Stiffness tensor condition-guided diffusion generation method for material crystal structures

By establishing a unique standard coordinate system and freezing the weights of the pre-trained network, the stability and stiffness control issues of generating two-dimensional material structures under small-scale labeled data are solved, achieving thermodynamically stable and stiffness-controllable generation results, which are applicable to flexible electronics and nanoelectromechanical systems.

CN123072575APending Publication Date: 2026-11-10SHANGHAI INST OF CERAMIC CHEM & TECH CHINESE ACAD OF SCI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611082338.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-21
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies struggle to generate thermodynamically stable and stiffness-controllable two-dimensional material structures with small-scale labeled data. Furthermore, the injection of stiffness tensor conditions can easily disrupt rotational isomorphism, leading to unreasonable generation results.

Method used

By establishing a unique standard coordinate system, synchronously transforming the lattice matrix, atomic coordinates, and in-plane stiffness tensor, a joint condition for multiple independent stiffness components is constructed. The stiffness tensor condition embedding module and condition adapter are used to inject a pre-trained isotropic diffusion denoising network, and transfer learning is performed by freezing the main weights of the network to generate a two-dimensional crystal structure that meets the target stiffness condition.

Benefits of technology

It enables the generation of thermodynamically stable and stiffness-controllable two-dimensional material structures under small-scale labeled data, solves the problem of rotational deformation failure, improves the physical rationality and stability of the generated results, and meets the anisotropic mechanical performance requirements of flexible electronics and nanoelectromechanical systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123072575A_ABST
    Figure CN123072575A_ABST
Patent Text Reader

Abstract

This invention relates to a method for generating material crystal structures through diffusion guided by stiffness tensor conditions, comprising: acquiring two-dimensional material crystal structure data; establishing a unique canonical coordinate system fixed to the crystal structure based on layer normals, in-plane lattice basis vectors, and preset sorting and orientation determination rules; synchronously transforming the lattice matrix, atomic coordinates, and in-plane stiffness tensor to the aforementioned coordinate system; constructing multiple independent stiffness components into joint conditions and mapping them to conditional embedding vectors, and injecting them layer by layer into a pre-trained isovariant diffusion denoising network; freezing the main weights of the network and training the conditional embedding module and conditional adapter; performing inverse denoising starting from random noise to generate a two-dimensional crystal structure in which each stiffness component is offset towards a target range. This invention can generate thermodynamically stable and stiffness-controllable two-dimensional material structures with small-scale labeled data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of computational materials science and generative artificial intelligence, and in particular to a method for generating material crystal structures by diffusion guided by stiffness tensor conditions. Background Technology

[0002] The stiffness tensor of a crystalline material is a fundamental physical quantity characterizing its elastic mechanical behavior, consisting of multiple independent components that describe the material's stiffness properties in different directions. Reverse engineering of crystal structures to meet specific stiffness requirements has significant application value in aerospace, flexible electronics, and functional composite materials. Traditionally, obtaining materials that meet specific stiffness indices relies on high-throughput first-principles calculations for screening, which is costly and the results are limited by the coverage of existing structure libraries. Recent machine learning-assisted methods are mostly forward prediction models used to accelerate screening but still cannot actively generate new structures. Reverse design based on generative models has gained attention in recent years. Existing conditional crystal generation works such as MatterGen support generation guided by scalar properties such as bulk modulus and shear modulus. However, the stiffness tensor is inherently a multi-component tensor; a single scalar constraint cannot distinguish the stiffness differences in different directions within a plane, and the generated results cannot systematically shift with different target anisotropy settings. Taking two-dimensional materials as an example, their in-plane stiffness tensor is described by multiple independent components, which is particularly crucial for applications such as flexible electronics, nanoelectromechanical systems, and composite material reinforcement that require precise control of in-plane anisotropic mechanical properties.

[0003] Introducing the complete stiffness tensor as a condition into the isotropic diffusion generation model faces two technical challenges. First, the components of the stiffness tensor change in tandem with the coordinate system under structural rotation. Directly concatenating these components into a regular input vector and injecting it into the network would cause inconsistencies in the output of the denoising network depending on the rotation of the input structure, thus violating the rotational isotropy upon which the isotropic diffusion model relies and resulting in physically unreasonable generated structures. Second, the scale of high-precision two-dimensional material stiffness tensor annotation data is much smaller than that of general crystal pre-training data. Directly training the conditional diffusion model from random initialization is prone to convergence difficulties or overfitting, leading to poor thermodynamic stability of the generated structures. Summary of the Invention

[0004] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a method for generating material crystal structures by diffusion guided by stiffness tensor conditions. By using tensor isovariant encoding and transfer learning strategies, stiffness tensor condition injection that maintains rotational isovariance is achieved, and thermodynamically stable and stiffness-controllable two-dimensional material structures are generated under small-scale labeled data. This provides an effective technical means for the on-demand design of anisotropic mechanical functional materials.

[0005] To achieve the above objectives, the present invention adopts the following technical solution.

[0006] This invention provides a method for diffusion generation of material crystal structures guided by stiffness tensor conditions, employing the following technical solution: Obtain training samples containing two-dimensional crystal structures and corresponding in-plane stiffness tensors, wherein the training samples include lattice matrices, atomic coordinates, and in-plane stiffness tensors; Based on the layer normals of the two-dimensional lattice, the lengths and angles of the in-plane lattice basis vectors, and the preset basis vector ordering and orientation determination rules, a unique standard coordinate system is established that is fixed to the two-dimensional crystal structure. Based on the rotational transformation relationship from the original coordinate system to the unique canonical coordinate system, the lattice matrix, atomic coordinates and in-plane stiffness tensor in the training samples are synchronously transformed, wherein the in-plane stiffness tensor is transformed according to the fourth-order tensor transformation rule; The transformed in-plane stiffness tensor is constructed into a joint condition containing multiple independent stiffness components, and then mapped into a fixed-dimensional condition embedding vector through the stiffness tensor condition embedding module. The conditional embedding vector is injected layer by layer into a pre-trained equivariant diffusion denoising network through a conditional adapter. With the main weights of the pre-trained isovariant diffusion denoising network frozen, the parameters of the stiffness tensor conditional embedding module and the conditional adapter are trained based on the training samples. Obtain the target stiffness condition expressed according to the component definition rules of the unique standard coordinate system; Map the target stiffness condition to a target condition embedding vector; and Starting from random noise, the target condition embedding vector is input into the trained network to perform a reverse denoising process, generating a two-dimensional crystal structure corresponding to the target stiffness condition.

[0007] Furthermore, in the above method, the step of constructing the transformed in-plane stiffness tensor into a joint condition containing multiple independent stiffness components includes: The transformed in-plane stiffness tensor is expressed as a stiffness matrix using Voigt notation; and The multiple matrix elements of the stiffness matrix are flattened into a vector in a fixed order, and the vector contains multiple independent stiffness components.

[0008] Furthermore, in the above method, the plurality of independent stiffness components include C 11 C 12 C 22 and C 66 .

[0009] Furthermore, in the above method, the step of mapping the stiffness tensor conditional embedding module to a fixed-dimensional conditional embedding vector includes: The stiffness components in the combined conditions are standardized; and The standardized stiffness component is input into a multilayer perceptron network including multiple fully connected layers and activation function layers, and the output is a fixed-dimensional conditional embedding vector. The dimension of the conditional embedding vector is consistent with the hidden dimension of the pre-trained equivariant diffusion denoising network.

[0010] Furthermore, in the above method, the conditional adapter is inserted into each interaction block of the pre-trained equivariant diffusion denoising network to inject the conditional embedding vector into the network in a layer-by-layer modulation manner.

[0011] Furthermore, in the above method, the step of training the parameters of the stiffness tensor conditional embedding module and the conditional adapter adopts a classifier-guided strategy, including: During training, stiffness tensor conditions are randomly discarded with a preset probability, enabling the network to learn both conditionally generated and unconditionally generated distributions simultaneously; and During the sampling process, a weighted interpolation of the conditional and unconditional scores is performed using a guiding coefficient to adjust the guiding strength of the stiffness tensor condition, where the guiding score s guide =s uncond +γ(s cond -s uncond ), where γ is the guiding coefficient, and s cond For conditional scoring, s uncond This is an unconditional score.

[0012] Furthermore, in the above method, the conditional adapter broadcasts the conditional embedding vector to each atom node and splices or modulates it only with the zero-order scalar feature channel in the pre-trained equivariant diffusion denoising network.

[0013] Furthermore, the above method also includes: The thermodynamic stability of the generated two-dimensional crystal structures is screened, and structures with energies below a preset threshold on the convex hull are retained.

[0014] Furthermore, the above method also includes: First-principles calculations were performed on the generated two-dimensional crystal structure to verify the degree of agreement between the actual elastic constants and the target stiffness condition; and When the generated result deviates from the target stiffness condition, adjust the guiding coefficient and regenerate.

[0015] Furthermore, in the above method, the steps for establishing a unique normalized coordinate system include: The gauge z-axis is determined based on the layer normals of the two-dimensional lattice; Sort the in-plane lattice basis vectors by length from smallest to largest, and select the direction corresponding to the shorter basis vector as the candidate x-axis direction; When the difference between the lengths of two basis vectors is less than a preset threshold, a unique canonical x-axis direction is determined according to the basis vector angle or a preset direction priority rule; and Determine the standard y-axis using a right-handed coordinate system.

[0016] In summary, compared with the prior art, the present invention has at least one of the following beneficial technical effects: This invention solves the rotational isotropic damage problem during stiffness tensor condition injection by establishing a unique canonical coordinate system and synchronously transforming the stiffness tensor according to tensor transformation rules. This makes the generated two-dimensional crystal structure physically reasonable and unaffected by the choice of coordinate system. By freezing the main weights of the pre-trained isotropic diffusion denoising network and training only the stiffness tensor condition embedding module and condition adapter, efficient transfer learning is achieved on small-scale labeled data. This avoids convergence difficulties and overfitting problems caused by training from random initialization, and the thermodynamic stability of the generated structure is significantly improved. By using multiple independent components of the complete stiffness tensor as joint conditions to guide the generation process, precise control of the in-plane anisotropic mechanical properties of two-dimensional materials is achieved. The generated results can be systematically shifted according to different target stiffness conditions, providing an effective technical means for the on-demand design of anisotropic mechanical functional materials in fields such as flexible electronics and nanoelectromechanical devices. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart of an embodiment of the material crystal structure diffusion generation method guided by the stiffness tensor condition of the present invention is shown.

[0019] Figure 2 A flowchart of an embodiment of the stiffness tensor condition construction and embedding process of the present invention is shown.

[0020] Figure 3 The flowchart shown is an embodiment of the training and sampling process of the present invention using a classifier-guided strategy.

[0021] Figure 4 A flowchart illustrating an embodiment of the process for screening and verifying generated structures according to the present invention is shown. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Furthermore, it should be understood that the specific embodiments described herein are only for illustration and explanation of this application and are not intended to limit this application.

[0023] It should be noted that the order of description of the following embodiments is not intended to limit the preferred order of the embodiments of this application. Furthermore, the descriptions of each embodiment in the following embodiments have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0024] The method steps described in this embodiment of the invention can be executed in the order described in the specific implementation, or the execution order of each step can be adjusted according to actual needs, provided that the technical problem can be solved. These are not listed one by one here.

[0025] The present invention will be further described in detail below with reference to the accompanying drawings.

[0026] Reference Figure 1 This invention provides a method 100 for generating material crystal structures through diffusion guided by stiffness tensor conditions. Method 100 includes the following steps: 102 acquiring two-dimensional material crystal structure data; 104 establishing a unique canonical coordinate system; 106 synchronously transforming the lattice matrix, atomic coordinates, and in-plane stiffness tensor; 108 constructing joint conditions and mapping them to conditional embedding vectors; 110 injecting the conditional embedding vectors layer by layer into a pre-trained isovariant diffusion denoising network; 112 freezing the main weights of the pre-trained network and training the parameters of newly added modules; and 114 performing inverse denoising from random noise to generate a two-dimensional crystal structure.

[0027] In step 102, training samples containing two-dimensional crystal structures and corresponding in-plane stiffness tensors are obtained. In some embodiments, the training samples are derived from the C2DB two-dimensional material database. Based on C2DB extended data obtained up to a specific date, 8454 training samples are obtained after deduplication, complete stiffness annotation screening, and stability filtering. For each training sample, a fractional coordinate matrix, atomic number list, lattice matrix, and in-plane stiffness tensor are recorded to constitute the target structure data of the diffusion model. In some embodiments, thermodynamic stability screening of the training data is performed using the energy on the convex hull, retaining stable structures with convex hull energies below a preset threshold to reduce the interference of unstable samples on the generated distribution.

[0028] Based on the layer normals of the two-dimensional lattice, the lengths and angles of the in-plane lattice basis vectors, and the preset basis vector sorting and direction determination rules, step 104 establishes a unique canonical coordinate system fixed to the two-dimensional crystal structure. The steps for establishing the unique canonical coordinate system include: determining the canonical z-axis based on the layer normals of the two-dimensional lattice; sorting the in-plane lattice basis vectors from shortest to longest, selecting the direction corresponding to the shorter basis vector as the candidate x-axis direction; when the length difference between two in-plane lattice basis vectors is less than a preset threshold, determining the unique canonical x-axis direction according to the basis vector angle or a preset direction priority rule; and determining the canonical y-axis according to a right-handed coordinate system. When multiple equivalent candidate coordinate systems still exist according to the aforementioned rules, the lattice matrix and atomic coordinates are transformed based on each equivalent candidate coordinate system. The transformed atoms are arranged according to the preset sorting order of atomic number and fractional coordinates to form a structure code corresponding to each equivalent candidate coordinate system. The candidate coordinate system corresponding to the structure code with the smallest lexicographical order is determined as the unique canonical coordinate system. For structures stored in C2DB using crystallographic standard orientation, first verify their orientation according to the above rules; if the rules are met, use them directly; otherwise, perform synchronous conversion.

[0029] Step 106 performs simultaneous transformations on the lattice matrix, atomic coordinates, and in-plane stiffness tensor in the training samples based on the rotational transformation relationship from the original coordinate system to the unique canonical coordinate system. The in-plane stiffness tensor is transformed according to the fourth-order tensor transformation rule, and the transformation formula is as follows: Where R is the rotation matrix, and C and C' are the stiffness tensor components before and after the transformation, respectively. The transformation of atomic coordinates is as follows: the original fractional coordinates are converted to Cartesian atomic coordinates, the same rotation transformation is applied to the lattice matrix and the Cartesian atomic coordinates, and then the fractional coordinates in the gauge coordinate system are calculated based on the transformed lattice matrix. Through synchronous transformation, the structural representation and the stiffness tensor components maintain a one-to-one correspondence in the same gauge coordinate system.

[0030] In step 108, the transformed in-plane stiffness tensor is constructed as a joint condition containing multiple independent stiffness components, which is then mapped into a fixed-dimensional conditional embedding vector by the stiffness tensor conditional embedding module. The stiffness tensor conditional embedding module inputs the standardized stiffness matrix elements in a unique canonical coordinate system into the multilayer perceptron, mapping them into a fixed-dimensional scalar conditional embedding vector. Since the same crystal structure and its stiffness tensor are synchronously normalized according to the same rules before entering the model, even if the original structure uses different global rotation representations, a consistent canonical structure representation and canonical stiffness conditions are obtained. Therefore, this conditional embedding does not carry additional directional information relative to the instantaneous attitude of the network input.

[0031] Step 110 injects the conditional embedding vector layer by layer into the pre-trained equivariant diffusion denoising network through the conditional adapter. The pre-trained equivariant diffusion denoising network is an E(3) equivariant graph neural network that maintains equivariance under structural rotation and translation transformations. In some implementations, the pre-trained equivariant diffusion denoising network uses the GemNet-T equivariant network as the denoising backbone network, containing 46.8M parameters. This network can map noisy atomic structures to estimates of the noise added to each variable. The diffusion process is combined for three types of variables: discrete diffusion for atomic types, periodic coordinate diffusion for fractional coordinates, and continuous diffusion for lattices. The denoising network predicts the scores of the three types of variables simultaneously at each diffusion time step. The conditional adapter is inserted into each interaction block of the pre-trained equivariant diffusion denoising network, broadcasting the conditional embedding vector to each atomic node, and splicing or affine modulation only with the zero-order scalar feature channel, so that the stiffness condition continues to participate in the feature update process without destroying the network's E(3) equivariance.

[0032] With the main weights of the pre-trained isovariant diffusion denoising network frozen, step 112 trains the parameters of the stiffness tensor conditional embedding module and the conditional adapter based on the training samples. This network has already learned general geometric laws and thermodynamic stability priors from large-scale crystal data during the pre-training phase; conditional adaptation is based on this, eliminating the need to learn the crystal structure generation laws from scratch. The number of trainable parameters is much smaller than full-scale fine-tuning, suppressing overfitting when the data scale is limited. In some implementations, training uses a learning rate of 5e-6 for a total of 200 epochs. The loss function is a weighted sum of the denoising objectives for three types of variables: atom type, fractional coordinates, and lattice. The model simultaneously optimizes the generation quality of these three types of variables through backpropagation.

[0033] Step 114 starts with random noise and performs a reverse denoising process through the trained network to generate a two-dimensional crystal structure corresponding to the target stiffness condition. During the sampling phase, the target stiffness condition expressed according to the component definition rules of the unique canonical coordinate system is first obtained. When the target stiffness tensor is given based on other coordinate systems, it is first transformed to the unique canonical coordinate system according to the rotation transformation relationship from that coordinate system to the unique canonical coordinate system. The transformed target stiffness condition is mapped to a target condition embedding vector according to the process in step 108, and used as the guiding condition input to the denoising network during the sampling phase. The reverse denoising process is executed for 1000 steps, gradually restoring the two-dimensional crystal structure that conforms to the target stiffness condition from pure noise. In some implementations, a classifier-guided strategy is used to adjust the guiding strength of the stiffness tensor condition. The guided coefficient is used to perform weighted interpolation of the conditional and unconditional scores; the larger the guided coefficient, the stronger the influence of the target stiffness on the generated structure.

[0034] Reference Figure 2This invention also provides a process 200 for constructing and embedding stiffness tensor conditions. Process 200 describes in detail... Figure 1 The process of constructing and embedding stiffness tensor conditions involved in step 108 includes steps 202 to 210.

[0035] Step 202 represents the transformed in-plane stiffness tensor as a stiffness matrix using Voigt notation. Voigt notation utilizes the symmetry of the fourth-order stiffness tensor to represent the in-plane stiffness tensor as a 3×3 stiffness matrix. The unit of the stiffness matrix is ​​N / m, reflecting the linear relationship between in-plane stress and strain in a two-dimensional material. In this embodiment, the Voigt notation defines the third component of the strain vector as the engineering shear strain γ. xy =2ε xy The corresponding stiffness matrix element C 16 C 26 C 66 The relationship between equal components and shear strain is calculated according to this convention.

[0036] Step 204 flattens the multiple matrix elements of the stiffness matrix into a vector in a fixed order. The stiffness matrix is ​​a 3×3 matrix, which is flattened into a 9-dimensional vector of matrix elements in a fixed order. The flattened vector contains multiple independent stiffness components. For orthorhombic crystal systems and two-dimensional materials with higher symmetry, the independent stiffness components include C... 11 C 12 C 22 and C 66 For materials with lower symmetry, the independent stiffness components may also include C. 16 and C 26 This is used to describe the coupling effect between in-plane shear and normal strain.

[0037] Step 206 standardizes the stiffness components in the joint conditions. Standardization involves applying zero-mean, unit-variance standardization to each matrix element based on the training set statistics, eliminating the impact of numerical scale differences between stiffness components on subsequent neural network training. The standardized stiffness components have a numerical range close to a standard normal distribution, facilitating the convergence of the multilayer perceptron network.

[0038] Step 208 inputs the standardized stiffness components into the multilayer perceptron network. The stiffness tensor conditional embedding module includes multiple fully connected layers and activation function layers, used to map the stiffness tensor into a conditional embedding vector with the same hidden dimension as the pre-trained isovariant diffusion denoising network. The stiffness tensor conditional embedding module consists of a three-layer multilayer perceptron with the structure Linear(9→512), SiLU activation, Linear(512→512), SiLU activation, and Linear(512→512). The first fully connected layer maps the 9-dimensional standardized stiffness vector to a 512-dimensional latent space. The subsequent two fully connected layers perform nonlinear transformations in the 512-dimensional space, and SiLU activation functions are inserted between layers to introduce nonlinearity.

[0039] Step 210 outputs a fixed-dimensional conditional embedding vector. The multilayer perceptron network outputs a 512-dimensional conditional embedding vector, with the same dimension as the hidden dimension of the pre-trained equivariant diffusion denoising network, facilitating subsequent injection into various layers of the network via a conditional adapter. The conditional embedding vector is a scalar vector and does not carry directional information, maintaining compatibility with the equivariance of the pre-trained network.

[0040] In some implementations, when it is necessary to simultaneously constrain the chemical system, the chemical system embedding and stiffness embedding are concatenated into a joint conditional vector input to the denoising network. The chemical system embedding maps the elemental composition to a fixed-dimensional vector through an independent embedding module, which, after being concatenated with the stiffness embedding vector, forms a higher-dimensional joint condition, thereby guiding the generation process to satisfy the dual constraints of chemical composition and mechanical properties.

[0041] In some implementations, the user-specified target stiffness condition is also expressed in a unique canonical coordinate system. When the target stiffness tensor is given based on another coordinate system, it is first transformed to the unique canonical coordinate system according to the rotation transformation relationship from that coordinate system to the unique canonical coordinate system. The transformed target stiffness tensor is constructed as a conditional embedding vector according to the process of steps 202 to 210, which serves as the guiding conditional input to the denoising network in the sampling stage, driving the generated structure to shift towards the target stiffness range.

[0042] Reference Figure 3 This invention also provides a training and sampling process 300 using a classifier free guidance strategy. Process 300 includes: a training phase step 302 where stiffness tensor conditions are randomly discarded with a preset probability; a determination 304 determining whether stiffness tensor conditions have been discarded; a step 306 where, when conditions are discarded, conditions are masked and an unconditional generation distribution is learned; a step 308 where, when conditions are retained, a conditional generation distribution is learned; a sampling phase step 310 where conditional and unconditional scores are calculated respectively; and a step 312 where the guidance strength is adjusted by weighted interpolation of conditional and unconditional scores using guidance coefficients.

[0043] During training, stiffness tensor conditions are randomly discarded with a preset probability, enabling the network to learn both conditional and unconditional generation distributions simultaneously. Step 302 involves independently performing a conditional discard decision for each sample in each training batch, with a preset probability of p=0.2, meaning approximately 20% of the training samples in that batch do not accept stiffness tensor conditions. Decision 304 determines whether to discard conditions for the current sample based on the random sampling results; if conditions are discarded, step 306 is executed; otherwise, step 308 is executed. Step 306 masks the stiffness tensor conditional embedding vector and inputs the zero vector or a special label vector into the conditional adapter, allowing the denoising network to learn the general generation distribution of two-dimensional crystal structures in unconditional mode. Step 308 retains the stiffness tensor conditional embedding vector and injects the complete conditional information into each layer of the denoising network through the conditional adapter, enabling the network to learn the conditional generation distribution under given stiffness constraints. By simultaneously learning both conditional and unconditional modes during the training phase, the network gains the ability to flexibly adjust the strength of conditional guidance during the sampling phase.

[0044] During sampling, the conditional and unconditional scores are weighted and interpolated using a guiding coefficient to adjust the guiding strength of the stiffness tensor condition. Step 310 calculates the conditional and unconditional scores at each time step of the reverse denoising process. The conditional score is output by the denoising network that takes the input complete stiffness tensor conditional embedding vector, while the unconditional score is output by the denoising network that takes the input zero vector or a specially labeled vector. Step 312 performs weighted interpolation on the two scores according to the guiding coefficient γ. The interpolation formula is: Guiding score s guide =s uncond +γ(s cond -s uncond The guiding coefficient γ controls the degree of influence of the target stiffness condition on the generated structure. γ=2.0 is the value adopted in this embodiment, determined through ablation by balancing condition compliance and structural diversity. A guiding coefficient γ=0 degenerates into unconditional generation, which is used as the ablation baseline to verify the effectiveness of stiffness condition guidance; a guiding coefficient γ=1 corresponds to conditional generation according to the standard. The larger the guiding coefficient γ, the closer the stiffness component of the generated structure is to the target range, but the structural diversity decreases; the smaller the guiding coefficient γ, the higher the diversity of the generated structure, but the higher the probability of the stiffness component deviating from the target range.

[0045] The loss function is a weighted sum of the denoising objectives for three types of variables: atomic type, fractional coordinates, and lattice. The model simultaneously optimizes the generation quality of these three variables through backpropagation. Cross-entropy loss is used for atomic type to measure the difference between the predicted type distribution and the true type; mean squared error loss is used for fractional coordinates to measure the difference between the predicted coordinate noise and the true noise; and mean squared error loss is used for lattice to measure the difference between the predicted lattice noise and the true noise. The three losses are summed according to preset weights to form the total loss. The weight settings reflect the relative contribution of each variable to the generation quality of the 2D crystal structure. Conditional adapters are set in multiple interaction blocks to ensure that stiffness conditions continuously participate in the feature update process at different levels.

[0046] Reference Figure 4 This invention also provides a process 400 for screening and verifying generated structures. Process 400 includes a step 402 of screening the thermodynamic stability of the generated two-dimensional crystal structure, a step 404 of performing first-principles calculations on the retained structure, a step 406 of determining whether it meets the target stiffness condition, a step 408 of adjusting the guiding coefficient and regenerating, and a step 410 of outputting a two-dimensional crystal structure that meets the target stiffness condition.

[0047] Step 402 involves thermodynamic stability screening of the generated two-dimensional crystal structures, retaining structures with convex hull energies below a preset threshold. In some embodiments, a surrogate model finely tuned with two-dimensional material energy data is used to predict the energy of the generated structure, and the energy E on the convex hull is calculated in conjunction with the energy of the reference phase convex hull. hull The preset threshold is 0.1 eV / atom. The energy on the convex hull reflects the thermodynamic stability of the material relative to a known stable phase with the same chemical composition; structures below the threshold are feasible for synthesis. The surrogate model is several orders of magnitude faster than density functional theory, making it suitable for rapid screening of a large number of generated structures.

[0048] Step 404 performs first-principles calculations on the retained structure to verify the degree of agreement between the actual elastic constants and the target stiffness condition. In some implementations, the first-principles method employs VASP, using the energy-strain method to calculate the elastic constants. The energy-strain method applies a small strain to the lattice, calculates the total energy of the strained state, and obtains the elastic constant matrix by fitting the second derivative of energy with respect to strain. For two-dimensional materials, based on the in-plane area A0 of the unit cell, the two-dimensional in-plane stiffness is obtained through the second derivative of energy-strain, using the following formula: This ensures that the stiffness unit is N / m. The calculation uses a PBE exchange-correlated functional, with the plane wave cutoff energy set at 520 eV, and the reciprocal space k-point spacing not exceeding 0.03 Å. -1 The energy convergence criterion is 10. -6eV, with a force convergence criterion of 0.01 eV / Å. Six independent strain modes were applied to each structure with strain amplitudes of ±0.5%, ±1.0%, and ±1.5%, and the components of the stiffness matrix were obtained by fitting energy-strain curves.

[0049] The determination 406 uses the actual elastic constants calculated based on first-principles calculations to determine whether the generated structure meets the target stiffness conditions. The compliance criterion is that the actual values ​​of the target stiffness components fall within the target range. For multi-component joint constraints, compliance is determined when all components simultaneously satisfy their corresponding ranges. In some implementations, each set of target stiffness conditions generates 64 candidate two-dimensional crystal structures. The target stiffness conditions include 8 sets of stiffness targets, covering C... 11 The range is from 10 to 200 N / m. For 80 representative structures selected from the pool, the actual elastic constants were calculated using first-principles methods, and the guiding effect of the conditions was quantified using the Pearson correlation coefficient. DFT verification results show that all four physically independent in-plane stiffness components are significantly correlated with their respective target values, C... 11 The Pearson correlation coefficient r = 0.839, C 12 The Pearson correlation coefficient r = 0.705, C 22 The Pearson correlation coefficient r = 0.788, C 66 The Pearson correlation coefficient was r=0.660, the sample size was n=80, and the significance level was p<0.001. The Pearson correlation coefficient measures the degree of linear correlation between the target stiffness component and the actual stiffness component. A coefficient close to 1 indicates that the condition guidance effectively drives the generated structure to shift toward the target.

[0050] When the generated result deviates from the target stiffness condition, step 408 adjusts the guiding coefficient and regenerates the structure. The adjustment range of the guiding coefficient is determined according to the degree of deviation; when the deviation is large, the guiding coefficient is increased to strengthen the condition constraint, and when the deviation is small, the guiding coefficient is fine-tuned to balance compliance and diversity. The adjusted guiding coefficient is used as a new sampling parameter, and the reverse denoising process in step 114 is re-executed to generate a new batch of candidate structures, which are then re-entered into process 400 for screening and verification. The iterative process continues until a sufficient number of structures that meet the target stiffness condition are obtained.

[0051] In some implementations, C is unconditionally generated (guiding coefficient γ=0). 11 The mean value is 33.9 N / m, and the condition-guided generation (guidance coefficient γ=2.0) yields C values ​​for each group. 11 The mean value monotonically increases with the target quantile to approximately 178 N / m, representing a maximum increase of approximately 5.25 times relative to the unconditionally generated baseline, thus verifying the effectiveness of the stiffness tensor condition guidance.

[0052] Step 410 outputs a two-dimensional crystal structure that satisfies the target stiffness condition. The output structure includes the lattice matrix, atomic coordinates, atom types, and elastic constants verified using first-principles calculations. In some implementations, the effectiveness, uniqueness, and novelty of the generated structures are statistically evaluated (SUN composite index). Effectiveness refers to the proportion of structures selected through thermodynamic stability screening; uniqueness refers to the proportion of independent structures after removing duplicate structures; and novelty refers to the proportion of new structures not appearing in the training set. The SUN composite index comprehensively evaluates the ability of the generated model to maintain structural diversity and innovation while meeting the constraints, providing a quantitative evaluation basis for the on-demand design of two-dimensional mechanical functional materials.

[0053] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a system including a processing module or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0054] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diffusion generation of material crystal structures guided by stiffness tensor conditions, comprising the following steps: Obtain training samples containing two-dimensional crystal structures and corresponding in-plane stiffness tensors, wherein the training samples include lattice matrices, atomic coordinates, and in-plane stiffness tensors; Based on the layer normals of the two-dimensional lattice, the lengths and angles of the in-plane lattice basis vectors, and the preset basis vector ordering and orientation determination rules, a unique standard coordinate system is established that is fixed to the two-dimensional crystal structure. Based on the rotational transformation relationship from the original coordinate system to the unique canonical coordinate system, the lattice matrix, atomic coordinates and in-plane stiffness tensor in the training samples are synchronously transformed, wherein the in-plane stiffness tensor is transformed according to the fourth-order tensor transformation rule; The transformed in-plane stiffness tensor is constructed into a joint condition containing multiple independent stiffness components, and then mapped into a fixed-dimensional condition embedding vector through the stiffness tensor condition embedding module. The conditional embedding vector is injected layer by layer into a pre-trained equivariant diffusion denoising network through a conditional adapter. With the main weights of the pre-trained isovariant diffusion denoising network frozen, the parameters of the stiffness tensor conditional embedding module and the conditional adapter are trained based on the training samples. Obtain the target stiffness condition expressed according to the component definition rules of the unique standard coordinate system; Map the target stiffness condition to a target condition embedding vector; as well as Starting from random noise, the target condition embedding vector is input into the trained network to perform a reverse denoising process, generating a two-dimensional crystal structure corresponding to the target stiffness condition.

2. The method according to claim 1, characterized in that, The steps to construct a joint condition containing multiple independent stiffness components from the transformed in-plane stiffness tensor include: The transformed in-plane stiffness tensor is expressed as a stiffness matrix using Voigt notation; and The multiple matrix elements of the stiffness matrix are flattened into a vector in a fixed order, and the vector contains multiple independent stiffness components.

3. The method according to claim 2, characterized in that, The multiple independent stiffness components include C 11 C 12 C 22 and C 66 .

4. The method according to claim 1, characterized in that, The steps of mapping the stiffness tensor conditional embedding module to a fixed-dimensional conditional embedding vector include: The stiffness components in the combined conditions are standardized; and The standardized stiffness component is input into a multilayer perceptron network including multiple fully connected layers and activation function layers, and the output is a fixed-dimensional conditional embedding vector. The dimension of the conditional embedding vector is consistent with the hidden dimension of the pre-trained equivariant diffusion denoising network.

5. The method according to claim 1, characterized in that, The conditional adapter is inserted into each interactive block of the pre-trained isovariant diffusion denoising network to inject the conditional embedding vector into the network in a layer-by-layer modulation manner.

6. The method according to claim 1, characterized in that, The steps for training the parameters of the stiffness tensor conditional embedding module and the conditional adapter employ a classifier-guided strategy, including: During training, stiffness tensor conditions are randomly discarded with a preset probability, enabling the network to learn both conditionally generated and unconditionally generated distributions simultaneously; and During the sampling process, a weighted interpolation of the conditional and unconditional scores is performed using a guiding coefficient to adjust the guiding strength of the stiffness tensor condition, where the guiding score s guide =s uncond +γ(s cond -s uncond ), where γ is the guiding coefficient, and s cond For conditional scoring, s uncond This is an unconditional score.

7. The method according to claim 1, characterized in that, The conditional adapter broadcasts the conditional embedding vector to each atom node and splices or modulates it only with the zero-order scalar feature channel in the pre-trained isovariant diffusion denoising network.

8. The method according to claim 1, characterized in that, Also includes: The thermodynamic stability of the generated two-dimensional crystal structures is screened, and structures with energies below a preset threshold on the convex hull are retained.

9. The method according to claim 1, characterized in that, Also includes: First-principles calculations were performed on the generated two-dimensional crystal structure to verify the degree of agreement between the actual elastic constants and the target stiffness conditions. as well as When the generated result deviates from the target stiffness condition, adjust the guiding coefficient and regenerate.

10. The method according to claim 1, characterized in that, The steps to establish a unique normalized coordinate system include: The gauge z-axis is determined based on the layer normals of the two-dimensional lattice; Sort the in-plane lattice basis vectors by length from smallest to largest, and select the direction corresponding to the shorter basis vector as the candidate x-axis direction; When the difference between the lengths of two basis vectors is less than a preset threshold, a unique canonical x-axis direction is determined according to the basis vector angle or a preset direction priority rule; and Determine the standard y-axis using a right-handed coordinate system.