A method, system, device, and storage medium for supervising fine-tuning data selection of a crystal material generation model.

By employing multidimensional hierarchical allocation, element distribution correction, and structural farthest point sampling methods, the imbalance problem in data selection during supervised fine-tuning of crystal material generation models is solved, thereby improving the effectiveness, uniqueness, and novelty of the generation results. This approach is applicable to various generation model backbones.

CN123120119APending Publication Date: 2026-11-17BEIJING LIANGKUN TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611438264.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-16
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for supervising fine-tuning data selection in crystal material generation models cannot effectively balance the coverage of elemental and crystal structure distributions, structural repeatability, and the effectiveness of the generation model, leading to unpredictable contradictions between the effectiveness, uniqueness, novelty, and stability of the generation results.

Method used

A multidimensional hierarchical allocation, element distribution correction, and structural farthest point sampling method is adopted. By using quality weight, element distribution correction weight, and structural descriptor formation selection priority, a target data subset is constructed to ensure that the data selection process achieves a balance between thermodynamic quality, distribution coverage, and structural diversity. The LeMat-GenBench index is used for feedback control.

Benefits of technology

It improves the MSUN+SUN index of generative models, enhances the effectiveness, uniqueness and novelty of the generated results, is applicable to the backbone of various generative models, and can explicitly control the impact of stability and metastability, and realizes interpretable control of data subsets.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120119A_ABST
    Figure CN123120119A_ABST
Patent Text Reader

Abstract

This invention discloses a method, system, device, and storage medium for supervising fine-tuning data selection in a crystal material generation model. The method includes: constructing metadata from acquired crystal material training corpus, including element quantity layers, atom quantity layers, crystal system layers, and structural descriptors; calculating quality weights based on convex hull energies; constructing multi-dimensional layers based on element quantity, atom quantity, and crystal system, and allocating layer quotas; calculating element distribution correction weights for candidate samples based on the element occurrence rate of the entire training corpus; forming selection priorities within each layer based on quality weights and distribution correction weights; and performing farthest-point sampling based on structural descriptors to obtain a target data subset for supervised fine-tuning of the crystal material generation model. This invention achieves a balance between quality improvement and distribution preservation through the coordinated use of quality weighting, layer quotas, element distribution correction, and farthest-point sampling. The method also supports explicit control of the ratio of samples on the convex hull to near-convex hull metastable samples.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence-assisted material design, crystal structure generation, and machine learning data engineering technology, and more specifically to a method, system, device, and storage medium for supervising fine-tuning data selection of crystal material generation models. Background Technology

[0002] Crystal material generation models typically learn atomic elements, atomic numbers, lattice structures, space groups, periodic coordinates, and their coupling distributions from crystal structure databases, and generate candidate crystals during the inference phase. Common generation models include flow-matching models, diffusion models, Transformers, autoregressive models, and isomorphic graph generation models. Supervised fine-tuning (SFT) can shift the generation distribution towards a specific target region without changing the model backbone, making it suitable for improving the generation rate of low-energy, stable, or metastable crystals.

[0003] However, the SFT data for crystal material generation models are not random samples or simple sets of high-scoring samples as in ordinary text models. If only the top-k samples are selected according to the energy on the convex hull from low to high, the elemental system, number of elements, number of atoms, crystal system, and structural morphology are easily shifted, leading to unpredictable trade-offs between the effectiveness, uniqueness, novelty, stability, and metastable novelty and uniqueness of the model generation results. If only the structure on the convex hull is retained, although it may improve stability, it may excessively shrink the structural space and reduce novelty and metastable uniqueness. If the energy threshold is set too wide, the SFT data will be too similar to the full corpus, making it difficult to generate effective distribution shifts.

[0004] After comparing existing publicly available technical approaches with locally developed three-stage schemes for crystal structure generation, periodic position encoding, and flow matching-SFT-reinforcement learning, the inventors found that existing schemes mainly focus on the backbone of the generation model, position encoding, flow matching paths, or post-training strategies. No publicly available scheme has yet been found that unifies the thermodynamic quality, layered distribution preservation, elemental coverage, structural diversity, and generation benchmark feedback of crystal material SFT data into a single executable data selection algorithm. In particular, the following is lacking:

[0005] First, a data selection mechanism is designed for the SFT crystal generation model rather than general classification / regression tasks; Second, it is a process that can explicitly control the ratio of thermodynamically stable structures (on-hull) to near-hull metastable structures and maintain their distribution in each pool; Third, it has a mechanism that can establish a systematic linear correlation between training subset statistics and LeMat-GenBench indices for reverse parameter tuning.

[0006] Furthermore, existing data selection methods for general image or text data cannot be directly adapted to crystal structure data because crystal data has a periodic symmetric structure, long-range interactions between different atoms, and the generative model is extremely sensitive to the energy distribution, elemental distribution, and structural diversity of the training data. Therefore, a data selection method for supervised fine-tuning of crystal material generative models is needed that does not rely on a specific generative model backbone, can achieve a balance between quality improvement and distribution preservation, and can control the output through structural diversity sampling and index feedback. Summary of the Invention

[0007] The objective of this invention is to address the technical problem of generating a subset of training data from the original training corpus that possesses low energy on the convex hull, maintains elemental and crystal structure distribution coverage, avoids structural repetition, and allows for interpretable adjustment of key LeMat-GenBench metrics during supervised fine-tuning of crystal material generation models. In view of this, this invention proposes a method, system, device, and storage medium for selecting training data for supervised fine-tuning of crystal material generation models.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, embodiments of the present invention provide a method for supervising and fine-tuning data selection for a crystal material generation model, comprising the following steps: Acquire training corpus of crystal materials, wherein each sample in the training corpus includes at least a sample identifier, energy on the convex hull, elemental composition, and crystal structure representation; Crystal metadata is constructed from the training corpus, and the crystal metadata includes element number hierarchy, atomic number hierarchy, crystal system hierarchy and structure descriptor; The mass weight is calculated based on the energy on the convex hull; Multidimensional hierarchical structures are constructed based on element quantity, atomic quantity, and crystal system, and hierarchical quotas are allocated accordingly. The element distribution correction weights for candidate samples are calculated based on the element occurrence rates of the entire training corpus. Within each stratified candidate pool, selection priorities are formed based on the quality weight and element distribution correction weight, and farthest point sampling is performed according to the structure descriptor to obtain a target data subset for supervised fine-tuning of the crystal material generation model.

[0009] In one embodiment, obtaining the crystal material training corpus includes: reading only the training partition and not the test partition as supervised fine-tuning data; deduplicating or marking duplicate samples, samples with missing key fields, and samples with failed structure parsing; and writing the original fields and filtered metadata together into the output file to form a traceable data record.

[0010] In one embodiment, the quality weight is: w Q (x)=exp(-E hull (x) / τ) Where x is the crystal sample, E hull (x) represents the energy on the convex hull, τ represents the mass temperature parameter; and E... hull Samples not exceeding a preset threshold T are defined as the low-energy core candidate pool, and E is... hull Samples greater than T are used as the backfill candidate pool.

[0011] In one embodiment, before performing farthest-point sampling based on the structural descriptor, the method further includes: explicitly dividing the candidate sample dataset into an on-hull pool and a near-hull pool, wherein the on-hull pool satisfies |E hull |≤ε, the near-hull pool satisfies ε <E hull ≤T; ε represents the set on-hull threshold; T represents the set near-hull threshold; Based on the target data subset size K and the preset on-hull ratio ρ, determine the target quotas k0 for the two pools respectively. ρ) and k1=K-k0, perform hierarchical quota, element distribution correction and farthest point sampling in the two pools respectively and then merge the output; round(·) is the rounding function.

[0012] In one embodiment, the multidimensional layering is determined by a ternary set s = (Nelement-bin, Natom-bin, crystal-system), wherein Nelement-bin includes at least the intervals of elements 1, 2, 3, 4, and 5+; Natom-bin includes at least the intervals of atoms 1-4, 5-8, 9-12, 13-16, 17-20, and 21+; and crystal-system includes at least triclinic, monoclinic, orthorhombic, tetragonal, trigonal, hexagonal, and cubic crystal systems; and each layering quota satisfies q. s With N s α Proportional, N s q is the stratified sample size, α is the quota exponent and is less than 1; s This represents the sampling ratio for each layer.

[0013] In one embodiment, the element distribution correction weight is:

[0014] Where x is the crystal sample, Z(x) represents the distribution of the entire crystal sample, Z represents sampling from the sample, and p full(Z) represents the occurrence rate of element Z in the entire training corpus, p core (Z) represents the occurrence rate of element Z in the low-energy core candidate pool, γ is the correction intensity, δ is the smoothing constant, and clip represents the cutoff function. , = (0.5, 0.5, 2.0).

[0015] In one embodiment, the selection priority is w(x) = w Q (x)w E (x); w Q (x) represents the quality weight; w E (x) represents the element distribution correction weight; In each stratum, a candidate pool is selected according to the selection priority. The size of the candidate pool is a preset multiple of the quota of the corresponding stratum, and the farthest point sampling is performed from the candidate pool.

[0016] In one embodiment, the structural descriptor includes at least two of the following: Dimensionless lattice ratio obtained by normalizing the geometric mean of lattice edge lengths; The normalized lattice angle is obtained by dividing the lattice angle by 180 degrees. The characteristic of the unit atomic volume obtained by dividing the unit cell volume by the number of atoms and then taking the logarithm. Truncation of atomic number characteristic; And the minimum mirror atom pair distance histogram calculated based on periodic boundary conditions and 27 periodic mirrors.

[0017] In one embodiment, the farthest point sampling employs a greedy iteration: First, select the candidate sample with the highest priority; Then, for each unselected sample, the minimum structural descriptor distance between it and the set of selected samples is calculated, and the minimum distance is multiplied by a preset power of the normalized selection priority to obtain the sampling score. The sample with the highest sampling score is selected until the stratified quota is reached.

[0018] In one embodiment, when the sample size of an on-hull pool or a near-hull pool is less than its target quota, all samples in that pool are used and the unmet quota is transferred to another pool. When the combined size of the two pools is still less than the target data subset size, samples from outside the low-energy core candidate pool are backfilled according to selection priority, and the actual proportion, actual quantity, and backfill quantity are recorded in the data summary.

[0019] In one embodiment, a candidate parameter control step is also included: By changing at least one of the following factors: mass temperature τ, energy threshold T, on-hull ratio ρ, quota index α, element correction intensity γ, structure candidate pool multiple, and farthest point sampling quality power, multiple candidate supervised fine-tuning data subsets are generated. Calculate the training distribution description vector u for each candidate subset, and obtain the LeMat-GenBench index y under the same generative model, the same number of fine-tuning steps, and the same evaluation metric, fitting y≈β0+β T u, select the target data subset based on the fitting results.

[0020] In one embodiment, the training distribution description vector u includes at least three of the following: on-hull proportion, average energy on the convex hull, low energy proportion, proportion of ternary structure, proportion of 9-12 atom structure, proportion of cubic crystal system, and element occurrence rate JS divergence. The LeMat-GenBench metrics include at least three of the following: Valid, Unique, Novel, Stable, Metastable, SUN, MSUN, and MSUN+SUN. The fitting is a single-factor or multi-factor linear regression, and the correlation coefficient is used as the basis for parameter selection rather than as proof of causality.

[0021] In one embodiment, the output target data subset includes the original crystal structure field, filtering mode, quality weight, element correction weight, selection priority, hierarchical affiliation, structure descriptor configuration, random seed, and a distribution statistical summary of the full corpus and the target subset.

[0022] Secondly, embodiments of the present invention provide a data selection system for supervised fine-tuning of crystal material generation models, comprising: The data acquisition module is used to acquire training data for crystal materials. Each sample in the training data includes at least a sample identifier, energy on the convex hull, elemental composition, and crystal structure representation. A metadata construction module is used to construct crystal metadata from the training corpus, wherein the crystal metadata includes element number hierarchy, atom number hierarchy, crystal system hierarchy and structure descriptor; A mass weight calculation module is used to calculate the mass weight based on the energy on the convex hull. The hierarchical quota allocation module is used to construct multidimensional hierarchies based on element quantity, atomic quantity, and crystal system, and to allocate hierarchical quotas. The distribution correction weight calculation module is used to calculate the element distribution correction weight of the candidate samples based on the element occurrence rate of the full training corpus. The priority formation module is used to form a selection priority in each hierarchical candidate pool based on the quality weight and element distribution correction weight. The diversity sampling module is used to perform farthest point sampling based on the structure descriptor to obtain a target data subset for supervised fine-tuning of the crystal material generation model.

[0023] Thirdly, embodiments of the present invention provide an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in any one of the embodiments of the first aspect.

[0024] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in any one of the embodiments of the first aspect.

[0025] As can be seen from the above technical solution, compared with the prior art, the present invention has the following technical advantages: (i) This invention is the first to define data selection for the SFT model of crystal material generation as an independent data algorithm problem, and incorporates thermodynamic quality, distribution coverage, elemental correction and structural diversity into the same executable process.

[0026] (ii) Compared with taking only the energy samples on the top-k lowest convex hull, this invention reduces the training distribution offset by using hierarchical quotas, element distribution correction and sampling of the farthest point of the structure, thereby reducing the loss of effectiveness, uniqueness and novelty of the generated model.

[0027] (iii) Through explicit on-hull / near-hull dual-pool control, the effects of stability and metastability on MSUN, SUN and their combined indices can be studied separately, and safe backfilling under different data pool sizes can be supported.

[0028] (iv) In the local experiment, the SFT data of this invention was used to perform supervised fine-tuning on MP-20 and Alex-MP20. MSUN+SUN improved by 2.84 and 2.32 percentage points respectively compared with the corresponding full corpus training base. The experiment adopted a unified evaluation caliber and conducted multi-random seed evaluation.

[0029] (v) This invention identifies systematic linear correlation factors between training subset statistics and LeMat-GenBench key indicators, and can use the influence of data subset parameters on Valid, Novel, Stable, SUN and MSUN for targeted adjustment.

[0030] (vi) This invention is located on the preprocessing side of the generative model and does not require changes to the generative model backbone, loss function or inference code. Therefore, it can be extended to other generative model backbones such as flow matching, diffusion, Transformer, autoregression and isovariate graph, and can be adapted to MP-20, Alex-MP20 and other crystal datasets with structure and energy labels. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0032] Figure 1 A flowchart of the method for supervising and fine-tuning data selection for crystal material generation models provided in this embodiment of the invention.

[0033] Figure 2 This diagram illustrates the linear correlation between training subset statistics and the LeMat-GenBench metric on the MP-20 dataset provided in this embodiment of the invention.

[0034] Figure 3 This diagram illustrates the linear correlation between training subset statistics and LeMat-GenBench metrics on the Alex-MP20 dataset, as provided in this embodiment of the invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Terminology Explanation: To facilitate understanding of this invention, the technical terms appearing in this invention will first be explained: Energy above hull (E hull E is a commonly used measure of the thermodynamic stability of crystalline materials, representing the energy difference of the crystal structure relative to the thermodynamic convex hull. hull =0 indicates that it is located on the convex hull and is a thermodynamically stable structure; E hull >0 indicates a metastable or unstable structure.

[0037] On-hull samples: These refer to samples whose energy on the convex hull is zero or close to zero (|E0).hull A crystal structure with |≤ε) is a thermodynamically stable structure. ε represents the set on-hull threshold, such as 1e-12ev / atom.

[0038] Near-hull sample: refers to a crystal structure on the convex hull with energy between ε and T, i.e., a metastable structure near the convex hull; T represents the set threshold for near-hull, such as 0.025 ev / atom.

[0039] Supervised Fine-Tuning (SFT): Based on a pre-trained generative model, supervised training is performed using carefully selected data to shift the model's output distribution toward the target distribution.

[0040] LeMat-GenBench metrics: A set of benchmark metrics for evaluating material generation models, including Valid (generated structure validity), Unique, Novel, Stable, Metastable, SUN (stable, unique, and novel), MSUN (metastable, unique, and novel), and MSUN+SUN (combined metrics).

[0041] Example 1: This invention provides a method for supervising the fine-tuning of SFT data in a crystal material generation model, comprising: breaking down SFT data selection into six operations: data pattern normalization, quality assessment, distribution balancing, structural diversity screening, controllable pool merging, and index feedback. The overall process flow is as follows: Figure 1 As shown: The first step is to obtain training data for crystal materials. Each sample in the training data includes at least a sample identifier, energy on the convex hull, elemental composition, and crystal structure representation.

[0042] During implementation, for example, when reading the input crystal material training corpus, the input file can be a CSV file in MP-20 or Alex-MP20 format, containing at least the following fields: sample identifier (material_id), energy / stability field (e_above_hull), elemental composition (elements), crystal structure file (cif), and space group field (spacegroup_number). A unified mapping is performed to address differences in field naming; for Hermann-Mauguin space group symbols in Alex-MP20, unique symbols are first cached before the symbol-to-space group number mapping is completed to avoid repetitive calculations line by line.

[0043] Step 2, Quality Assessment and Quality Weighting: Crystal metadata is constructed from the training corpus mentioned above. This crystal metadata includes element number hierarchy, atomic number hierarchy, crystal system hierarchy, and structure descriptor.

[0044] For example, the number of elements is obtained by composing elements; atomic coordinates, lattice edge lengths, lattice angles, unit cell volumes, and number of atoms are resolved from the crystal structure file (cif); and the crystal system is mapped by the space group (spacegroup_number). In particular, when the symmetry field in the structure file header is inconsistent with the original space group field in the data table, the space group field in the data table is used for crystal system layering, and the structure file is used for structure descriptor calculation.

[0045] Mass weights are calculated based on the energy on the convex hull.

[0046] Calculate w for each sample Q (x)=exp(-E hull (x) / τ) Where x is the crystal sample, E hull (x) represents the energy on the convex hull, τ represents the mass temperature parameter; and E... hull Samples not exceeding a preset threshold T are defined as the low-energy core candidate pool, and E is... hull Samples greater than T are used as the backfill candidate pool.

[0047] Unlike directly deleting non-low-energy samples, mass weighting elevates low-energy samples to a higher selection probability while retaining non-core samples within the same stratum as a source of coverage and backfill. The recommended energy threshold T is between 0.01 and 0.05 eV / atom, and the recommended mass temperature τ is between 0.01 and 0.05 eV / atom.

[0048] The third step is distribution balancing: A multi-dimensional hierarchy is constructed based on element count, atom count, and crystal system, and hierarchical quotas are allocated. Element distribution correction weights are calculated for candidate samples based on the element occurrence rate of the entire training corpus. This achieves the following: maintaining the element count distribution, maintaining the atom count distribution, maintaining the crystal system distribution, and controlling element coverage to prevent excessive shifts.

[0049] The above process uses elements to parse the number of elements, cif to parse the number of atoms, and spacegroup_number to map the crystal system. These three elements are combined to form a multi-dimensional hierarchy, and a hierarchical quota is used to ensure that rare types are covered. Then, a correction weight w is calculated based on the occurrence rate of elements in the full corpus and the core pool. E (x).

[0050] in: (1) Layered quota: Multidimensional layering is constructed based on the number of elements, the number of atoms, and the crystal system, and layered quotas are allocated.

[0051] Mapping each sample to the Cartesian product of element number layering, atomic number layering, and crystal system layering yields the layer s. For the number of layers N... s Temperature-based quota q s ∝N s α α is less than 1, to moderately increase the retention chances of rare strata without forcibly homogenizing a large number of minimal strata. Optionally, a minimum guaranteed quota can be set for each stratum, and integerization can be performed using the maximum remainder method.

[0052] q s N represents the target sampling quota obtained from layer s. s This represents the sample size of this stratum in the candidate corpus, where α is the quota index. According to q... s ∝N s α Calculate and ensure that the total quota equals K by integerization.

[0053] (2) Element distribution correction: Calculate the element distribution correction weight for candidate samples based on the element occurrence rate of the full training corpus.

[0054] The element distribution correction weights are:

[0055] Where x is the crystal sample, Z(x) represents the distribution of the entire crystal sample, Z represents sampling from the sample, and p full (Z) represents the occurrence rate of element Z in the entire training corpus, p core (Z) represents the occurrence rate of element Z in the low-energy core candidate pool, γ is the correction intensity, δ is the smoothing constant, and clip represents the cutoff function. , = (0.5, 0.5, 2.0). and These represent the minimum and maximum values ​​of the element distribution correction weights, respectively. In this embodiment, they are 0.5 and 2.0. The truncation operation ensures that the correction weights are within this range, preventing rare elements from causing the weights to become out of control.

[0056] Low-energy cores typically contain or lack specific elements. Calculate the occurrence rate p of elements across the entire corpus. full (Z) and the occurrence rate p of the core candidate pool elements core (Z), w is composed of the geometric mean of the ratios of the occurrence rates of elements. E (x), then adjust the correction intensity with γ and truncate the result to [w min ,w max ].

[0057] The fourth step is structural diversity screening.

[0058] Figure 1 In step 4, RDF stands for Radial Distribution Function, used to statistically analyze the distance distribution of atomic pairs in periodic crystals and form structure descriptors; FPS stands for Farthest Point Sampling, used to iteratively select samples least similar to the already selected structures to reduce duplication. Specifically, it includes the following sub-steps: (1) Priority selection: Within each stratified candidate pool, weights are adjusted based on quality weights and element distribution to form selection priorities: w(x)=w Q (x)w E (x) Among them, w Q (x) represents the quality weight; w E (x) represents the element distribution correction weight. This operation is used to suppress systematic loss of element coverage, rather than forcing the element distribution to be exactly the same.

[0059] (2) Structural description: For each CIF (cif field), parse the structure descriptor, including the scale-independent lattice ratio, normalized lattice angle, logarithm of unit atomic volume, number of atoms, and the periodic minimum mirror pair distance histogram.

[0060] (3) Screening: After dimensionally standardizing all descriptors, weighted farthest-point sampling is used in each hierarchical quality candidate pool. The highest-priority sample is selected first, and subsequent samples are selected based on the product of its minimum descriptor distance to the selected set and the power of its quality priority. This avoids outputting a large number of similar lattices or similar local geometries.

[0061] (4) Explicit pool ratio control: For SFT experiments that require distinguishing between stable and metastable structures, the data is split into on-hull pools (|E hull |≤ε) and the near-hull pool (ε) <E hull ≤T), perform stratification, element correction, and structural sampling respectively, and then merge according to the target ratio. This design makes the on-hull ratio an explicit control variable, no longer indirectly determined by soft quality weights; when a pool is insufficient, perform auditable quota transfer and backfilling.

[0062] Step 5: Output the target SFT subset: This ultimately yields a high-quality, distribution-preserving, and structurally diverse data subset for supervised fine-tuning of the downstream generative model. Specifically, the selected samples have low energy on their convex hulls (E...).hull The selected samples are mainly thermodynamically stable or near-stable structures; and the selected samples are significantly different from each other in structure, avoiding the repetition of a large number of similar lattices, similar local atomic arrangements or similar geometric configurations; and ensuring that the distribution of the selected subset in terms of elemental composition, number of atoms, crystal system and other dimensions is close to that of the full training corpus, avoiding systematic “unbalance” due to only selecting low-energy samples.

[0063] Step 6: Indicator feedback control.

[0064] Multiple candidate parameter combinations generate data subsets of the same size, and the statistical descriptive vector u of each subset is recorded. Using the same generative model backbone, the same number of SFT steps, and the same sampling and relaxation methods, the LeMat-GenBench index y is obtained. A single-factor or multi-factor linear relationship y≈β0+βTu is fitted, and the parameters for the next round of data selection are selected based on the correlation coefficient, fitting direction, and constraints. This step extends data selection from static screening to interpretable closed-loop control.

[0065] In other words, the evaluation results of the generative model on LeMat-GenBench are correlated with the statistical descriptive vectors of the subset, and the parameters of the selected data are adjusted in reverse to form a closed loop. After selecting the target data subset, supervised fine-tuning is performed by continuing flow matching as the training objective.

[0066] This invention achieves a balance between quality improvement and distribution preservation through quality weighting, hierarchical quotas, element distribution correction, and collaborative sampling of the farthest point of the structure. The method also supports explicit control over the ratio of samples on the convex hull to metastable samples near the convex hull. The core of the supervised fine-tuning data selection method for crystal material generation models provided by this invention is not to change the generation model architecture, but to construct a target SFT subset on the training data side using quality weighting, hierarchical distribution preservation, element coverage correction, and structural diversity sampling. Furthermore, controllable parameter tuning is achieved through explicit energy pool ratio and LeMat-GenBench linear correlation analysis. This method is auditable, reproducible, scalable, and decoupled from the model backbone.

[0067] The above-described method steps of the present invention are further described in detail below: Example 1: MP-20 format normalization and metadata construction; like Figure 1As shown, the MP-20 training CSV is read, with fields including material_id, e_above_hull, elements, cif, and spacegroup_number. Elements are parsed into sets of elements, with the length of each set mapped to five categories: 1, 2, 3, 4, and 5+. Lattice parameters and atomic fraction coordinates are read from the cif to obtain the number of atoms, which is then mapped to six categories: 01-04, 05-08, 09-12, 13-16, 17-20, and 21+. Seven crystal systems are mapped based on the spacegroup_number. In cases where the space group field in the CSV differs from the cif header, the crystal system is determined by the CSV field; the cif is only used for the structure descriptor.

[0068] In one specific implementation, the field naming and spatial group representation methods differ between datasets from different sources. Therefore, pattern normalization is performed before metadata construction. Table 1 shows the field mapping relationships between two typical corpora: MP-20 and Alex-MP20.

[0069] Table 1

[0070] After completing the space group number mapping, seven crystal systems were mapped according to the number intervals. Table 2 shows the correspondence between space group numbers and crystal systems.

[0071] Table 2

[0072] For fields represented in Hermann-Mauguin notation, this invention first traverses the corpus to collect all unique symbols, constructs a "symbol → number" mapping cache, and then performs a table lookup for each row to complete the conversion, thereby avoiding the computational overhead caused by repeatedly calling the space group resolution library for the same symbol. For samples that fail to resolve the structure, such as those with invalid CIF format, singular lattice matrix, or out-of-bounds atomic coordinates, they are marked as unusable and excluded in subsequent steps, while the reasons for exclusion are recorded for traceability.

[0073] Example 2: Joint screening of quality, distribution, and diversity; The target data subset size is set to K=5000, the mass threshold T=0.025 eV / atom, the mass temperature τ=0.02 eV / atom, the quota index α=0.7, the elemental correction intensity γ=0.5, the elemental correction cutoff range is 0.5 to 2.0, the structure distance histogram uses 32 intervals with a maximum distance of 8 Å, the candidate pool multiplier is 4, and the mass power of the farthest point sampling is 0.25. For each sample in the full training corpus, a mass weight and elemental correction weight are calculated. Approximately 4K candidate pools are generated in each stratum, and weighted farthest point sampling is performed. The output file retains the original structure fields and outputs a separate JSON summary.

[0074] A summary of a set of 5000 sample data points is as follows: The average E of the selected data is... hull It is 0.00436 eV / atom, with a median of 0, an on-hull percentage of 0.668, and E hull The proportion of atoms with an intensity not exceeding 0.025 eV / atom is 0.946, and the element occurrence rate JS divergence is 0.015839. This abstract is used to demonstrate that the selected data results are auditable and does not constitute a limitation on the scope of protection.

[0075] The following describes the specific calculation methods for the quality weight, element distribution correction weight, and structural descriptor in this embodiment.

[0076] (1) Quality weights. Calculate w for each sample x. Q (x)=exp( E hull (x) / τ), where x is the crystal sample, E hull (x) represents the energy on the convex hull, τ represents the mass temperature parameter; and E... hull Samples not exceeding a preset threshold T are defined as the low-energy core candidate pool, and E is... hull Samples greater than T are used as the backfill candidate pool.

[0077] When E hull =0 when w Q =1; when E hull =τtimew Q =e 1 ≈0.368; when E hull =3τ when w Q ≈0.050. It can be seen that the smaller τ is, the more concentrated the weights are on extremely low-energy samples, and the more aggressive the screening; the larger τ is, the more gradual the weight decay is, and the closer it is to uniform sampling. Therefore, τ is a continuous knob that controls the intensity of quality guidance, while T is a hard cutoff that controls the boundary of the candidate pool. The combination of the two can achieve a continuous transition from approximate top-k to approximate full sampling.

[0078] (2) Element distribution correction weights. Let the set of elements of sample x be Z(x), then:

[0079] Where x is the crystal sample, Z(x) represents the distribution of the entire crystal sample, Z represents sampling from the sample, and p full (Z) represents the occurrence rate of element Z in the entire training corpus (number of samples containing this element / total number of samples), p core (Z) represents the occurrence rate of element Z in the low-energy core candidate pool. If an element is systematically enriched in the low-energy core (p... core (Z)>p full If (Z) is less than 1, the sample is appropriately downweighted; conversely, if an element is scarce in a low-energy core, the sample is appropriately upweighted, thereby suppressing the systematic loss of element coverage. The smoothing constant δ is used to avoid the ratio of division by zero to infinity caused by the zero occurrence rate; in this embodiment, δ = 10. 3 The truncation interval [0.5, 2.0] ensures that the weights will not get out of control due to a few rare elements.

[0080] (3) Structural descriptor. For each sample, the following components are analyzed: scale-independent lattice scale characteristics, i.e., each component of (a,b,c) divided by its geometric mean (abc). 1 / 3 The three dimensionless ratios obtained; normalized lattice angle characteristics, i.e., the three normalized angles obtained by dividing each component of (α, β, γ) by 180 degrees; unit atomic volume characteristics, i.e., ln(V / N), where V is the unit cell volume and N is the number of atoms; atomic number characteristics, i.e., N or its normalized form; and the minimum mirror image atom pair distance histogram calculated based on periodic boundary conditions and 27 periodic mirror images, i.e., within a cutoff distance of 8 Å, the distance of all atom pairs is calculated according to the minimum mirror image convention, and then normalized after statistical histograms are calculated according to 32 equal-width intervals. The minimum mirror image convention means that for atoms i and j in the unit cell, their distance is d. ij =min n∈{ 1,0,1} 3 |r i r j +n·L|, where L is the lattice matrix, and the 27 periodic mirrors, i.e., the three components of n, are each taken as... All combinations of 1, 0, and 1. This histogram remains unchanged under translation, cell selection, and atomic ordering, and can characterize the statistical features of local coordination environments.

[0081] (4) Weighted farthest point sampling. Within each layer s, first select according to the selection priority w(x)=w Q (x)·w E(x) Take the first m·q from high to low. s The candidate pool consists of 10 samples. s (m is a multiple of the candidate pool, which is 4 in this embodiment), and then perform greedy farthest point sampling: the first step is to select C s The highest-priority sample is selected and added to the selected set S; in each subsequent step, a sampling score score(x) = min is calculated for each unselected sample x in the candidate pool. y∈S d(f(x),f(y))· (x) η Where f(·) is the standardized structure descriptor vector, and d(·,·) is the descriptor distance. (x) represents the selection priority for normalization to the [0,1] interval, and η is the quality power (0.25 in this embodiment); the sample with the highest sampling score is added to S, and this process is repeated until |S|=q. s In this process, standardization can employ z-score standardization (with statistics taken from the full training corpus or the current candidate pool), descriptor distance can use standardized weighted Euclidean distance, and histogram components can be synthesized using L1 distance, L2 distance, χ² distance, or bulldozer distance. Since the sampling score is the product of distance and quality rather than a weighted sum, the quality term plays a modulating role in a power-law manner, which can maintain structural diversity while avoiding the selection of outliers with excessively low quality.

[0082] Example 3: Explicit control of on-hull and near-hull ratios; Assuming the on-hull target ratio ρ = 0.7, then an on-hull pool E is established from the full training corpus. hull ≤10 8 and near-hull pool 10 8 <E hull ≤0.025. For K=5000, 3500 and 1500 entries are allocated respectively; the two pools are respectively processed for element count, atom count, crystal system stratification quota, element distribution correction, and sampling of the farthest point of the structure, and finally merged and deduplicated. If the near-hull pool is insufficient, the unmet quota is transferred to the on-hull pool; if the total amount of the low-energy dual pool is insufficient, samples outside the threshold are backfilled, and the actual proportion is output in the abstract. ρ=0 and ρ=1 degenerate into pure near-hull and pure on-hull configurations respectively.

[0083] In a specific implementation, quota transfer and backfill are performed according to the following rules to ensure the process is auditable: let the number of available samples in the on-hull pool be A0, and the number of available samples in the near-hull pool be A1. If A0≥k0 and A1≥k1, the original quota is executed; if A0<k0, all A0 entries in the on-hull pool are taken, and k0 The unmet quota of A0 entries is transferred to the near-hull pool; if A1<k1, all A1 entries in the near-hull pool are taken, and k1 The unmet quota of A1 entries is transferred to the on-hull pool; if the total of the two pools is still less than K, the deficiency is supplemented from the backfill candidate pool with selection priority from E hull >T, and the actual on-hull ratio, the actual quantity of the two pools, and the backfill quantity are recorded in the data summary. Through this mechanism, regardless of the energy distribution of the corpus, the output scale K is always guaranteed, and the deviation between the actual composition and the expected composition is explicitly recorded.

[0084] It should be noted that the on-hull ratio ρ is a continuously adjustable experimental variable. In other embodiments, candidate subsets of the same scale can be generated respectively according to the configuration set of ρ∈{0, 0.3, 0.5, 0.7, 1.0}, and the Valid, Unique, Novel, Stable, Metastable, SUN, MSUN and MSUN+SUN indicators of each subset are evaluated under the same generative model backbone, the same SFT step count and the same evaluation standard, so as to obtain the influence curve of ρ on each indicator, and select the ρ value for the specific target accordingly. Since SUN and MSUN usually change in opposite directions with ρ, in practical applications, a trade-off can be made between the two according to whether the target is more stable or more novel, or the total value of MSUN+SUN can be taken as the comprehensive target.

[0085] It should be further noted that when explicit dual-pool ratio control is not enabled, the process degenerates into a single-pool soft weight mode: at this time, ρ, k0 and k1 are not preset, but the full low-energy core candidate pool is directly weighted by the quality weight w Q (x), the selected samples of each layer are determined by both the selection priority and structural diversity, and the on-hull proportion in the subset is naturally formed by the energy distribution of the corpus itself and the quality temperature τ, rather than being an artificially set explicit control quantity. In other words, excluding on-hull in this context specifically refers to the configuration where explicit dual-pool ratio control is not enabled and the on-hull proportion is not taken as an independent constraint, which does not mean that all on-hull samples are excluded from the corpus; the subset obtained under this configuration can still contain a relatively high proportion of on-hull samples, and the proportion is naturally generated by the screening process and truthfully recorded in the summary.

[0086] Example 4: Systematic identification of linearly correlated factors; Training distribution statistics were recorded on multiple candidate SFT data subsets, and the same crystal generation model, the same number of SFT steps, and the same LeMat-GenBench evaluation process were used for each subset. The on-hull ratio and average E0 of the training subset were used as the benchmarks. hull Using the proportion of low-energy elements, the proportion of three elements, the proportion of 9-12 atoms, the proportion of cubic crystal system, and the JS divergence of element occurrence rate as independent variables, and Valid, Unique, Novel, Stable, Metastable, SUN, MSUN, and MSUN+SUN as dependent variables, Pearson correlation coefficients were calculated and least squares lines were fitted.

[0087] like Figure 2 The figure shows the trend of the percentage of MSUN and MSUN+SUN metrics among the 2500 generated structures after performing SFT on the MP20 dataset, as a function of the on-hull fraction. The horizontal axis represents on-hull fraction (ρ), and the vertical axis represents the percentage of MSUN and MSUN+SUN in the generated structures. Each point corresponds to a subset of SFT data and one unified evaluation, with a sample size of 2500 generated structures.

[0088] Local MP-20 experiments (one SFT model per point, n=10) show that the correlation coefficient between the on-hull ratio and MSUN is [missing value]. 0.90, compared to +0.83 for Sun; the proportions of the three elements compared to Sun are... 0.95, compared to +0.93 for SUN; cubic crystal system percentage compared to MSUN is... 0.95, compared to +0.93 for SUN; the percentage of atoms 9-12 is the same as that for MSUN. 0.88, and SUN is +0.82; element JS divergence and Valid are... 0.88, and Unique are 0.88.

[0089] like Figure 3 The figure shows the trend of MSUN and MSUN+SUN indices as a function of the on-hull fraction in 2500 structures generated after performing SFT on the model on the Alex-MP20 dataset. Local Alex-MP20 experiments (one SFT model per point, n=6) show that the correlation coefficient between the on-hull fraction and MSUN is [value missing]. 0.84, compared to +0.77 for SUN; the proportions of the three elements compared to MSUN are... 0.82, compared to +0.91 for SUN; the cubic crystal system ratio is the same as that for MSUN. 0.68, compared to +0.57 for SUN; average E hull The JS divergence of the element is +0.53 with MSUN and +0.58 with MSUN+SUN; the JS divergence of the element with MSUN is... 0.46, compared to MSUN+SUN 0.51. The above results are used to establish the interpretable regulatory direction of data distribution to indicators; the correlation results themselves are not to be interpreted as strict causation.

[0090] Table 3 shows the representative MSUN+SUN performance of the SFT data selection method of this invention on two types of crystal corpora. The improvements in the table are expressed in percentage points, and the data comes from local experimental records; further supplementation may be made based on the final publicly available experimental records, sample size, and evaluation protocol during the formal application process. The data in the table comes from a local unified comparative experiment on MP-20 and Alex-MP20, with improvements expressed in percentage points, and the comparison object is the training base of the same generative model on the full corpus.

[0091] Table 3

[0092] Under the same generative model backbone, the same number of SFT steps, the same evaluation criteria, and the use of multiple random seeds, the SFT data selection method improves the MSUN+SUN of the full corpus training base on MP-20 and Alex-MP20 by 2.84 and 2.32 percentage points, respectively.

[0093] Implementation 5: Comparison with simple screening; Comparison Option 1 uses random sampling; Comparison Option 2 uses only E. hull The selection process proceeds from low to high; in contrast, scheme three only selects on-hull samples; the present invention employs a combined method of quality weighting, stratified quotas, element distribution correction, and sampling at the structural farthest point. Experimental results show that while selecting only the lowest Ehull or only on-hull samples can improve Stability, it easily reduces Novel and MSUN; the present invention achieves a more favorable overall effect in terms of stability, novelty, and structural diversity by retaining near-convex hull samples and controlling their distribution and structure.

[0094] Table 4 presents representative linear correlation results between training subset statistics and MSUN. The sample sizes for the correlation coefficients are n=10 and n=6, respectively, primarily used for interpretable screening of candidate data subsets.

[0095] Table 4

[0096] It should be noted that the linear correlation analysis described above is used for interpretable directional screening among candidate subsets, rather than establishing a rigorous causal model. When the sample size is limited (e.g., n=6), the absolute values ​​of some correlation coefficients may be lower than the critical value required for statistical significance. In such cases, they should be used as weak evidence and combined with the direction of the effect and domain knowledge for comprehensive judgment. In other embodiments, the number of candidate subsets can be increased, multiple random seed evaluations can be used for each subset to estimate the variance of the index, or leave-one-out cross-validation and regularized regression can be used to improve the stability of the fit. When there is collinearity among multiple independent variables (e.g., the on-hull ratio, the average Ehull ratio, and the low-energy ratio are highly correlated), only the representative variable can be retained, or a single-variable regression can be performed directly.

[0097] Based on the above correlations, this invention provides interpretable control directions as shown in the last column of Table 4: if the goal is to improve the stability of the generated structure (Stable, SUN), then increase the on-hull ratio, the proportion of ternary structures, and the proportion of cubic crystal systems; if the goal is to improve metastable novelty and uniqueness (MSUN), then decrease the above control variables; if the goal is to achieve the optimal combination between the two, then use MSUN+SUN as the optimization objective and search on the candidate set of parameters such as ρ, α, and γ. Thus, data selection is extended from a one-time static screening to an interpretable closed-loop control process.

[0098] In one specific implementation, the four schemes mentioned above maintain complete consistency in target subset size K, generative model backbone, SFT steps, sampling quantity, relaxor, and evaluation criteria, changing only the data subset itself to ensure fairness in the comparison. In addition to the final metrics, the JS divergence of the occurrence rate of subset elements relative to the full corpus, the JS divergence of the crystal system distribution, and the average nearest neighbor distance of the internal structure descriptors of each scheme are recorded to characterize the degree of distribution preservation and structural redundancy: random sampling maintains the best distribution but has the worst quality; Ehull hard threshold filtering and only taking on-hull samples have the best quality but the most severe distribution shift and structural redundancy; the scheme of this invention is in the middle in all three characterization metrics and has the best overall performance.

[0099] Example 6: Extension of other generative model backbones; The data selection method of this invention relies solely on the training corpus fields and structure parsing results, without depending on the network layers, noise scheduler, streaming time parameterization, or decoder within the generative model. For streaming matching models, the output subset can be directly replaced with the streaming matching pre-training corpus; for diffusion models, the output subset can be used for noise-conditional supervised training; for Transformer, autoregressive, or isomorphic graph generation models, the original crystal samples can be passed to the corresponding data loader with the same material_id and structure field. As long as the supervised fine-tuning of the target model can receive crystal structure samples, the target data subset output by this invention can be used.

[0100] In one specific implementation, the interface is as follows: For autoregressive and language model backbones that take CIF text or crystal strings as input, the output subset is serialized according to the original fields and directly replaces its fine-tuning data file; for isovariant graph neural network backbones that take graph representation as input, the lattice and fractional coordinates are first parsed and the graph is constructed according to material_id in the loader, and then the index is filtered according to the output subset; for diffusion and flow matching backbones that take continuous coordinates as input, the training index file is directly replaced with the output subset, and the rest of the training scripts, hyperparameters and inference code remain unchanged. Since this invention does not change the data format but only the data set, the same selected results can be reused for horizontal comparison experiments between different backbones.

[0101] Example 7: Extending to other datasets; For datasets with formation energy, energy above hull, stability classification, or other sortable quality signals, the quality signals can be normalized to E. hull Equivalent quantities; for datasets without spacegroup_number but with spacegroup symbols, spacegroup number mapping can be completed through unique symbol caching; for datasets without standard CIF but with lattice matrix and atomic coordinates, lattice ratio and periodic atom pair distance histograms can be directly calculated; for datasets without energy labels but with experimental syntheticability, relaxation energy, or other mass labels, these labels can replace E in the mass weights. hull .

[0102] In one specific implementation, the normalization method is as follows: Let the original quality signal be q(x), and the smaller the value, the better. First, map it to the [0,1] interval according to the quantile of the entire corpus, then multiply it by a scaling factor with the same dimensions as the recommendation energy threshold T to obtain E. hull An equivalent quantity is obtained, thus reusing all subsequent steps of the invention without modifying the algorithm itself. For cases where a larger quality signal is better (e.g., synthetic probability scores), the inverse or complement of q(x) is taken first before performing the above normalization.

[0103] Example 2: This invention also provides a data selection system for supervised fine-tuning of crystal material generation models, which can implement the method steps as described in Embodiment 1, specifically including: The data acquisition module is used to acquire training data for crystal materials. Each sample in the training data includes at least a sample identifier, energy on the convex hull, elemental composition, and crystal structure representation. A metadata construction module is used to construct crystal metadata from the training corpus, wherein the crystal metadata includes element number hierarchy, atom number hierarchy, crystal system hierarchy and structure descriptor; A mass weight calculation module is used to calculate the mass weight based on the energy on the convex hull. The hierarchical quota allocation module is used to construct multidimensional hierarchies based on element quantity, atomic quantity, and crystal system, and to allocate hierarchical quotas. The distribution correction weight calculation module is used to calculate the element distribution correction weight of the candidate samples based on the element occurrence rate of the full training corpus. The priority formation module is used to form a selection priority in each hierarchical candidate pool based on the quality weight and element distribution correction weight. The diversity sampling module is used to perform farthest point sampling based on the structure descriptor to obtain a target data subset for supervised fine-tuning of the crystal material generation model.

[0104] The system of this invention can be deployed in a single-machine or multi-node environment. The data acquisition module is responsible for field mapping, spatial group symbol caching, and training partitioning checks; the metadata construction module is responsible for parsing composition, number of atoms, crystal system, and structure descriptors; the mass weight calculation module, the hierarchical quota allocation module, and the distribution correction weight calculation module together form a mass and distribution selector; the selection priority formation module forms a selection priority in each hierarchical candidate pool based on the mass weight and element distribution correction weight; and the diversity sampling module outputs a subset of structural diversity.

[0105] In one specific implementation, the system is organized into a pipeline as follows: Input paths, target subset size K, various parameters (T, τ, ρ, α, γ, δ, candidate pool multiple, quality power η, histogram interval number and cutoff distance) are declared in a configuration file, along with a random seed. The pipeline sequentially executes reading, metadata construction, weight calculation, hierarchical quotas, candidate pool generation, weighted farthest point sampling, dual-pool merging, and output. At each step, intermediate results are persisted to disk as reproducible intermediate files, allowing any step to be rerun independently without repeating upstream computations. In a multi-node environment, candidate samples can be sharded by hierarchy or by pool, with each shard independently performing weighted farthest point sampling before global merging and deduplication.

[0106] Example 3: The present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in Embodiment 1 above.

[0107] Example 4: The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in Embodiment 1 above.

[0108] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0109] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for supervising and fine-tuning data selection for a crystal material generation model, characterized in that, Includes the following steps: Acquire training corpus of crystal materials, wherein each sample in the training corpus includes at least a sample identifier, energy on the convex hull, elemental composition, and crystal structure representation; Crystal metadata is constructed from the training corpus, and the crystal metadata includes element number hierarchy, atomic number hierarchy, crystal system hierarchy and structure descriptor; The mass weight is calculated based on the energy on the convex hull; Multidimensional hierarchical structures are constructed based on element quantity, atomic quantity, and crystal system, and hierarchical quotas are allocated accordingly. The element distribution correction weights for candidate samples are calculated based on the element occurrence rates of the entire training corpus. Within each stratified candidate pool, selection priorities are formed based on the quality weight and element distribution correction weight, and farthest point sampling is performed according to the structure descriptor to obtain a target data subset for supervised fine-tuning of the crystal material generation model.

2. The method according to claim 1, characterized in that, The acquisition of crystal material training corpus includes: reading only the training partition and not the test partition as supervised fine-tuning data; deduplicating or marking duplicate samples, samples with missing key fields, and samples with failed structure parsing; and writing the original fields and filtered metadata together into the output file to form a traceable data record.

3. The method according to claim 1, characterized in that, The quality weight is: w Q (x)=exp(-E hull (x) / τ) Where x is the crystal sample, E hull (x) represents the energy on the convex hull, τ represents the mass temperature parameter; and E... hull Samples not exceeding a preset threshold T are defined as the low-energy core candidate pool, and E is... hull Samples greater than T are used as the backfill candidate pool.

4. The method according to claim 3, characterized in that, Before performing farthest point sampling based on the structural descriptor, the method further includes: explicitly dividing the candidate sample dataset into an on-hull pool and a near-hull pool, wherein the on-hull pool satisfies |E hull |≤ε, the near-hull pool satisfies ε <E hull ≤T; ε represents the set on-hull threshold; T represents the set near-hull threshold; Based on the target data subset size K and the preset on-hull ratio ρ, determine the target quotas k0 for the two pools respectively. ρ) and k1=K-k0, perform hierarchical quota, element distribution correction and farthest point sampling in the two pools respectively and then merge the output; round(·) is the rounding function.

5. The method according to claim 1, characterized in that, The multidimensional layering is determined by the ternary set s = (Nelement-bin, Natom-bin, crystal-system), where Nelement-bin includes at least the intervals of elements 1, 2, 3, 4, and 5+; Natom-bin includes at least the intervals of atoms 1-4, 5-8, 9-12, 13-16, 17-20, and 21+; and crystal-system includes at least triclinic, monoclinic, orthorhombic, tetragonal, trigonal, hexagonal, and cubic crystal systems. Each layer's quota satisfies q. s With N s α Proportional, N s q is the stratified sample size, α is the quota exponent and is less than 1; s This represents the sampling ratio for each layer.

6. The method according to claim 1, characterized in that, The element distribution correction weight is: Where x is the crystal sample, Z(x) represents the distribution of the entire crystal sample, Z represents sampling from the sample, and p full (Z) represents the occurrence rate of element Z in the entire training corpus, p core (Z) represents the occurrence rate of element Z in the low-energy core candidate pool, γ is the correction intensity, δ is the smoothing constant, and clip represents the cutoff function. , = (0.5, 0.5, 2.0).

7. The method according to claim 1, characterized in that, The selection priority is w(x) = w Q (x)w E (x); w Q (x) represents the quality weight; w E (x) represents the element distribution correction weight; In each stratum, a candidate pool is selected according to the selection priority. The size of the candidate pool is a preset multiple of the quota of the corresponding stratum, and the farthest point sampling is performed from the candidate pool.

8. The method according to claim 1, characterized in that, The structural descriptor includes at least two of the following: Dimensionless lattice ratio obtained by normalizing the geometric mean of lattice edge lengths; The normalized lattice angle is obtained by dividing the lattice angle by 180 degrees. The characteristic of the unit atomic volume obtained by dividing the unit cell volume by the number of atoms and then taking the logarithm. Truncation of atomic number characteristic; And the minimum mirror atom pair distance histogram calculated based on periodic boundary conditions and 27 periodic mirrors.

9. The method according to claim 1, characterized in that, The sampling of the farthest point adopts a greedy iteration: First, select the candidate sample with the highest priority; Then, for each unselected sample, the minimum structural descriptor distance between it and the set of selected samples is calculated, and the minimum distance is multiplied by a preset power of the normalized selection priority to obtain the sampling score. The sample with the highest sampling score is selected until the stratified quota is reached.

10. The method according to claim 4, characterized in that, When the sample size of an on-hull pool or a near-hull pool is less than its target quota, use all the samples in that pool and transfer the unmet quota to another pool. When the combined size of the two pools is still less than the target data subset size, samples from outside the low-energy core candidate pool are backfilled according to selection priority, and the actual proportion, actual quantity, and backfill quantity are recorded in the data summary.

11. The method according to claim 1, characterized in that, It also includes candidate parameter control steps: By changing at least one of the following factors: mass temperature τ, energy threshold T, on-hull ratio ρ, quota index α, element correction intensity γ, structure candidate pool multiple, and farthest point sampling quality power, multiple candidate supervised fine-tuning data subsets are generated. Calculate the training distribution description vector u for each candidate subset, and obtain the LeMat-GenBench index y under the same generative model, the same number of fine-tuning steps, and the same evaluation metric, fitting y≈β0+β T u, select the target data subset based on the fitting results.

12. The method according to claim 11, characterized in that, The training distribution description vector u includes at least three of the following: on-hull ratio, average energy on the convex hull, low energy ratio, ternary structure ratio, 9-12 atom structure ratio, cubic crystal system ratio, and element occurrence rate JS divergence. The LeMat-GenBench metrics include at least three of the following: Valid, Unique, Novel, Stable, Metastable, SUN, MSUN, and MSUN+SUN. The fitting is a single-factor or multi-factor linear regression, and the correlation coefficient is used as the basis for parameter selection rather than as proof of causality.

13. The method according to claim 1, characterized in that, The output target data subset includes the original crystal structure field, filtering mode, quality weight, element correction weight, selection priority, hierarchical affiliation, structure descriptor configuration, random seed, and a statistical summary of the distribution of the full corpus and the target subset.

14. A data selection system for supervised fine-tuning of a crystal material generation model, characterized in that, include: The data acquisition module is used to acquire training data for crystal materials. Each sample in the training data includes at least a sample identifier, energy on the convex hull, elemental composition, and crystal structure representation. A metadata construction module is used to construct crystal metadata from the training corpus, wherein the crystal metadata includes element number hierarchy, atom number hierarchy, crystal system hierarchy and structure descriptor; A mass weight calculation module is used to calculate the mass weight based on the energy on the convex hull. The hierarchical quota allocation module is used to construct multidimensional hierarchies based on element quantity, atomic quantity, and crystal system, and to allocate hierarchical quotas. The distribution correction weight calculation module is used to calculate the element distribution correction weight of the candidate samples based on the element occurrence rate of the full training corpus. The priority formation module is used to form a selection priority in each hierarchical candidate pool based on the quality weight and element distribution correction weight. The diversity sampling module is used to perform farthest point sampling based on the structure descriptor to obtain a target data subset for supervised fine-tuning of the crystal material generation model.

15. An electronic device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1 to 14.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 14.