transform of matrix can be written as .
Additional Numerical Experiments.
In order to highlight the consistency of results that are obtained using scattering and transmission matrix, when applied to data of specimens with highly different thickness, we present additional numerical experiments, that are, diffraction patterns shown in Figure 11, phase of eigenvalues as in Figure 9, as well as projected potential as in Figure 10.
Figure 9. Phase distribution of eigenvalues obtained from transmission matrix and scattering matrix for two different thicknesses . a, GaAs of Å (SD = 0.098 rad). d, GaAs of Å (SD = 0.043 rad). b, SrTiO of Å (SD = 0.068 rad). e, SrTiO of Å (SD = 0.0384). c, Au of Å (SD = 0.06 rad). f, Au of Å (SD = 0.05 rad).
Figure 10. Projected potential of unit cells in [1] zone axis using the transmission matrix (1st, 3rd and 5th column) and the scattering matrix (2nd, 4th and last column) for different thicknesses . a, GaAs with thickness 141.33 Å (25 unit cells) and, d, 197.87 Å (35 unit cells). b, SrTiO₃ with thickness 58.58 Å (15 unit cells) and, e, 78.1 Å (20 unit cells). c, Au with thickness 20.4 Å (5 unit cells) and, f, 40.8 Å (10 unit cells).
Figure 11. Diffraction patterns along the [1] zone axis calculated using the transmission matrix (1st and 3rd row) and the scattering matrix (2nd and last row) for different thicknesses . a, GaAs with Å and, d, Å. b, SrTiO₃ with simulated thickness Å and, e, Å. c, Au with simulated thickness Å and, f, Å.
References
- [1]B. G. Mendis, “A physical optics formulation of Bloch waves and its application to 4D STEM, 3D ED and inelastic scattering simulations,” Foundations of Crystallography 81 (2025).DOI
- [2]C. Doberstein and P. Binev, “Lattice multislice algorithm for fast simulation of scanning transmission electron microscopy images,” Microscopy and Microanalysis, ozae116 (2025).arxiv.org/abs/2310.16829
- [3]B. Mendis, “Modelling dynamical 3D electron diffraction intensities. I. A scattering cluster algorithm,” Foundations of Crystallography 80, 167–177 (2024).DOI
- [4]B. Diederichs, Z. Herdegen, A. Strauch, F. Filbir, and K. Müller-Caspary, “Exact inversion of partially coherent dynamical electron scattering for picometric structure retrieval,” Nature Communications 15, 101 (2024).DOI
- [5]5H. Bethe, “Theorie der Beugung von Elektronen an Kristallen,” Annalen der Physik 392, 55–129 (1928).DOI
- [6]6J. M. Cowley and A. F. Moodie, “The scattering of electrons by atoms and crystals. I. A new theoretical approach,” Acta Crystallographica 10, 609–619 (1957).DOI
- [7]7K. Ishizuka and N. Uyeda, “A new theoretical and practical approach to the multislice method,” Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography 33, 740–749 (1977).DOI
- [8]8D. V. Dyck, “Fast computational procedures for the simulation of structure images in complex or disordered crystals: a new approach,” Journal of Microscopy 119, 141–152 (1980).
- [9]9P. Goodman and A. F. Moodie, “Numerical evaluations of N-beam wave functions in electron scattering by the multi-slice method,” Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography 30, 280–290 (1974).DOI
- [10]10L. Sturkey, “The calculation of electron diffraction intensities,” Proceedings of the Physical Society 80, 321 (1962).DOI
- [11]11F. Fujimoto, “Dynamical theory of electron diffraction in Laue-case, I. General theory,” Journal of the Physical Society of Japan 14, 1558–1568 (1959).DOI
- [12]12L. J. Allen, H. Leeb, and A. Spargo, “Retrieval of the projected potential by inversion from the scattering matrix in electron–crystal scattering,” Acta Crystallographica Section A: Foundations of Crystallography 55, 105–111 (1999).pubmed.ncbi.nlm.nih.gov/10927239
- [13]13J. C. Spence, B. Calef, and J.-M. Zuo, “Dynamic inversion by the method of generalized projections,” Acta Crystallographica Section A: Foundations of Crystallography 55, 112–118 (1999).pubmed.ncbi.nlm.nih.gov/10927240
- [14]14A. M. Maiden, M. J. Humphry, and J. M. Rodenburg, “Ptychographic transmission microscopy in three dimensions using a multi-slice approach,” Journal of the Optical Society of America A 29, 1606–1614 (2012).DOI
- [15]15H. G. Brown, Z. Chen, M. Weyland, C. Ophus, J. Ciston, L. J. Allen, and S. D. Findlay, “Structure retrieval at atomic resolution in the presence of multiple scattering of the electron probe,” Physical Review Letters 121, 266102 (2018).DOI
- [16]16J. J. Donatelli and J. C. Spence, “Inversion of many-beam Bragg intensities for phasing by iterated projections: Removal of multiple scattering artifacts from diffraction data,” Physical Review Letters 125, 065502 (2020).DOI
- [17]17S. D. Findlay, H. G. Brown, P. M. Pelz, C. Ophus, J. Ciston, and L. J. Allen, “Scattering matrix determination in crystalline materials from 4D scanning transmission electron microscopy at a single defocus value,” Microscopy and Microanalysis 27, 744–757 (2021).DOI
- [18]18Z. Chen, Y. Jiang, Y.-T. Shao, M. E. Holtz, M. Odstrčil, M. Guizar-Sicairos, I. Hanke, S. Ganschow, D. G. Schlom, and D. A. Muller, “Electron ptychography achieves atomic-resolution limits set by lattice vibrations,” Science 372, 826–831 (2021).DOI
- [19]19A. Bangun, O. Melnyk, B. März, B. Diederichs, A. Clausen, D. Weber, F. Filbir, and K. Müller-Caspary, “Inverse multislice ptychography by layer-wise optimisation and sparse matrix decomposition,” IEEE Transactions on Computational Imaging 8, 996–1011 (2022).DOI
- [20]20A. Sadri and S. D. Findlay, “Determining the projected crystal structure from four-dimensional scanning transmission electron microscopy via the scattering matrix,” Microscopy and Microanalysis 29, 967–982 (2023).DOI
- [21]21M. De Graef, Introduction to conventional transmission electron microscopy (Cambridge University Press, 2003).
- [22]22Y. Yang, Q. Yang, J. Huang, C. Cai, and J. Lin, “Quantitative comparison between real space and Bloch wave methods in image simulation,” Micron 100, 73–78 (2017).DOI
- [23]23P. G. Self, M. A. O’keefe, P. R. Buseck, and A. Spargo, “Practical computation of amplitudes and phases in electron diffraction,” Ultramicroscopy 11, 35–52 (1983).DOI
- [24]24C. Koch and J.-M. Zuo, “Comparison of multislice computer programs for electron scattering simulations and the Bloch wave method,” Microscopy and Microanalysis 6, 126–127 (2000).DOI
- [25]25D. B. Durham, C. Ophus, K. M. Siddiqi, A. M. Minor, and D. Filippetto, “Accurate quantification of lattice temperature dynamics from ultrafast electron diffraction of single-crystal films using dynamical scattering simulations,” Structural Dynamics 9 (2022).DOI
- [26]26F. Winkler, A. H. Tavabi, J. Barthel, M. Duchamp, E. Yucelen, S. Borghardt, B. E. Kardynal, and R. E. Dunin-Borkowski, “Quantitative measurement of mean inner potential and specimen thickness from high-resolution off-axis electron holograms of ultra-thin layered {WSe₂},” Ultramicroscopy in press, 1–10 (2016).DOI
- [27]27T. Lorenzen, B. März, T. Xue, A. Beyer, K. Volz, T. Bein, and K. Müller-Caspary, “Imaging built-in electric fields and light matter by fourier-precession tem,” Scientific Reports 14, 1320 (2024).DOI
- [29]29R. Wyckoff, Crystal Structures, Vol. 2 (Interscience Publishers, New York, 1963) pp. 85–237.
- [30]30R. H. Mitchell, A. R. Chakhamouradian, and P. M. Woodward, “Crystal chemistry of perovskite-type compounds in the tausonite-loparite series,(Sr1−2xNaxLax)TiO3,” Physics and Chemistry of Minerals 27, 583–589 (2000).
- [31]31B. Schönfeld, J. J. Huang, and S. C. Moss, “Anisotropic mean-square displacements (MSD) in single-crystals of 2H- and 3R-MoS₂,” Acta Crystallographica Section B 39, 404–407 (1983).DOI
- [32]32P. Kruse, A. Rosenauer, and D. Gerthsen, “Determination of the mean inner potential in III–V semiconductors by electron holography,” Ultramicroscopy 96, 11–16 (2003).pubmed.ncbi.nlm.nih.gov/12623168
- [33]33P. Kruse, M. Schowalter, D. Lamoen, A. Rosenauer, and D. Gerthsen, “Determination of the mean inner potential in III–V semiconductors, Si and Ge by density functional theory and electron holography,” Ultramicroscopy 106, 105–113 (2006).pubmed.ncbi.nlm.nih.gov/16125323
- [34]Gaidardziska-Josifovska, M. R. McCartney, W. De Ruijter, D. J. Smith, J. Weiss, and J.-M. Zuo, “Accurate measurements of mean inner potential of crystal wedges using digital electron holograms,” Ultramicroscopy 50, 285–299 (1993).
- [35]N. Yamamoto and J. C. Spence, “Surface imaging of III–V semiconductors by reflection electron microscopy and inner potential measurements,” MRS Online Proceedings Library (OPL) 18, 43 (1982).
- [36]C. A. Mizzi and L. D. Marks, “The role of surfaces in flexoelectricity,” Journal of Applied Physics 129, 202101 (2021).DOI
- [37]H. Sun, Z. Mao, T. Zhang, L. Han, T. Zhang, X. Cai, X. Guo, Y. Li, Y. Zang, W. Guo, et al., “Chemically specific termination control of oxide interfaces via layer-by-layer mean inner potential engineering,” Nature Communications 9, 2965 (2018).DOI
- [38]A. Auslender, N. Pandey, A. Kohn, and O. Diéguez, “Mean inner potential of elemental crystals from density-functional theory calculations: Efficient computation and trends,” Ultramicroscopy 255, 113862 (2024).DOI
- [39]V. Tull, “Some experimental measurements of the inner potentials of various crystals,” Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences 206, 219–232 (1951).DOI
- [40]R. Buhl, “Interferenzmikroskopie mit Elektronenwellen,” Zeitschrift für Physik 155, 395–412 (1959).DOI
- [41]M. Keller, “Ein Biprisma-Interferometer für Elektronenwellen und seine Anwendung,” Zeitschrift für Physik 164, 274–291 (1961).DOI
- [42]E. Kerschbaumer, “Ein Biprisma-Interferometer für 100 keV Elektronen und seine Anwendung,” Zeitschrift für Physik 201, 200–208 (1967).DOI
- [43]A. Goswami and N. Lisgarten, “The measurement of inner potentials for copper, silver and gold,” Journal of Physics C: Solid State Physics 15, 4217 (1982).DOI
- [44]K. Okazaki, S. Ichikawa, Y. Maeda, M. Haruta, and M. Kohyama, “Electronic structures of Au supported on TiO₂,” Applied Catalysis A: General 291, 45–54 (2005).DOI
- [45]R. Popescu, E. Müller, M. Wanner, D. Gerthsen, M. Schowalter, A. Rosenauer, A. Böttcher, D. Löffler, and P. Weis, “Increase of the mean inner coulomb potential in au clusters induced by surface tension and its implication for electron scattering,” Physical Review B – Condensed Matter and Materials Physics 76, 235411 (2007).DOI
- [46]V. Boureau, B. Sklenard, R. McLeod, D. Ovchinnikov, D. Dumcenco, A. Kis, and D. Cooper, “Quantitative mapping of the charge density in a monolayer of MoS₂ at atomic resolution by off-axis electron holography,” ACS nano 14, 524–530 (2019).DOI
- [47]A. J. F. Metherell, “Diffraction of electrons by perfect crystals,” Electron Microscopy in Materials Science, part 2 397 (1976).
- [48]C. J. Humphreys, “The scattering of fast electrons by crystals,” Reports on Progress in Physics 42, 1825 (1979).DOI
- [49]E. J. Kirkland, Advanced computing in electron microscopy, Vol. 12 (Springer, 1998).
- [50]M. A. Herráez, D. R. Burton, M. J. Lalor, and M. A. Gdeisat, “Fast two-dimensional phase-unwrapping algorithm based on sorting by reliability following a noncontinuous path,” Applied optics 41, 7437–7444 (2002).pubmed.ncbi.nlm.nih.gov/12502301
- [51]A. K. Jain, Fundamentals of digital image processing (Prentice-Hall, Inc., 1989).