betaFeedback

WorksPhysics and Astronomy

Learning Interatomic Force Coefficients from X-ray Thermal Diffuse Scattering Data

Abstract

We present a fully automated framework for extracting interatomic force constants (IFCs) directly from X-ray thermal diffuse scattering (TDS) data. By formulating scattering intensity as a differentiable function of a symmetry-reduced IFC parameterization, we enable gradient-based optimization via direct, Cholesky-based sampling of correlated atomic displacements at thermal equilibrium. This approach bypasses the computational bottleneck of repeated Hessian matrix diagonalizations, significantly accelerating the inversion process. Benchmark tests demonstrate that the framework accurately recovers ground-truth IFCs and phonon dispersion relations, providing a robust, high-throughput pathway for studying lattice dynamics across diverse crystalline materials. This method bridges the gap between experimental observations and computational modeling, enabling the direct integration of TDS data into the refinement of high-fidelity inter-atomic potentials.

Introduction

The interactions between atoms are fundamental to all properties of solid-state materials, ranging from elasticity and thermal conductivity to emergent phenomena like charge density waves and superconductivity [1, 2, 3, 4, 5, 6]. For crystalline materials, these interactions can be formally quantified by a set of interatomic force constants (IFCs), also known as Born-von Kármán force constants [1, 7]. While IFCs are traditionally derived from computational models spanning a range of sophistication—from first-principles density functional theory [8] to machine-learned empirical potentials [9]—extracting them directly from experimental measurements is essential. Such direct inversion provides the definitive reference data required to validate and refine theoretical models against physical reality.

The significance of IFCs also stems from their direct mathematical connection to the phonon spectrum, which represents the collective vibrations of the atoms in the crystal lattice. Since phonons scatter incident radiation across a wide range of probes (e.g., X-rays, neutrons, electrons), they directly link IFCs to experimentally measurable quantities [10]. Obtaining quantitative, experimentally-derived information on IFCs is crucial for advancing our understanding of lattice dynamics, improving predictive atomistic simulations, and enabling the knowledge-driven design of new functional materials.

Despite their central role in materials physics, the accurate and systematic determination of IFCs directly from experiments remains a challenge. Experimental access to phonon dispersions and thereby to the IFCs, is commonly sought using Raman spectroscopy, inelastic neutron scattering (INS) [11, 12], and inelastic X-ray scattering (IXS) [13, 14]. However, each method has its own limitations.

Raman spectroscopy is inherently restrictive, as it probes only zone-center phonons and only those that are Raman-active, leaving most of the Brillouin zone inaccessible. INS offers the desired full-zone phonon dispersions, but it typically requires large single crystals, involves long measurement times, and requires access to specialized, limited-availability neutron sources [10]. IXS offers data richness comparable to INS, however, it suffers from intrinsically low flux, limited reciprocal-space coverage, and point-by-point mapping, requiring extensive sample rotation and energy scans. As a result, full Brillouin-zone phonon mapping using IXS is extremely time-consuming and often impractical [10, 15].

Thermal diffuse scattering (TDS) captures the inelastic scattering occurring away from the Bragg peaks. This scattering arises from the instantaneous atomic displacements in a thermally vibrating crystal (see Figure 1). Consequently, TDS provides a continuous map of phonon intensities across the entire reciprocal space, encompassing both acoustic and optical modes. These unique characteristics allows the TDS to serve as a robust probe of the IFCs. The utility of phonon scattering for validating machine-learned interatomic potentials (MLIP) has recently been demonstrated using electron TDS [16, 17, 18]. X-ray TDS (X-TDS) extends this potential by leveraging the versatility and accessibility of modern X-ray facilities. Unlike INS or IXS, X-TDS measurements can be acquired rapidly over vast regions of reciprocal space and can be readily performed under diverse in situ conditions (e.g. varying temperature, pressure, or illumination). However, extracting information on IFC embedded in the X-TDS data is not trivial; it requires sophisticated modeling and computational inversion to decode the atomic interactions from the scattering signal.

Schematic representation of thermal diffuse scattering

Figure 1. Schematic representation of thermal diffuse scattering. In the absence of atomic vibrations, elastic scattering leads to sharp Bragg peaks in the scattering pattern. Atomic vibrations introduce diffuse signal around the Bragg peaks due to inelastic scattering. The specific pattern of the diffuse scattering signal reflects the correlation of atomic vibrations governed by interatomic force, making TDS a direct probe of the interatomic force constants (IFCs) in materials.

Computational X-TDS images are typically generated from atomic trajectories via molecular dynamics (MD) simulations [19]. Because X-TDS is a time-averaged quantity, producing a statistically reliable image requires MD trajectories consisting of millions of time steps to sample a sufficient number of independent configurations. This requirement for statistical rigor renders the direct MD approach computationally prohibitive, particularly when address the inverse problem: the quantitative extraction of IFCs from experimental data. An efficient inversion would typically require a gradient-based optimization algorithm, which requires calculating the gradients of the X-TDS intensity with respect to the IFC parameters. Using automatic differentiation (autograd) in this context would necessitate backpropagating derivatives through a multi-million-step trajectory, which is computationally infeasible. Alternatively, estimating the gradient via the finite-difference method introduces significant numerical noise and scales poorly with the number of IFC parameters. As a result, MD-based X-TDS simulations are primarily confined to generating images for qualitative comparison with experiments, making the method unsuitable for the task of quantitative extraction of IFCs from experimental X-TDS data.

To address the inverse problem, a quantitative X-TDS analysis has been developed by Holt et al. [20, 21, 22]. This approach is based on an analytical derivation of the scattering intensities from the eigenvalues of the Hessian matrix. It has been demonstrated for selected model materials like Ni, Si, Nb, PuGa TiSe2, and MgO [20, 21, 22, 23, 24, 25]. However, this eigenvalue-based approach faces intrinsic limitations: it requires repeated diagonalization of the dynamic matrix, resulting in a computationally costly cubic scaling with system size. Furthermore, the model treats phonon modes as independent entities, which complicate the inclusion of higher-order scattering effects (like phonon-phonon scattering) that can contribute up to 40% of the total intensity at elevated temperatures [26, 23]. While Wehinger et al. [27] developed an alternative approach to obtain the elastic tensor from the TDS successfully, it has not led to the extraction of the entire interatomic force constant matrix. Collectively, these limitations have prevented X-TDS from becoming a routine quantitative tool for recovering IFCs. Despite its remarkable experimental accessibility and information content, experimental TDS images are thus largely analyzed qualitatively, by visually comparing broad qq-space trends to other experimental methods such as INS or computational X-TDS images [28].

A fast, robust, and general method capable of extracting IFCs from X-TDS data would provide direct experimental access to the fundamental forces governing atomic interactions in solids. This information is essential to the understanding of a material’s intrinsic dynamics and response to external stimuli. To enable the direct extraction of the IFCs from X-TDS data, we develop a fully differentiable end-to-end framework connecting IFCs to the X-TDS images. Our approach combines and builds on three key advances. First, we develop an algorithm to utilize the crystal’s space group symmetry to construct a symmetry-reduced Born-von Kármán parametrization of the Hessian, yielding a minimal and physically consistent set of independent IFCs. These ensure both physical consistency and reduction of computational complexity. Second, we evaluate the TDS without the phonon eigen-modes by directly sampling correlated thermal displacements from the Hessian via a differentiable Cholesky factorization and computing scattering intensities using FFT-based methods. Third, this differentiable forward model enables automatic gradient computation with respect to the IFCs. This strategy eliminates the need for repeated dynamical-matrix diagonalizations and enables efficient, gradient-based refinement of interatomic force constants from X-TDS. Furthermore, this method paves the way for routine, quantitative X-TDS analysis for both fundamental phonon studies and data-driven materials modeling [23, 29].

The remainder of this paper is structured as follows. In § 2, we describe our computational method, including the complexity reduction techniques using crystal symmetry, the direct sampling approach for generating atomic displacements, and the resulting gradient descent algorithm for optimizing the IFCs. § 3 presents the results of our method applied to benchmark cases, demonstrating convergence and accuracy in IFC extraction. The implications of the results and limitations of the approach are discussed in § 4. Finally, in § 5, we summarize our key findings and outline potential avenues for further development and applications for this computational framework.

Methods

To enable efficient and quantitative extraction of the IFCs directly from measured TDS data, we reformulate the forward scattering model (from the IFCs to TDS images) to be both fully differentiable and computationally efficient. This differentiability is the key innovation: it allows errors between simulated and target intensities to be back-propagated to enable gradient-based optimization.

The optimization workflow is illustrated in Figure 2. We start by initializing the values of a symmetryreduced minimal set of IFCs (ϕ\phi) that describes the atomic interactions within a chosen cut-off radius. This minimal set is then used to explicitly and linearly construct the entire Hessian matrix (H\mathbf{H}) for a crystal supercell through an inherently differentiable mapping. The next critical step involves efficiently and differentially generating a set of random atomic displacements that satisfy the Boltzmann distribution dictated by the Hessian matrix. This is accomplished through a Cholesky decomposition of the Hessian matrix, which ensures both efficiency and differentiability. Subsequently, the X-ray scattering intensity image is computed for each generated atomic configuration using the Fast Fourier Transform (FFT) approach [30]. The average of these intensities yields the computational prediction I(q)I(\mathbf{q}) for the X-TDS image corresponding to the current IFC values. This predicted image is then compared with the experimentally measured X-TDS image, and the loss is computed as the root mean square error. The derivatives of the loss function with respect to the IFCs are calculated through automatic differentiation (autograd). Finally, a gradient-descent algorithm uses these derivatives to update the IFC values. This entire cycle constitutes one epoch of the training process, which is typically repeated for thousands of epochs until the IFCs converge to the optimal target values.

Schematic of the differentiable optimization loop

Figure 2. Schematic of the differentiable optimization loop. The forward pass (bold arrows) generates predicted TDS intensities through stochastic displacement sampling and FFT-based scattering calculations. During the backward pass (thin arrow), gradients of the loss function are propagated to update the parameters ϕ\phi iteratively.

Minimal IFC Set

As a concrete example, we consider computational X-TDS from an atomistic model of face-centered cubic (FCC) Ni. We choose a supercell of size 5[1 0 0]×5[0 1 0]×5[0 0 1]5[1\ 0\ 0] \times5[0\ 1\ 0] \times5[0\ 0\ 1], which contains N=500N = 500 atoms and is subjected to periodic boundary conditions in all three directions.

This large supercell size is chosen to minimize finite-size effects. A smaller simulation cell results in a much coarser resolution in the q\mathbf{q}-space over which the X-TDS images will be computed. Let U({xi})U(\{x_i\}) be the interatomic interaction potential function we wish to learn from measured TDS images, where x≡{xi}\mathbf{x} \equiv\{x_i\}, i=1⋯3Ni = 1 \cdots3N is the array of all atomic coordinates. We can Taylor expand U({xi})U(\{x_i\}) around the perfect crystal positions xi0x_i^0, for which the forces on all atoms are zero.

U({xi})=U({xi0})+12∑i,jHij δxi δxj+O({δxi})3,(1)U(\{x_i\}) = U(\{x_i^0\}) + \frac{1}{2} \sum_{i,j} H_{ij}\,\delta x_i\,\delta x_j + \mathcal{O}(\{\delta x_i\})^3, \tag*{(1)}

where δxi=xi−xi0\delta x_i = x_i - x_i^0 is the atomic displacements from the perfect crystal positions and HijH_{ij} are the second-order partial derivatives of the potential,

Hij≡∂2U∂xi∂xj∣{xi=xi0}.(2)H_{ij} \equiv\left.\frac{\partial^2 U}{\partial x_i \partial x_j}\right|_{\{x_i=x_i^0\}}. \tag*{(2)}

At temperatures much lower than the melting temperature, the atomic displacements {δxi}\{\delta x_i\} are sufficiently small that the higher-order terms, O({δxi})3\mathcal{O}(\{\delta x_i\})^3, in the Taylor expansion are negligible. By truncating the expansion after the second-order term, we employ the harmonic approximation (HA), a standard approach in TDS analysis. The coefficients HijH_{ij} are also called the Born-von Kármán force constants. These coefficients collectively form the 3N×3N3N \times3N Hessian matrix H={Hij}\mathbf{H} = \{H_{ij}\}. This Hessian matrix completely describes the interatomic forces within the HA. The central goal of our TDS image analysis is to learn the force constants, hence the Hessian matrix, from the experimental data.

Illustration of the IFC hierarchy for the first five neighbor shells

Figure 3. Illustration of the IFC hierarchy for the first five neighbor shells. The 0-th shell (a0a_0) corresponds to the self-interaction term of the Hessian matrix. Coefficients bkb_k, ckc_k, dkd_k, eke_k, and fkf_k correspond to the force constants for the 1st, 2nd, 3rd, 4th, and 5th coordination shells, respectively.

A first challenge one encounters in learning the Hessian matrix is the large number of its components. In our current example of N=500N = 500 atoms, the Hessian matrix contains 2,250,000 components, which is too many to deal with. Fortunately, the number of independent matrix components is drastically smaller due to the symmetry inherent in the crystal structure.

First, FCC Ni has a simple crystal structure, with only one atom per unit cell. Consequently, all atoms are symmetry-equivalent due to the crystal’s translational symmetry. This means the relationship between any two atoms, kk and ll, is identical to the relationship between any other two atoms, mm and nn, which are separated by the same lattice vector. Specifically, the translational symmetry implies that the 3×33 \times3 sub-block of the Hessian matrix relating atoms kk and ll only depends on the difference in the equilibrium positions between these two atoms. Therefore, if we determine the force constants between the first atom (atom 0) and all other atoms in the supercell, we can construct all the remaining components of the Hessian matrix. This allows us to focus our learning efforts exclusively on the 9N9N components of the first three rows of the Hessian matrix, HijH_{ij} (where i=0,1,2i = 0,1,2, j=1,…,3Nj = 1,\ldots,3N), which contains 4,500 components.

Without loss of generality, we choose the perfect crystal position of atom 0 to be the origin of the coordinate system. To systematically organize the force constants, we sort all atoms into shells based on their distance from atom 0. For example, atom 0 itself is referred to as shell 0. The 12 nearest neighbor atoms of atom 0 form shell 1; the 6 second nearest neighbor atoms form shell 2, etc. We expect the strength of the interatomic interaction, hence the magnitude of the Hessian matrix elements HijH_{ij}, to rapidly decay as the distance between the corresponding atom kk and atom 0 increases. This justifies the use of a cut-off distance beyond which the force constants are approximated as zero. In this example, we assume all Hessian matrix elements to be zero beyond the fifth nearest neighbor shell. As a result, the number of Hessian matrix elements that we still need to consider is now limited to 711. We labeled the force constants as aka_k, bkb_k, ckc_k, dkd_k, eke_k, fkf_k, corresponding to shell 0, 1, 2, 3, 4, 5, respectively.

The matrix elements corresponding to the same shell are related to each other through the rotational symmetries of the crystal’s point group. For the FCC crystal structure, the corresponding point group is OhO_h. For example, the 3×33 \times3 sub-block of the Hessian matrix, corresponding to shell 0, represents the restoring force when atom 0 is displaced. The OhO_h group requires that this matrix to be isotropic, i.e. a scalar times the identity matrix. Hence, the 9 matrix elements for shell 0 contain only one independent constant, a0a_0. It can be shown that there are only three independent constants, {b0,b1,b2}\{b_0,b_1,b_2\}, for shell 1; two independent constants, {c0,c1}\{c_0,c_1\}, for shell 2; four independent constants, {d0,d1,d2,d3}\{d_0,d_1,d_2,d_3\} for shell 3; three independent constants, {e0,e1,e2}\{e_0,e_1,e_2\} for shell 4; and four independent constants, {f0,f1,f2,f3}\{f_0,f_1,f_2,f_3\}. Thus the point group Oh\mathrm{O}_{h} reduced the number of independent force constants to 17.[31]

There is one more symmetry satisfied by the Hessian matrix beyond the ones considered above. Because the potential energy UU only depends on the relative positions of the atoms with respect to each other, no force is produced on any atoms if all atoms are subjected to a rigid-body translation from their perfect crystal positions. As a result, every row of the HijH_{ij} matrix sum up to zero. In the FCC example considered here, this leads to one more constraint on the force constants.

a0+8b0+4b1+2c0+4c1+8d0+16d1+8e0+4e1+8f0+8f1+8f2=0.(3)a_0 + 8b_0 + 4b_1 + 2c_0 + 4c_1 + 8d_0 + 16d_1 + 8e_0 + 4e_1 + 8f_0 + 8f_1 + 8f_2 = 0 . \tag*{(3)}

In other words, we can always express a0a_0 as a linear combination of the other coefficients. Thus, the number of irreducible IFCs for FCC crystal up to the fifth nearest neighbors is finally reduced to 16[32]; they are

ϕ={b0,b1,b2,c0,c1,d0,d1,d2,d3,e0,e1,e2,f0,f1,f2,f3}.(4)\boldsymbol{\phi} = \{b_0,b_1,b_2,c_0,c_1,d_0,d_1,d_2,d_3,e_0,e_1,e_2,f_0,f_1,f_2,f_3\} . \tag*{(4)}

These are the set of parameters we aim to learn from measured TDS images. Once obtained, they fully determine the phonon dispersion relations as well as many other properties of the crystal, such as all the elastic constants.

We have implemented the above symmetry considerations into a Python program using the SymPy symbolic manipulation package. The program takes the number of nearest neighbor shells to be considered as input, and correctly deduces the set of irreducible IFCs (16 IFCs for up to 5 nearest neighbors). It also constructs the mapping tensor Msym\mathbf{M}_{\mathrm{sym}} that constructs the full Hessian matrix H\mathbf{H} from the set of IFCs,

H=Msym⋅ϕ.(5)\mathbf{H} = \mathbf{M}_{\mathrm{sym}} \cdot\boldsymbol{\phi} . \tag*{(5)}

By construction, the mapping from ϕ\boldsymbol{\phi} to the Hessian matrix H\mathbf{H} is differentiable.

Sampling of Atomic Configurations

The full Hessian matrix H\mathbf{H} constructed from the set of irreducible IFCs uniquely defines a computational X-TDS image, I(q)I(\mathbf{q}). Given that the typical detector exposure times far exceed the atomic vibration correlation times, the measured X-TDS image corresponds to the time-averaged diffraction signals from many atomic configurations. Let I(q;x)I(\mathbf{q};\mathbf{x}) be the X-ray diffraction signal from a specific atomic configuration x\mathbf{x}; I(q;x)I(\mathbf{q};\mathbf{x}) can be computed from x\mathbf{x} using a fast Fourier Transform algorithm from the corresponding atomic density field [30]. In the language of statistical mechanics, the X-TDS image I(q)I(\mathbf{q}) is the ensemble average of I(q;x)I(\mathbf{q};\mathbf{x}), where x\mathbf{x} satisfies the canonical distribution at the given temperature TT with probability density,

f(x)=1Zexp⁡[−U(x)kBT].(6)f(\mathbf{x}) = \frac{1}{Z}\exp\left[-\frac{U(\mathbf{x})}{k_{\mathrm{B}}T}\right] . \tag*{(6)}

where Z=∫R3Nexp⁡(−U(x)kBT) dxZ = \int_{\mathbb{R}^{3N}}\exp\left(-\frac{U(\mathbf{x})}{k_{\mathrm{B}}T}\right)\,d\mathbf{x} is the partition function. Within the HA,

f(x)∝exp⁡[−12(x−x0)T⋅H⋅(x−x0)kBT].(7)f(\mathbf{x}) \propto\exp\left[-\frac{\frac{1}{2}(\mathbf{x}-\mathbf{x}^{0})^{\mathsf{T}}\cdot\mathbf{H}\cdot(\mathbf{x}-\mathbf{x}^{0})}{k_{\mathrm{B}}T}\right] . \tag*{(7)}

This means that f(x)f(\mathbf{x}) is equivalent to a multivariate Gaussian distribution, x∼N(μ,Σ)\mathbf{x}\sim\mathcal{N}(\boldsymbol{\mu},\boldsymbol{\Sigma}), with mean μ=x0\boldsymbol{\mu}=\mathbf{x}^{0} and variance Σ=kBT⋅H−1\boldsymbol{\Sigma}=k_{\mathrm{B}}T\cdot\mathbf{H}^{-1}. In our approach, we compute an estimate of I(q)I(\mathbf{q}) by averaging I(q;x)I(\mathbf{q};\mathbf{x}) over a set of NsN_{\mathrm{s}} atomic configurations sampled over the Gaussian distribution, (7),

Icomp(q)=1Ns∑n=1NsI(q;x(n))(8)I_{\mathrm{comp}}(\mathbf{q}) = \frac{1}{N_{\mathrm{s}}}\sum_{n=1}^{N_{\mathrm{s}}} I(\mathbf{q};\mathbf{x}^{(n)}) \tag*{(8)}

To do so requires the generation of samples from the distribution, (7). To enable gradient-based optimization of the IFCs, this sampling step must be accomplished in a way that allows efficient computation of the derivative of the result with respect to the Hessian matrix. We accomplish this by performing the Cholesky decomposition of the Hessian matrix,

H=L⋅LT.(9)\mathbf{H} = \mathbf{L} \cdot\mathbf{L}^{\mathrm{T}}. \tag*{(9)}

where L\mathbf{L} is a lower triangular matrix. Recall that the atomic displacement δx=x−x0\delta\mathbf{x} = \mathbf{x} - \mathbf{x}^{0} satisfies the Gaussian distribution, δx∼N(0,kBTH−1)\delta\mathbf{x} \sim\mathcal{N}(\mathbf{0}, k_{\mathrm{B}}T\mathbf{H}^{-1}). Let u∼N(0,I)\mathbf{u} \sim\mathcal{N}(\mathbf{0},\mathbf{I}) be a random variable satisfying the standard Gaussian distribution with zero mean and identity matrix as the variance. The atomic displacement can be obtained by solving the matrix equation,

LT⋅δx=kBTu.(10)\mathbf{L}^{\mathrm{T}} \cdot\delta\mathbf{x} = \sqrt{k_{\mathrm{B}}T}\mathbf{u}. \tag*{(10)}

For each random sample of u(n)\mathbf{u}^{(n)}, an independent atomic configuration can be obtained by solving (10), and then adding the displacement to the perfect crystal positions, x(n)=x0+δx(n)\mathbf{x}^{(n)} = \mathbf{x}^{0} + \delta\mathbf{x}^{(n)}. By implementing the above steps in PyTorch, the derivatives of the atomic configuration x(n)\mathbf{x}^{(n)} with respect to the Cholesky factor L\mathbf{L} and then to the Hessian matrix H\mathbf{H} can be computed by backward propagation.

IFC Optimization

The methods described above are combined to give an efficient algorithm that takes an irreducible set of IFCs ϕ\phi as input, and outputs the X-TDS images Ifit(q)I_{\mathrm{fit}}(\mathbf{q}) over a 3-dimensional grid of reciprocal space vectors q\mathbf{q}. Importantly, the algorithm also computes the derivatives of the output Icomp(q)I_{\mathrm{comp}}(\mathbf{q}) with respect to the input ϕ\phi efficiently through backward propagation. Given a target X-TDS image, Itar(q)I_{\mathrm{tar}}(\mathbf{q}), e.g., from experiments, the goal is to search for the IFC values, ϕ\phi, that minimize the variance of the difference between the logarithm of the computed and target X-TDS images. This variance serves as the primary loss function, and is defined as

LVar=Var⁡[log⁡Itar(q)−log⁡Ifit(q)]q∈Q,(11)L_{\mathrm{Var}} = \operatorname{Var}\left[\log I_{\mathrm{tar}}(\mathbf{q}) - \log I_{\mathrm{fit}}(\mathbf{q})\right]_{\mathbf{q} \in\mathcal{Q}}, \tag*{(11)}

where Q\mathcal{Q} is the set of q\mathbf{q}-space points at which the images are compared. The use of the log function in (11) is essential. This is because the signal containing the information about atomic motion primarily resides in the q\mathbf{q}-space between the intense Bragg peaks. The X-ray intensity in such spaces is orders of magnitude smaller than that at the Bragg peaks. The minimization of the variance of intensity on a logarithmic scale effectively applies a homoscedastic transformation on the signal’s variance over the range of the data, contributing more meaningfully to the total error, which would otherwise be dominated by the Bragg peak.

For a physically stable crystal structure, the IFC array ϕ\phi must result in a Hessian matrix H\mathbf{H} that is positive semi-definite. If this condition is not met, the Cholesky decomposition of H\mathbf{H} cannot be completed, and the rest of the workflow cannot proceed. To enforce this constraint, we introduce a penalty term,

Lpenalty=∑imax⁡(0,−λi),(12)L_{\mathrm{penalty}} = \sum_i \max(0,-\lambda_i), \tag*{(12)}

where λi\lambda_i are the 3N3N eigenvalues of the Hessian matrix. Lpenalty=0L_{\mathrm{penalty}} = 0 if all eigenvalues are non-negative, i.e. when H\mathbf{H} is positive semi-definite. The final loss function for the optimization algorithm is a combination of the variance error and the penalty term,

L(ϕ)=LVar⋅1(Lpenalty=0)+α⋅Lpenalty,(13)L(\phi) = L_{\mathrm{Var}} \cdot\mathbb{1}(L_{\mathrm{penalty}} = 0) + \alpha\cdot L_{\mathrm{penalty}}, \tag*{(13)}

where α\alpha is a hyper-parameter controlling the strength of the penalty, and 1(⋅)\mathbb{1}(\cdot) is the indicator function, which equals 1 if the condition inside is true and 0 otherwise. This loss function is implemented in PyTorch and the Adam algorithm is used to minimize the loss and optimize the IFCs.

Results

Here we demonstrate that our method can successfully extract IFC values from TDS data. As a benchmark, we first compute the target TDS signals for Ni modeled by an embedded-atom method (EAM) potential [33] at 300 K, and use this TDS data as ground truth to extract the IFC values. This approach allows us to compare the extracted IFC values with the ground truth, and quantify the error of the numerical approach. The target IFC values are obtained from the Hessian matrix of the embedded-atom method (EAM) potential evaluated at the perfect crystal configuration, with the lattice constant set to its equilibrium value at 300 K (a=3.52a = 3.52 Å [34]), corresponding to the quasi-harmonic approximation. The target TDS signal used for learning the IFC values is generated from an atomistic model of a face-centered cubic (FCC) crystal of Ni of size [5 0 0]×[0 5 0]×[0 0 5][5\ 0\ 0]\times[0\ 5\ 0]\times[0\ 0\ 5], containing 500 atoms, and subjected to periodic boundary conditions (PBC) in all three directions. The target TDS signal Itar(q)I_{\mathrm{tar}}(\mathbf{q}) is computed directly with the same forward model (shown in Figure 2), by sampling 100,000 statistically independent atomic configurations from the quasi-harmonic approximation and averaging their FFT-based scattering intensities. The real-space extent of the supercell in our simulation is L=17.6L = 17.6 Å along each direction. The corresponding reciprocal-space discretization is therefore Δq=2π/L≈0.357\Delta q = 2\pi/L \approx0.357 Å−1^{-1}. The TDS intensity is evaluated on a uniform 81×81×8181 \times81 \times81 grid in reciprocal space centered at q=0\mathbf{q} = 0, spanning wavevectors from −14.28-14.28 Å−1^{-1} to 14.2814.28 Å−1^{-1} along each reciprocal-space direction. This grid provides a sufficiently dense and symmetric sampling of the diffuse scattering signal in the reciprocal space, capturing the relevant phonon-driven intensity modulations over multiple Brillouin zones.

To learn IFC values from the target TDS signal, we use the workflow depicted in Figure 2. As an initial guess, the values of the IFCs are set to those computed from a different interatomic potential of Ni, namely the modified embedded-atom method (MEAM) potential [35]. The IFC values are optimized using the Adam optimizer for 6000 epochs, and the learning rate is reduced by a factor of 2 from 0.4 to 0.1 every 2000 epochs. During each epoch, we generated 15 atomic configurations from which we calculated an average scattering image for the fitting process to have a robust estimate of the TDS whilst limiting computational costs. Computations were performed on an NVIDIA A100 GPU (Ampere architecture) using CUDA 12.5. A single epoch executes in approximately 40 s on a single GPU, and convergence was typically achieved after 4000 epochs (<50< 50 GPU hours), at which point the loss function has stabilized.

Learning IFCs from Full TDS Data in 3D q-Space

Figure 4(a) compares the TDS signal (on a 2D slice of qz=0q_z = 0) for the target EAM potential and the MEAM potentials used as initial guess; qualitatively the images look similar. However, the differences in the IFC values for the EAM and MEAM potentials lead to discernible patterns when the differences between their TDS predictions are plotted in Figure 4(b) (left half), as well as clear differences in the phonon dispersion relation shown in Figure 4(d). Figure 4(c) shows the evolution of the loss function over the training period, demonstrating steady convergence towards the end of the training phase. Already after 1000 optimization epochs, the loss in the simulated TDS intensity image converges to below 8×10−48 \times10^{-4}, indicating that the relative spectral shape of the scattering image is well described.

Comparison of computed TDS signals, errors, training loss, and phonon dispersion relation

Figure 4. a) Comparison of the computed TDS signals at 300 K on the plane of qz=0q_z = 0 (a.k.a. the HK(L=0) slice in the reciprocal space). The sections marked as “Target” correspond to the U3 EAM potential. The section marked as “Initial” corresponds to the MEAM potential. The section marked as “Fit” corresponds to the IFCs after the training. b) The error of the computed TDS signals based on the initial and fitted IFCs when compared with the target. c) The training loss (light purple) is shown together with the moving average (dark purple) and the final value (dashed). d) The phonon dispersion relation computed from the Target, Initial, and Fitted IFCs. The curves corresponding to the Target and Fitted IFCs are nearly indistinguishable here. The inset shows the first Brillouin zone with the paths connecting high symmetry points shown in red.

Figure 4(b) shows the spatial loss distribution of the TDS images across reciprocal space before and after optimization, confirming that the fitted IFCs substantially reduce the intensity mismatch over the full q-space region. The reconstructed phonon dispersion relation along the high symmetry path, shown in Figure 4d, closely resembles the reference EAM dispersion. The dispersion relations agree at points near the Brillouin center, only high frequency phonon modes at the Brillouin zone edge (e.g. at X, X*) exhibit any noticeable (but slight) deviation. This demonstrates that the approach enables accurate and precise reconstruction of the phonon dispersion across the Brillouin zone, consistent with reported results for comparable, non differentiable methods [20, 22, 21].

Figure 5 shows the evolution of the individual IFC values during optimization. The dominant first-nearest-neighbor shell (bkb_k) converges rapidly and stabilizes after approximately 1000 epochs, while the second- and third-nearest-neighbor shells converge more gradually. All IFCs converge to within the statistical uncertainty of the target values, with residual fluctuations reflecting the finite learning rate and sampling. Quantitative convergence errors for all shells are summarized in Table 1. The absolute errors for force constants b0b_0 through f3f_3 are all on the order of 10−310^{-3}. The a0a_0 parameter describing self-interaction, derived from the force constants according to Eq. (3), exhibits an absolute error on the order of 10−210^{-2}, though its relative error remains less than 10−310^{-3}. All parameters corresponding to the interaction of close neighbor shells 0 to 4 (a0a_0 to e2e_2) demonstrate progressive error reduction through the optimization progress. In contrast, fifth-shell constants (f1f_1-f5f_5) maintain relatively constant error magnitudes, indicating limited convergence improvement for these distant interactions. This behavior reflects the decaying sensitivity of the diffuse scattering signal to increasingly distant interatomic interactions. While the dominant short-range force constants are accurately recovered, interactions beyond the fourth neighbor shell contribute only weakly to the scattering intensity and therefore cannot be reliably determined from TDS alone.

Evolution of the fitted IFCs corresponding to different neighboring shells during training

Figure 5. The evolution of the fitted IFCs corresponding to different neighboring shells during training. The initial values of IFCs correspond to the MEAM potential, while the target values correspond to the U3 EAM potential.

StartingTargetConvergedError
a0a_07.5697.4057.406−1.5×10−3-1.5 \times10^{-3}
b0b_0−9.546×10−1-9.546 \times10^{-1}−9.907×10−1-9.907 \times10^{-1}−9.903×10−1-9.903 \times10^{-1}−3.480×10−4-3.480 \times10^{-4}
b1b_1−1.131×10−1-1.131 \times10^{-1}5.773×10−25.773 \times10^{-2}5.557×10−25.557 \times10^{-2}2.155×10−32.155 \times10^{-3}
b2b_2−7.934×10−1-7.934 \times10^{-1}−1.041-1.041−1.035-1.035−5.471×10−3-5.471 \times10^{-3}
c0c_08.487×10−28.487 \times10^{-2}1.358×10−11.358 \times10^{-1}1.318×10−11.318 \times10^{-1}3.978×10−33.978 \times10^{-3}
c1c_1−6.711×10−2-6.711 \times10^{-2}−1.119×10−2-1.119 \times10^{-2}−1.6346×10−2-1.6346 \times10^{-2}5.160×10−35.160 \times10^{-3}
d0d_07.344×10−27.344 \times10^{-2}8.787×10−38.787 \times10^{-3}7.479×10−37.479 \times10^{-3}1.308×10−31.308 \times10^{-3}
d1d_15.354×10−25.354 \times10^{-2}2.866×10−32.866 \times10^{-3}4.682×10−34.682 \times10^{-3}−1.816×10−3-1.816 \times10^{-3}
d2d_2−1.877×10−2-1.877 \times10^{-2}4.220×10−44.220 \times10^{-4}−1.256×10−3-1.256 \times10^{-3}1.678×10−31.678 \times10^{-3}
d3d_33.866×10−23.866 \times10^{-2}5.112×10−35.112 \times10^{-3}6.068×10−36.068 \times10^{-3}−9.563×10−4-9.563 \times10^{-4}
e0e_0−1.083×10−1-1.083 \times10^{-1}−5.230×10−3-5.230 \times10^{-3}−4.565×10−3-4.565 \times10^{-3}−6.646×10−4-6.646 \times10^{-4}
e1e_11.009×10−11.009 \times10^{-1}1.864×10−31.864 \times10^{-3}1.214×10−31.214 \times10^{-3}−1.027×10−3-1.027 \times10^{-3}
e2e_2−1.184×10−1-1.184 \times10^{-1}−5.371×10−3-5.371 \times10^{-3}−4.391×10−3-4.391 \times10^{-3}−9.795×10−4-9.795 \times10^{-4}
f0f_0−1.975×10−8-1.975 \times10^{-8}−1.530×10−3-1.530 \times10^{-3}−1.660×10−3-1.660 \times10^{-3}1.293×10−41.293 \times10^{-4}
f1f_1−1.975×10−8-1.975 \times10^{-8}−5.232×10−4-5.232 \times10^{-4}3.958×10−43.958 \times10^{-4}−4.481×10−4-4.481 \times10^{-4}
f2f_2−1.975×10−8-1.975 \times10^{-8}6.624×10−56.624 \times10^{-5}3.586×10−43.586 \times10^{-4}−2.923×10−4-2.923 \times10^{-4}
f3f_3−1.446×10−21-1.446 \times10^{-21}−6.790×10−4-6.790 \times10^{-4}−8.322×10−4-8.322 \times10^{-4}1.532×10−41.532 \times10^{-4}

Table 1. Comparison between IFC values in eV/A˚2\mathring{\mathrm{A}}^2 for training using TDS data on the entire 3D grid of q-space. The “Starting” column corresponds to the IFCs at the beginning of the training (taken from the MEAM potential). The “Target” column corresponds to the ground truth (here taken as the U3 EAM potential). The “Converged” column corresponds to the IFCs at the end of the training; its difference from the “Target” column is listed in the “Error” column.

The successful recovery of the dominant force constant corresponding to the close neighbor interactions confirms the effectiveness of the approach to recover the correct interaction potential even when starting from a different interatomic potential yielding very similar TDS images, emphasizing the power of quantitative TDS analysis over purely visual comparison. We hereby show that, in addition to reproducing the phonon dispersion relation, the vibrational information contained in the scattering images allows for the precise extraction of the dominant IFCs, including interactions up to the fourth nearest neighbors.

Learning IFCs from Partial TDS Data in 3D q-Space

In practice, acquiring TDS data across the full 3D reciprocal space typically requires sample rotation. Although full 3D TDS acquisition is now routine for ideal specimens, many experimental scenarios render sample rotation impractical or impossible. For instance, in situ monitoring of rapid phase transformations or the study of polycrystalline materials often prohibits sample rotation during observation, limiting data collection to sparse reciprocal-space coverage. Here, we investigate the feasibility of extracting interatomic force constants (IFCs) from TDS data when full 3D coverage is unavailable. We find that fitting a single 2D detector image generally results in incomplete parameter recovery due to a fundamental symmetry limitation inherent to two-dimensional datasets passing through the Brillouin zone center (Γ\Gamma). At the Γ\Gamma-point, phonon eigenvectors decouple into pure Cartesian displacement directions; consequently, a single reciprocal-space slice only constrains force-constant components that project onto phonons polarized within that plane. Force components perpendicular to the slice remain mathematically unconstrained, necessitating multi-slice data for full determination of IFCs.

A ±10∘\pm10^\circ wedge provides a subset of Q which is a practical compromise between data completeness, fit-accuracy, and experimental feasibility. Experimentally, a double-wedge of ±10∘\pm10^\circ around a high symmetry plane, e.g. H(K=0)L can be accessed by rocking the angle between the sample and incoming X-ray beam. An idealized wedge is shown in Figure 6(a). It should be noted that in the optimization procedure we apply all symmetry operations of the lattice (group Oh\mathrm{O}_h) to the computational TDS intensity, such that a 10∘10^\circ wedge is equivalent to a larger fraction of the reciprocal space in a highly symmetric FCC lattice. To assess if IFC recovery is possible on such an incomplete subset of Q, phonon dispersion reproduction, and agreement of TDS intensity, are analyzed for IFCs obtained by fitting only to the q contained in the 10∘10^\circ wedge. Figure 6(b) shows the reconstructed phonon dispersion relation along the high symmetry path. The reconstructed phonon dispersion closely resembles the reference EAM dispersion, with an accuracy comparable to the full three-dimensional fit.

Wedge in reciprocal space and phonon dispersion relation

Figure 6. (a) A wedge in the reciprocal space. Only the TDS data at q points inside the wedge are used to fit the IFCs. The wedge spans ±10∘\pm10^\circ around the qy=0q_y=0, a.k.a. the H(K=0)L, plane, allowing sampling of out-of-plane phonon modes. (b) The phonon dispersion relation computed from the Target, Initial, and Fitted IFCs using only TDS data from the wedge region. The curves corresponding to the Target and Fitted IFCs are nearly indistinguishable here. The inset shows the first Brillouin zone with the paths connecting high symmetry points shown in red.

Figure 7 shows the evolution of the individual IFC values during optimization with 6000 epochs. The dominant first-nearest-neighbor shell converges rapidly and stabilizes after approximately 1000 epochs, while the second- and third-nearest-neighbor shells converge more gradually. The IFCs in the firstnearest neighbor shell converge to the target values within the statistical uncertainty. Quantitative convergence errors for all shells are summarized in Table 2. The absolute errors for force constants b0b_0 through f3f_3 are all on the order of 10−310^{-3}. The a0a_0 parameter describing self-interaction, derived from the force constants according to Eq. (3), exhibits an absolute error on the order of 10−210^{-2}, though its relative error remains at 2×10−32 \times10^{-3}. All parameters corresponding to the interaction of close neighbor shells 0 to 4 (a0a_0 to e2e_2) demonstrate progressive error reduction through the optimization progress. In contrast, fifth-shell constants (f1f_1-f5f_5) maintain relatively constant error magnitudes, indicating limited convergence improvement for these distant interactions. Similar to the full 3D dataset, the error on the individual IFCs for the ±10∘\pm10^\circ wedge after 6000 epochs remain on the order of 10−310^{-3}–10−410^{-4}. Despite the comparably low values in the remaining error of the individual IFCs, the error in the self-interaction term a0a_0 increases from 1.6×10−31.6 \times10^{-3} to 1.5×10−21.5 \times10^{-2}. Since a0a_0 is derived from all IFCs, it serves as an overall measure of the fit quality. Its significant increase, signifies an overall compromised fit quality. Despite this, the agreement of the individual IFCs and the phonon dispersion relation confirms that the wedge geometry captures the essential information needed for an accurate reconstruction. Our results therefore demonstrate that a sample rotation of ±10∘\pm10^\circ provides a feasible path toward IFC refinement in cases where full sample rotation is not achievable, such as in in situ experiments.

The IFCs for the individual shells

Figure 7. The IFCs for the individual shells show that after 6000 epochs, the force constants converge and closely resemble the target U3 EAM IFCs after fitting the initialized MEAM IFCs with a wedge of ±10∘\pm10^\circ around the high symmetry HK0 slice (qz=0q_z = 0).

StartingTargetConvergedError
a0a_07.5697.4057.3891.602×10−21.602 \times10^{-2}
b0b_0−9.546×10−1-9.546 \times10^{-1}−9.907×10−1-9.907 \times10^{-1}−9.894×10−1-9.894 \times10^{-1}−1.311×10−3-1.311 \times10^{-3}
b1b_1−1.131×10−1-1.131 \times10^{-1}5.773×10−25.773 \times10^{-2}5.709×10−25.709 \times10^{-2}6.413×10−46.413 \times10^{-4}
b2b_2−7.934×10−1-7.934 \times10^{-1}−1.041-1.041−1.034-1.034−6.130×10−3-6.130 \times10^{-3}
c0c_08.487×10−28.487 \times10^{-2}1.358×10−11.358 \times10^{-1}1.315×10−11.315 \times10^{-1}4.199×10−34.199 \times10^{-3}
c1c_1−6.711×10−2-6.711 \times10^{-2}−1.119×10−2-1.119 \times10^{-2}−1.845×10−2-1.845 \times10^{-2}7.262×10−37.262 \times10^{-3}
d0d_07.344×10−27.344 \times10^{-2}8.787×10−38.787 \times10^{-3}7.340×10−37.340 \times10^{-3}1.388×10−31.388 \times10^{-3}
d1d_15.354×10−25.354 \times10^{-2}2.866×10−32.866 \times10^{-3}5.232×10−35.232 \times10^{-3}−2.367×10−3-2.367 \times10^{-3}
d2d_2−1.877×10−2-1.877 \times10^{-2}4.220×10−44.220 \times10^{-4}−4.112×10−4-4.112 \times10^{-4}8.223×10−48.223 \times10^{-4}
d3d_33.866×10−23.866 \times10^{-2}5.112×10−35.112 \times10^{-3}6.347×10−36.347 \times10^{-3}−1.235×10−3-1.235 \times10^{-3}
e0e_0−1.083×10−1-1.083 \times10^{-1}−5.230×10−3-5.230 \times10^{-3}−4.252×10−3-4.252 \times10^{-3}−9.771×10−4-9.771 \times10^{-4}
e1e_11.009×10−21.009 \times10^{-2}1.864×10−41.864 \times10^{-4}1.687×10−31.687 \times10^{-3}−1.500×10−3-1.500 \times10^{-3}
e2e_2−1.184×10−1-1.184 \times10^{-1}−5.371×10−3-5.371 \times10^{-3}−4.919×10−3-4.919 \times10^{-3}−4.512×10−4-4.512 \times10^{-4}
f0f_0−1.975×10−8-1.975 \times10^{-8}−1.530×10−3-1.530 \times10^{-3}−2.144×10−3-2.144 \times10^{-3}6.138×10−46.138 \times10^{-4}
f1f_1−1.975×10−8-1.975 \times10^{-8}−5.232×10−5-5.232 \times10^{-5}3.088×10−43.088 \times10^{-4}−3.611×10−4-3.611 \times10^{-4}
f2f_2−1.975×10−8-1.975 \times10^{-8}6.624×10−56.624 \times10^{-5}9.345×10−49.345 \times10^{-4}−8.682×10−4-8.682 \times10^{-4}
f3f_3−1.446×10−21-1.446 \times10^{-21}−6.790×10−4-6.790 \times10^{-4}−6.098×10−5-6.098 \times10^{-5}−6.180×10−4-6.180 \times10^{-4}

Table 2. Comparison between IFC values in eV/Å2^2 for training using TDS data on a wedge region of the qq-space shown in Figure 6(a). The meanings of the columns are the same as in Table 1.

Discussion

Our results demonstrate that thermal diffuse scattering contains rich information on lattice dynamics that can be systematically exploited to recover both phonon spectra and interatomic force constants. The differentiable framework developed herein enables the direct, quantitative extraction of this information, inaccessible by qualitative TDS analysis alone. A direct visual comparison of the TDS images obtained from the target U3 EAM potential and the initial guess obtained from the MEAM potential (Figure 4a) illustrates a fundamental limitation of qualitative TDS analysis. Despite markedly different phonon dispersions (Figure 4d), the HK0 TDS slices of the two potentials are visually indistinguishable. The overall intensity distributions are similar, and the differences are subtle, amounting to a mean absolute deviation of only 2×10−32 \times10^{-3}. These deviations manifest as spatially distributed intensity variations rather than qualitative differences in the scattering pattern. Although the error in any individual feature is small, deviations accumulate coherently across reciprocal space, yielding a quantitative distinction that is statistically significant. This demonstrates that apparent agreement in TDS images does not necessarily imply correct phonon physics, and that qualitative inspection alone can be misleading. A quantitative framework is therefore essential for reliably validating interatomic potentials against diffuse scattering data.

The present framework operates within the harmonic approximation, which assumes that atomic displacements are small and that interatomic forces scale linearly with displacement. This approximation is accurate at low to moderate temperatures where thermal energies remain small. For the FCC nickel system studied here at 300 K, anharmonic contributions to TDS are negligible. At elevated temperatures, however, higher-order phonon–phonon interactions become significant, producing temperature-dependent phonon linewidths and frequency shifts that are especially prominent near phase transitions and cannot be captured by a static harmonic force constant matrix [26, 23]. An important advantage of our approach is that, unlike eigenvalue-based methods, it is formulated directly in terms of atomic displacements. This provides additional entry points for incorporating temperature-dependent corrections and higher-order interactions through modified sampling schemes. In particular, normalizing flows can be used to sample atomic configurations directly from the interatomic potential [36]. The resulting configurations are inherently anharmonic while the sampling procedure remains differentiable. Replacing the harmonic sampling scheme would, therefore, extend the framework to high-temperature regimes without sacrificing differentiability.

The ability to directly fit interatomic force constants to experimental TDS data is particularly relevant for the rapidly growing field of machine learning interatomic potentials (MLIPs) [16, 37, 38, 39], whose accuracy depends critically on the representativeness of their training data. Conventional benchmarks like elastic moduli, phonon frequencies at high-symmetry points, or macroscopic thermal conductivity probe only bulk material properties and may fail to expose errors in the underlying potential energy landscape. Direct validation against experimentally derived IFCs, accessible through the present framework, would additionally provide an atomistic-level assessment of the fidelity across the full Brillouin zone. This positions TDS-based IFC extraction as a powerful complement to existing MLIP validation strategies, and suggests that diffuse scattering datasets could become a valuable addition to the training and benchmarking pipelines used in the development of next-generation interatomic potentials.

Conclusion

We have demonstrated that interatomic force constants and phonon dispersion relations can be recovered directly from thermal diffuse scattering data through gradient-based optimization. The key enabler is a fully differentiable pipeline, from a symmetry-constrained Born–von Kármán parameterization of the Hessian, through Cholesky-based configuration sampling, to simulated TDS intensities. Applied to FCC nickel, the framework accurately recovers force constants up to the fourth nearest-neighbor shell from both full 3D reciprocal-space data and an experimentally realistic ±10∘\pm10^\circ wedge domain, demonstrating robustness under partial data conditions, representative of experimentally constrained acquisition. Basing the framework directly on atomic configuration space is not merely a computational convenience but makes the method extensible to anharmonic regimes and directly applicable to the validation of machine-learned interatomic potentials. Diffuse scattering data, long predominantly used qualitatively, can thus serve as a quantitative target for atomistic refinement.

Conflict of Interest

The authors declare no competing interests.

Data Availability

All data can be obtained on request from the corresponding author. The library for the X-ray scattering calculations can be found at C-XPCS [30, 40].

Acknowledgements

K.S. thanks the Wenner-Gren Foundations (grant number WGF2023-0063). This work is partly supported by the Air Force Office of Scientific Research under award number FA9550-20-1-0397 (S.M.) and FA9550-24-1-0237 (H.Z. and W.C.).

References

  1. [1]Max Born, Kun Huang, and M. Lax. “Dynamical Theory of Crystal Lattices”. In: American Journal of Physics 23.7 (Oct. 1955), p. 474. ISSN: 0002-9505. DOI: 10.1119/1.1934059. URL: https://doi.org/10.1119/1.1934059 (visited on 10/03/2025).
  2. [2]Herbert Fröhlich. “On the theory of superconductivity: the one-dimensional case”. In: Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences 223.1154 (Jan. 1997). Publisher: Royal Society, pp. 296–305. DOI: 10.1098/rspa.1954.0116. URL: https://royalsocietypublishing.org/doi/10.1098/rspa.1954.0116 (visited on 11/18/2025).
  3. [3]J. Bardeen, L. N. Cooper, and J. R. Schrieffer. “Theory of Superconductivity”. In: Physical Review 108.5 (Dec. 1957). Publisher: American Physical Society, pp. 1175–1204. DOI: 10.1103/PhysRev.108.1175. URL: https://link.aps.org/doi/10.1103/PhysRev.108.1175 (visited on 11/18/2025).DOI
  4. [4]J.M. Ziman. Electrons and Phonons: The Theory of Transport Phenomena in Solids. Oxford University Press, Feb. 01, 2018-0-19-850779-6. DOI: 10.1093/acprof:oso/9780198507796.001.0001. URL: https://doi.org/10.1093/acprof:oso/9780198507796.001.0001 (visited on 11/18/2025).
  5. [5]Feliciano Giustino. “Electron-phonon interactions from first principles”. In: Reviews of Modern Physics 89.1 (2017). DOI: 10.1103/RevModPhys.89.015003.DOI
  6. [6]Pinchen Xie et al. “Thermal disorder and phonon softening in the ferroelectric phase transition of lead titanate”. In: Physical Review B 111.9 (Mar. 2025). Publisher: American Physical Society, p. 094113. DOI: 10.1103/PhysRevB.111.094113. URL: https://link.aps.org/doi/10.1103/PhysRevB.111.094113 (visited on 03/28/2025).DOI
  7. [7]W. Cochran. “Lattice vibrations”. en. In: Reports on Progress in Physics 26.1 (Jan. 1963), p. 1. ISSN: 0034-4885. DOI: 10.1088/0034-4885/26/1/301. URL: https://doi.org/10.1088/0034-4885/26/1/301 (visited on 11/18/2025).
  8. [8]K. Parlinski, Z. Q. Li, and Y. Kawazoe. “First-Principles Determination of the Soft Mode in Cubic ${\mathrm{ZrO}}_{2}$”. In: Physical Review Letters 78.21 (May 1997). Publisher: American Physical Society, pp. 4063–4066. DOI: 10.1103/PhysRevLett.78.4063. URL: https://link.aps.org/doi/10.1103/PhysRevLett.78.4063 (visited on 09/09/2025).DOI
  9. [9]Huiju Lee and Yi Xia. “Machine learning a universal harmonic interatomic potential for predicting phonons in crystalline solids”. In: Applied Physics Letters 124.10 (Mar. 2024), p. 102202. ISSN: 0003-6951. DOI: 10.1063/5.0199743. URL: https://doi.org/10.1063/5.0199743 (visited on 10/03/2025).
  10. [10]M. d’Astuto et al. “Phonon dispersion studies of crystalline materials using high-energy resolution inelastic X-ray scattering (IXS)”. In: Physica B: Condensed Matter. Proceedings of the 10th International Conference on Phonon Scattering in Condensed Matter 316-317 (May 2002), pp. 150–153. ISSN: 0921-4526. DOI: 10.1016/S0921-4526(02)00445-3. URL: https://www.sciencedirect.com/science/article/pii/S0921452602004453 (visited on 11/20/2025).DOI
  11. [11]R. Nicklow, N. Wakabayashi, and H. G. Smith. “Lattice Dynamics of Pyrolytic Graphite”. en. In: Physical Review B 5.12 (June 1972), pp. 4951–4962. ISSN: 0556-2805. DOI: 10.1103/PhysRevB.5.4951. URL: https://link.aps.org/doi/10.1103/PhysRevB.5.4951 (visited on 09/09/2025).DOI
  12. [12]Jürgen Neuhaus et al. “Role of vibrational entropy in the stabilization of the high-temperature phases of iron”. In: Physical Review B 89.18 (May 2014). Publisher: American Physical Society, p. 184302. DOI: 10.1103/PhysRevB.89.184302. URL: https://link.aps.org/doi/10.1103/PhysRevB.89.184302 (visited on 09/09/2025).DOI
  13. [13]M. Schwoerer-Böhning, A. T. Macrander, and D. A. Arms. “Phonon Dispersion of Diamond Measured by Inelastic X-Ray Scattering”. In: Physical Review Letters 80.25 (June 1998). Publisher: American Physical Society, pp. 5572–5575. DOI: 10.1103/PhysRevLett.80.5572. URL: https://link.aps.org/doi/10.1103/PhysRevLett.80.5572 (visited on 11/18/2025).
  14. [14]M. Mohr et al. “Phonon dispersion of graphite by inelastic x-ray scattering”. In: Physical Review B 76.3 (July 2007). Publisher: American Physical Society, p. 035439. DOI: 10.1103/PhysRevB.76.035439. URL: https://link.aps.org/doi/10.1103/PhysRevB.76.035439 (visited on 11/18/2025).DOI
  15. [15]Alfred Q. R. Baron. “High-Resolution Inelastic X-Ray Scattering I: Context, Spectrometers, Samples, and Superconductors”. en. In: Synchrotron Light Sources and Free-Electron Lasers. Springer, Cham, 2016, pp. 1643–1719. ISBN: 978-3-319-14394-1. DOI: 10.1007/978-3-319-14394-1_41. URL: https://link.springer.com/rwe/10.1007/978-3-319-14394-1_41 (visited on 11/20/2025).DOI
  16. [16]Chunfeng Cui et al. “Machine learning interatomic potentials as efficient tools for obtaining reasonable phonon dispersions and accurate thermal conductivity: A case study of typical two-dimensional materials”. In: Applied Physics Letters 123.15 (2023).DOI
  17. [17]Dennis S. Kim, Michael Xu, and James M. LeBeau. “Modeling Temperature-Dependent Electron Thermal Diffuse Scattering via Machine-Learned Interatomic Potentials and Path-Integral Molecular Dynamics”. en. In: Physical Review Letters 132.8 (Feb. 2024), p. 086301. ISSN: 0031-9007, 1079-7114. DOI: 10.1103/PhysRevLett.132.086301. URL: https://link.aps.org/doi/10.1103/PhysRevLett.132.086301 (visited on 10/23/2024).DOI
  18. [18]Shaswat Mohanty, Yifan Wang, and Wei Cai. “Generalizability of Graph Neural Network Force Fields for Predicting Solid-State Properties”. In: arXiv preprint arXiv:2409.09931 (2024).arxiv.org/abs/2409.09931
  19. [19]JP Bergsma et al. “Transient x-ray scattering calculated from molecular dynamics”. In: The Journal of chemical physics 84.11 (1986), pp. 6151–6160.
  20. [20]M. Holt et al. “Determination of Phonon Dispersions from X-Ray Transmission Scattering: The Example of Silicon”. en. In: Physical Review Letters 83.16 (Oct. 1999), pp. 3317–3319. ISSN: 0031-9007, 1079-7114. DOI: 10.1103/PhysRevLett.83.3317. URL: https://link.aps.org/doi/10.1103/PhysRevLett.83.3317 (visited on 03/02/2025).DOI
  21. [21]M. Holt et al. “X-Ray Studies of Phonon Softening in TiSe 2”. en. In: Physical Review Letters 86.17 (Apr. 2001), pp. 3799–3802. ISSN: 0031-9007, 1079-7114. DOI: 10.1103/PhysRevLett.86.3799. URL: https://link.aps.org/doi/10.1103/PhysRevLett.86.3799 (visited on 10/23/2024).DOI
  22. [22]M. Holt et al. “Phonon dispersions in niobium determined by x-ray transmission scattering”. en. In: Physical Review B 66.6 (Aug. 2002), p. 064303. ISSN: 0163-1829, 1095-3795. DOI: 10.1103/PhysRevB.66.064303. URL: https://link.aps.org/doi/10.1103/PhysRevB.66.064303 (visited on 10/23/2024).DOI
  23. [23]Ruqing Xu and Tai C. Chiang. “Determination of phonon dispersion relations by X-ray thermal diffuse scattering”. en. In: Zeitschrift für Kristallographie - Crystalline Materials 220.12 (Dec. 2005). Publisher: De Gruyter (O), pp. 1009–1016. ISSN: 2196-7105. DOI: 10.1524/zkri.2005.220.12.1009. URL: https://www.degruyter.com/document/doi/10.1524/zkri.2005.220.12.1009/html (visited on 10/23/2024).DOI
  24. [24]J. Wong et al. Intensity Pattern of Diffuse X-Ray Scattering From Thermally Populated Phonons in Fcc d-Pu-Ga. ENGLISH. Tech. rep. G. Djarvinen, eds.;AIP Publishing Center;Vol.673;07;(2003), Jan. 2016. URL: https://www.osti.gov/biblio/1235452 (visited on 10/23/2024).DOI
  25. [25]A. B. Mei et al. “Reflection thermal diffuse x-ray scattering for quantitative determination of phonon dispersion relations”. en. In: Physical Review B 92.17 (Nov. 2015), p. 174301. ISSN: 1098-0121, 1550-235X. DOI: 10.1103/PhysRevB.92.174301. URL: https://link.aps.org/doi/10.1103/PhysRevB.92.174301 (visited on 10/23/2024).DOI
  26. [26]Rozaliya I. Barabash, Gene E. Ice, and Patrice E. A. Turchi. Diffuse Scattering and the Fundamental Properties of Materials. en. Momentum Press, 2009. ISBN: 978-1-60650-000-2.
  27. [27]Björn Wehinger et al. “Full Elasticity Tensor from Thermal Diffuse Scattering”. In: Physical Review Letters 118.3 (Jan. 2017). Publisher: American Physical Society, p. 035502. DOI: 10.1103/PhysRevLett.118.035502. URL: https://link.aps.org/doi/10.1103/PhysRevLett.118.035502 (visited on 09/08/2025).DOI
  28. [28]Nicholas J. Weadock et al. “The nature of dynamic local order in CH3NH3PbI3 and CH3NH3PbBr3”. In: Joule 7.5 (May 2023), pp. 1051–1066. ISSN: 2542-4351. DOI: 10.1016/j.joule.2023.03.017. URL: https://www.sciencedirect.com/science/article/pii/S2542435123001290 (visited on 09/11/2023).DOI
  29. [29]Pol Benítez et al. Why Physics Still Matters: Improving Machine Learning Prediction of Material Properties with Phonon-Informed Datasets. arXiv:2511.15222 [cond-mat]. Nov. 2025. DOI: 10.48550/arXiv.2511.15222. URL: http://arxiv.org/abs/2511.15222 (visited on 11/24/2025).
  30. [30]Shaswat Mohanty et al. “Computational Approaches to Model X-ray Photon Correlation Spectroscopy from Molecular Dynamics”. In: Modelling and Simulation in Materials Science and Engineering (Aug. 2022).DOI
  31. [31]V. Frei. “Direct determination of born — von Kármán force constants for face-centred cubic elements from measured phonon frequencies”. en. In: Czechoslovak Journal of Physics B 35.9 (Sept. 1985), pp. 1049–1052. ISSN: 1572-9486. DOI: 10.1007/BF01676365. URL: https://doi.org/10.1007/BF01676365 (visited on 03/29/2026).
  32. [32]L. Enkhtor and V. M. Silonov. “The calculation of radial and tangential force constants for nickel using an experimental phonon spectrum”. en. In: Moscow University Physics Bulletin 71.1 (Jan. 2016), pp. 123–127. ISSN: 1934-8460. DOI: 10.3103/S0027134916010069. URL: https://doi.org/10.3103/S0027134916010069 (visited on 03/29/2026).
  33. [33]SM Foiles, MI Baskes, and Murray S Daw. “Embedded-atom-method functions for the fcc metals Cu, Ag, Au, Ni, Pd, Pt, and their alloys”. In: Physical review B 33.12 (1986), p. 7983.pubmed.ncbi.nlm.nih.gov/9938188
  34. [34]Jinn-Wen Hwang. “Thermal expansion of nickel and iron, and the influence of nitrogen on the lattice parameter of iron at the Curie temperature”. In: (1972).
  35. [35]Michael I Baskes. “Determination of modified embedded atom method parameters for nickel”. In: Materials Chemistry and Physics 50.2 (1997), pp. 152–158.DOI
  36. [36]Rasool Ahmad and Wei Cai. “Free energy calculation of crystalline solids using normalizing flows”. en. In: Modelling and Simulation in Materials Science and Engineering 30.6 (July 2022). Publisher: IOP Publishing, p. 065007. ISSN: 0965-0393. DOI: 10.1088/1361-651X/ac7f4b. URL: https://doi.org/10.1088/1361-651X/ac7f4b (visited on 11/18/2025).
  37. [37]Andrew Rohskopf et al. “Empirical interatomic potentials optimized for phonon properties”. English. In: npj Computational Materials 3.1 (2017). ISSN: 2057-3960. DOI: 10.1038/s41524-017-0026-y.DOI
  38. [38]Pjotrs Žguns, Inga Pudza, and Alexei Kuzmin. “Benchmarking CHGNet Universal Machine Learning Interatomic Potential against DFT and EXAFS: The Case of Layered WS2and MoS2”. English. In: Journal of Chemical Theory and Computation 21.16 (2025), pp. 8142–8150. ISSN: 1549-9618. DOI: 10.1021/acs.jctc.5c00955.DOI
  39. [39]Bowen Han and Yongqiang Cheng. “Benchmarking universal machine learning interatomic potentials for rapid analysis of inelastic neutron scattering data”. English. In: Machine Learning: Science and Technology 6.3 (2025). ISSN: 2632-2153. DOI: 10.1088/2632-2153/adfa68.DOI
  40. [40]Shaswat Mohanty et al. “Evaluating the Transferability of Machine-Learned Force Fields for Material Property Modeling”. In: arXiv preprint arXiv:2301.03729 (2023).DOI

Paper details

vlnc:W_ye7brutp
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents