-
近年来, 机器学习在材料科学中的应用显著加快了新材料的发现, 特别是在结合第一性原理计算等传统方法后, 能够高效筛选已有数据库中的潜在高性能材料. 然而, 此类方法大多局限于已有化学空间, 难以实现对全新材料结构的主动设计. 为突破这一瓶颈, 基于生成模型的材料逆向设计方法逐渐兴起, 成为探索未知结构与性质空间的重要手段. 尽管当前生成模型在晶体结构生成方面取得了初步进展, 但如何实现目标性质导向的材料生成仍面临显著挑战. 本文首先介绍了近年来在材料生成领域中具有代表性的生成模型, 包括CDVAE, MatGAN以及MatterGen, 分析其在结构生成上的基本能力与局限. 随后重点探讨如何将目标性质有效引入生成模型, 实现性质导向的结构生成, 具体包括基于目标性质向量的Con-CDVAE、融合结构约束与引导机制的SCIGEN、通过适配器实现性质调控的微调版MatterGen以及结合隐空间搜索优化的CDVAE隐变量优化策略. 最后总结当前性质导向生成机制面临的挑战, 并展望其未来的发展方向. 本文旨在为研究者深入理解和拓展性质驱动的材料生成方法提供系统性参考和启发.In recent years, the application of machine learning in materials science has significantly accelerated the discovery of new materials. In particular, when combined with traditional methods such as first-principles calculations, machine learning models have proven effective in screening potential high-performance materials from existing databases. However, these methods are largely limited by the known chemical spaces, making it difficult to achieve the active design of novel material structures. To overcome this limitation, generative models have become a promising tool for inverse material design, providing new avenues for exploring unknown structures and property spaces. Although existing generative models have achieved initial progress in crystal structure generation, achieving property-guided material generation remains a significant challenge. In this review paper, we first introduce the representative generative models recently applied to materials generation, including CDVAE, MatGAN, and MatterGen, and analyzes their basic abilities and limitations in structural generation. We then focus on strategies for incorporating target properties into generative models to generate the property-guided structure. Specifically, we discuss four representative methods: Con-CDVAE based on target property vectors, SCIGEN with integrated structural constraints and guidance mechanisms, a fine-tuned version of MatterGen leveraging adapter-based property control, and a CDVAE latent space optimization strategy guided by property objectives. Finally, we summarize the key challenges faced by property-guided generative models and provide an outlook on future research directions. This review aims to offer researchers a systematic reference and inspiration for advancing property-driven generative approaches in material design and provides researchers with a systematic reference and insight into the advancement of property-driven generative methods for materials design.
-
Keywords:
- machine learning /
- generative models /
- inverse design /
- property-guided
-
图 1 数据驱动材料设计的两种范式: 正向设计与逆向设计. 图中黑色区域表示材料化学空间中已探索的区域, 彩色点所在的区域表示未探索区域, 黄色点所在的区域表示潜在的高性能材料区域; 正向设计流程(上)从已知材料结构出发, 经第一性原理与机器学习预测逐步筛选候选材料; 逆向设计流程(下)从目标性质出发, 通过生成模型直接生成候选结构
Fig. 1. Two paradigms of data-driven materials design: forward design and inverse design.The black region in the figure represents the explored part of the materials chemical space, the region with colored dots indicates the unexplored space, and the region with yellow dots denotes the potential high-performance materials. The forward design process (top) starts from known material structures and gradually screens candidate materials via first-principles calculations and machine learning predictions; the inverse design process (bottom) starts from target properties and directly generates candidate structures through generative models.
图 2 (a) CDVAE模型的整体框架流程[30]. 模型的输入为晶体的图结构表示, 通过周期性图神经网络(PGNN)进行编码, 结合性质聚合预测器预测结构聚合信息, 并由得分网络[33](图中使用PGNN进行解码的部分)结合朗之万动力学(Langevin dynamics)[34]生成稳定的三维晶体结构; (b) 不同的生成模型在不同的数据集上经过随机采样生成的晶体结构图
Fig. 2. (a) Framework of the CDVAE[30]. The model takes the crystal graph representation as input, encodes it using a periodic graph neural network (PGNN) and predicts aggregated structural attributes through a property aggregation predictor. A score-based network[33] (implemented using PGNN in the decoder) combined with Langevin dynamics[34] is then used to generate stable three-dimensional crystal structures. (b) Crystal structures randomly generated by different generative models on different datasets.
图 3 (a) MatGAN的生成对抗网络架构示意图[31], 模型通过生成器从隐空间中采样隐向量, 生成新的材料组成样本, 判别器判断样本是否来源于真实数据分布, 通过对抗训练优化生成效果; (b) 使用t-SNE[50]方法对材料成分分布可视化降维图, 图中蓝色点为MatGAN生成的样本, 绿色点为训练集样本, 红色点为保留未参与训练的leave-out样本
Fig. 3. (a) Schematic diagram of the generative adversarial network architecture of MatGAN[31]. The model samples latent vectors from the latent space via the generator to produce new material compositions, while the discriminator determines whether the samples come from the real data distribution, thereby optimizing the generator through adversarial training. (b) t-SNE[50] visualization of the material composition distribution after dimensionality reduction. Blue points represent samples generated by MatGAN, green points denote training set samples, and red points indicate leave-out samples not used during training.
图 4 (a) MatterGen的扩散建模流程示意图[25], 正向过程(forward process)通过逐步加噪将稳定结构材料(A0, X0, L0)转化为完全随机结构(AT, XT, LT), 反向过程(reverse process)则利用得分网络(score network)[53]引导去噪, 逐步重建出稳定晶体结构; (b) 等变性得分网络(equivariant score network)结构示意图; (c) 不同生成模型在Materials Project[9]数据集上新颖稳定生成材料的比例对比图; (d) MatterGen生成材料在不同采样规模下的新颖性与唯一性统计图; (e) MatterGen生成的部分代表性新型晶体结构图
Fig. 4. (a) Schematic diagram of the diffusion modeling process of MatterGen[25]The forward process gradually corrupts a stable material structure (A0, X0, L0) into a completely random structure (At, Xt, Lt) by adding noise step-by-step, while the reverse process leverages a score network[53] to guide denoising and progressively reconstruct a stable crystal structure. (b) Schematic diagram of the equivariant score network structre. (c) Comparison of the proportion of novel and stable structures generated by different models on the Materials Project[9] dataset. (d) Statistical analysis of the novelty and uniqueness of materials generated by MatterGen at different sampling scales. (e) Representative examples of novel crystal structures generated by MatterGen.
图 5 (a) Con-CDVAE的训练与生成流程图[22]. 训练阶段包括结构与目标性质的联合编码(encoder)、性质预测器(predictor)和条件解码(decoder), 同时训练先验(prior)模块[14,33]用于实现从目标性质生成隐变量z. 生成阶段中, 输入目标性质后通过先验模块生成隐变量z, 并与通过propemb模块得到的性质嵌入向量进行拼接后得到含有目标性质信息的隐变量zcond, 将zcond送入解码器后生成晶体结构. (b) MP-20[9]训练数据集中材料的带隙分布与形成能分布图. (c) Con-CDVAE在设定的目标性质下的生成材料的性质分布图. (d) Con-CDVAE生成的部分代表性新型晶体结构图
Fig. 5. (a) Training and generation workflow of Con-CDVAE[22]. The training phase includes joint encoding of crystal structures and target properties via the encoder, a property predictor, and conditional decoding through the decoder. Simultaneously, a prior module [14,33] is trained to enable the generation of latent variable z from target property inputs. In the generation phase, a target property is first provided, from which a latent variable z is generated using the prior module. This latent variable is then concatenated with a property embedding vector obtained from the propemb module to derive the latent variable zcond conditioned on the target property, which is passed to the decoder to generate crystal structures. (b) Distribution of band gap and formation energy in the MP-20 training dataset[9]. (c) Property distributions of materials generated by Con-CDVAE under different target property settings. (d) Representative novel crystal structures generated by Con-CDVAE.
图 6 (a) SCIGEN模型的材料结构生成过程示意图[23], 该模型以几何规则为引导, 通过多步扩散与去噪过程, 结合掩码机制, 将具有几何约束的结构与待生成结构融合, 引导生成结构逐步逼近目标结构; (b) SCIGEN的结构生成中几何约束与非约束结构的合并与掩码机制过程图; (c) Lieb晶格的二维示意图; (d) SCIGEN在Lieb几何约束下生成的代表性结构Ce3NaTb4; (e) Ce3NaTb4的能带结构图
Fig. 6. (a)Schematic illustration of the material structure generation process in the SCIGEN model[23]. The model is guided by geometric rules and utilizes a multi-step diffusion and denoising process, combined with a masking mechanism, to integrate geometrically constrained structures with unconstrained ones, thereby gradually guiding the generation toward the target structure. (b) Diagram showing the merging of constrained and unconstrained structures and the masking mechanism used in SCIGEN during structure generation. (c) A 2D schematic of the Lieb lattice. (d) A representative structure, Ce3NaTb4, generated by SCIGEN under Lieb lattice constraints. (e) Band structure of the generated Ce3NaTb4 material.
图 7 (a) MatterGen微调阶段的模型结构示意图[25], 目标性质c被送入适配器模块(adapter module), 并与等变性得分网络(equivariant score network)的输出结合, 引导结构从随机状态逐步恢复至符合目标性质的晶体结构; (b) MatterGen在不同目标性质条件控制下生成满足不同目标性质的晶体结构; (c)—(e) MatterGen在目标性质分别为磁密度(c)、带隙(d)、体积模量(e)的条件下生成材料与对应微调数据集的材料的性质分布图; (f) 实验上成功合成的MatterGen以体积模量200 GPa为目标性质生成的TaCr2O6结构图; (g) MatterGen在目标性质为不同体积模量条件下生成材料的性质的DFT验证结果图
Fig. 7. (a) Schematic illustration of the MatterGen model during the fine-tuning stage[25]. The target property condition c is fed into the adapter module and combined with the output of the equivariant score network to guide the generation process from random structures toward crystal structures satisfying the target property. (b) MatterGen-generated crystal structures under different target property constraints, demonstrating the model’s ability to satisfy various design conditions. (c)–(e) Property distribution comparisons between MatterGen-generated materials and the fine-tuning dataset under target conditions of magnetic density (c), bandgap (d), and bulk modules (e). (f) Structure of TaCr2O6, successfully synthesized experimentally, generated by MatterGen with a target bulk modulus of 200 GPa. (g) DFT-validated property results of structures generated by MatterGen under different target bulk modulus values, illustrating the model’s predictive accuracy and target controllability.
图 8 (a) MAGECS(materials generation with efficient global chemical space search)框架结构图[26], 该框架包括生成模型CDVAE、监督图神经网络GNN和鸟群算法BSA, 通过在隐空间中搜索最优隐变量以生成满足目标性质的材料结构; (b) MAGECS, CDVAE生成结构和原始数据库结构在目标性质区间$ \left| {\Delta {E_{{\text{CO}}}} + {\text{ }}0.67} \right|{\text{ }} \leqslant {\text{ }}0.2{\text{ eV}} $ 的分布对比图; (c) 实验上成功合成的MAGECS生成的5种具有目标性质的结构; (d) CuAl和Pd5Sn2的电催化CO2还原反应性能曲线
Fig. 8. (a) Framework diagram of MAGECS (materials generation with efficient global chemical space search)[26]. The framework integrates a generative model (CDVAE), a supervised graph neural network (GNN), and a bird swarm algorithm (BSA) to search for optimal latent vectors in the latent space and generate materials that meet target properties. (b) Comparison of the distribution of structures from MAGECS, CDVAE, and the original database within the target property range $ \left| {\Delta {E_{{\text{CO}}}} + {\text{ }}0.67} \right|{\text{ }} \leqslant {\text{ }}0.2{\text{ eV}} $. (c) Five target-property-aligned structures generated by MAGECS and successfully synthesized in experiments. (d) Electrocatalytic CO2 reduction performance curves of CuAl and Pd5Sn2.
图 9 三种将目标性质导向机制引入生成模型的方法
Fig. 9. Three methods for incorporating property-guidance mechanisms into generative models.
表 1 不同的生成模型在不同的数据集上生成的结构在多个评估指标下的性能对比[30]
Table 1. Comparison of generative model performance across datasets and evaluation metrics[30].
Method Data Validity/% COV/% Property statistics Struc. Comp. R. P. ρ E # elem. FTCP Perov-5 0.24 54.24 0.10 0.00 10.27 156.0 0.6297 Carbon-24 0.08 — 0.00 0.00 5.206 19.05 — MP-20 1.55 48.37 4.72 0.09 23.71 160.9 0.7363 Cond-DFC-VAE Perov-5 73.60 82.95 73.92 10.13 2.268 4.111 0.8373 G-SchNet Perov-5 99.92 98.79 0.18 0.23 1.625 4.746 0.03684 Carbon-24 99.94 — 0.00 0.00 0.9427 1.320 — MP-20 99.65 75.96 38.33 99.57 3.034 42.09 0.6411 P-G-SchNet Perov-5 79.63 99.13 0.37 0.25 0.2755 1.388 0.4552 Carbon-24 48.39 — 0.00 0.00 1.533 134.7 — MP-20 77.51 76.40 41.93 99.74 4.04 2.448 0.6234 CDVAE Perov-5 100.0 98.59 99.45 98.46 0.1258 0.0264 0.0628 Carbon-24 100.0 — 99.80 83.08 0.1407 0.2850 — MP-20 100.0 86.70 99.15 99.49 0.6875 0.2778 1.432 
下载: 导出CSV
表 2 MatGAN在OQMD[8], Materials Project[9]与ICSD[7]三个数据库上的结构复现与新颖性统计[31]
Table 2. Statistical results of structural recovery and novelty of MatGAN on the OQMD[8], Materials Project[9], and ICSD[7] databases[31].
GAN-OQMD GAN-MP GAN-ICSD Training sample # 251368 57530 25323 Leave out sample # 27929 6392 2813 Generated sample # 2000000 2000000 2000000 Recovery of training
samples/%60.26 47.36 59.54 Recovery of leave out/% 60.43 48.82 60.13 New samples 1831648 1969633 1983231
下载: 导出CSV
表 3 三类性质导向方法在不同性质上的引导效果
Table 3. Effectiveness of three property-targeted methods across various properties.
下载: 导出CSV
-
[1] Jiang X, Xue D Z, Bai Y, Wang W Y, Liu J J, Yang M L, Su Y J 2025 Rev. Mater. Res. 1 100010
[2] Wu M F, Zhang S Y, Ren J 2025 APL Mater. 13 020601
Google Scholar
[3] Merchant A, Batzner S, Schoenholz S S, Aykol M, Cheon G, Cubuk E D 2023 Nature 624 80
Google Scholar
[4] Butler K T, Davies D W, Cartwright H, Isayev O, Walsh A 2018 Nature 559 547
Google Scholar
[5] Zhong M, Tran K, Min Y M, Wang C H, Wang Z Y, Dinh C T, De Luna P, Yu Z Q, Rasouli A S, Brodersen P, Sun S, Voznyy O, Tan C S, Askerka M, Che F L, Liu M, Seifitokaldani A, Pang Y J, Lo S C, Ip A, Ulissi Z, Sargent E H 2020 Nature 581 178
Google Scholar
[6] Rao Z Y, Tung P Y, Xie R W, Wei Y, Zhang H B, Ferrari A, Klaver T P C, Körmann F, Sukumar P T, Kwiatkowski da Silva A, Chen Y, Li Z M, Ponge D, Neugebauer J, Gutfleisch O, Bauer S, Raabe D 2022 Science 378 78
Google Scholar
[7] Hellenbrandt M 2004 Crystallogr. Rev. 10 17
Google Scholar
[8] Kirklin S, Saal J E, Meredig B, Thompson A, Doak J W, Aykol M, Rühl S, Wolverton C 2015 npj Comput. Mater. 1 1
[9] Jain A, Ong S P, Hautier G, Chen W, Richards W D, Dacek S, Cholia S, Gunter D, Skinner D, Ceder G, Persson K A 2013 APL Mater. 1 011002
Google Scholar
[10] Haastrup S, Strange M, Pandey M, Deilmann T, Schmidt P S, Hinsche N F, Gjerding M N, Torelli D, Larsen P M, Riis-Jensen A C, Gath J, Jacobsen K W, Jørgen Mortensen J, Olsen T, Thygesen K S 2018 2D Mater. 5 042002
Google Scholar
[11] Lu S H, Zhou Q H, Chen X Y, Song Z L, Wang J L 2022 Natl. Sci. Rev. 9 nwac111
Google Scholar
[12] Kingma D P, Welling M 2019 Found. Trends Mach. Learn. 12 307
Google Scholar
[13] Goodfellow I J, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, Courville A, Bengio Y 2014 arXiv: 1406.2661v1 [stat. ML]
[14] Ho J, Jain A, Abbeel P 2020 arXiv: 2006.11239v2 [cs. LG]
[15] Stimper V, Liu D, Campbell A, Berenz V, Ryll L, Schölkopf B, Hernández-Lobato J M 2023 J. Open Source Softw. 8 5361
Google Scholar
[16] Hoogeboom E, Gritsenko A A, Bastings J, Poole B, Berg R van den, Salimans T 2022 arXiv: 2110.02037v2 [cs. LG]
[17] Hong T, Chen T K, Jin D L, Zhu Y, Gao H, Zhao K, Zhang T Y, Ren W, Cao G X 2025 npj Quantum Mater. 10 12
Google Scholar
[18] Jin L Z, Du Z J, Shu L, Cen Y, Xu Y F, Mei Y F, Zhang H 2025 Nat. Commun. 16 1210
Google Scholar
[19] Sanchez-Lengeling B, Aspuru-Guzik A 2018 Science 361 360
Google Scholar
[20] Zhang K, Chen T, Abbas Y, Jan S U, Zhou Z H, Chu S Q, Xie G C, Ullah S, Akram M Z, Zhang J, Xuan Y M, Gong J R 2021 Matter 4 1054
Google Scholar
[21] Manzoor A, Zhang Y, Aidhy D S 2021 Comput. Mater. Sci 198 110669
Google Scholar
[22] Ye C Y, Weng H M, Wu Q S 2024 Comput. Mater. Today 1 100003
Google Scholar
[23] Okabe R, Cheng M, Chotrattanapituk A, Hung N T, Fu X, Han B, Wang Y, Xie W, Cava R J, Jaakkola T S, Cheng Y, Li M 2024 arXiv: 2407.04557v1 [cond-mat. mtrl-sci]
[24] Ye C, Wang Y, Xie X, Zhu T, Liu J, He Y, Zhang L, Zhang J, Fang Z, Wang L, Liu Z, Weng H, Wu Q 2025 arXiv: 2505.00076v1 [cond-mat. mtrl-sci]
[25] Zeni C, Pinsler R, Zügner D, Fowler A, Horton M, Fu X, Wang Z, Shysheya A, Crabbé J, Ueda S, Sordillo R, Sun L, Smith J, Nguyen B, Schulz H, Lewis S, Huang C W, Lu Z, Zhou Y, Yang H, Hao H, Li J, Yang C, Li W, Tomioka R, Xie T 2025 Nature 639 624
Google Scholar
[26] Song Z L, Fan L F, Lu S H, Ling C Y, Zhou Q H, Wang J L 2025 Nat. Commun. 16 1053
Google Scholar
[27] Chen X Y, Lu S H, Chen Q, Zhou Q H, Wang J L 2024 Nat. Commun. 15 5391
Google Scholar
[28] Choudhary K 2024 arXiv: 2405.03680v2 [cond-mat. mtrl-sci]
[29] Lu S H, Zhou Q H, Guo Y L, Wang J L 2022 Chem 8 769
Google Scholar
[30] Xie T, Fu X, Ganea O E, Barzilay R, Jaakkola T 2021 arXiv: 2110.06197v3 [cs. LG]
[31] Dan Y B, Zhao Y, Li X, Li S B, Hu M, Hu J J 2020 npj Comput. Mater. 6 84
Google Scholar
[32] Kingma D P, Welling M 2022 arXiv: 1312.6114v11 [stat. ML]
[33] Song Y, Ermon S 2019 Proceedings of the 32nd Annual Conference on Neural Information Processing Systems Vancouver, Canada, December 8-14, 2019 p11895
[34] Lemons D S, Gythiel A 1997 Am. J. Phys. 65 1079
Google Scholar
[35] Chen S, Ge C, Zhang S, Sun P, Luo P 2025 arXiv: 2504.07963v1 [cs. CV]
[36] Shi C, Xu M, Zhu Z, Zhang W, Zhang M, Tang J 2020 arXiv: 2001.09382v2 [cs. LG]
[37] Xu M, Yu L, Song Y, Shi C, Ermon S, Tang J 2022 arXiv: 2203.02923v1 [cs. LG]
[38] Castelli I E, Landis D D, Thygesen K S, Dahl S, Chorkendorff I, Jaramillo T F, Jacobsen K W 2012 Energy Environ. Sci. 5 9034
Google Scholar
[39] Pickard C J https://archive.materialscloud.org/record/2020.0026/v1 [2024-04-20]
[40] Ren Z K, Tian S I P, Noh J, Oviedo F, Xing G Z, Li J L, Liang Q H, Zhu R M, Aberle A G, Sun S J, Wang X N, Liu Y, Li Q X, Jayavelu S, Hippalgaonkar K, Jung Y, Buonassisi T 2022 Matter 5 314
Google Scholar
[41] Court C J, Yildirim B, Jain A, Cole J M 2020 J. Chem. Inf. Model. 60 4518
Google Scholar
[42] Gebauer N W A, Gastegger M, Schütt K T 2020 arXiv: 1906.00957v3 [stat. ML]
[43] Davies D W, Butler K T, Jackson A J, Skelton J M, Morita K, Walsh A 2019 J. Open Source Softw. 4 1361
Google Scholar
[44] Xu M, Luo S, Bengio Y, Peng J, Tang J 2021 arXiv: 2102.10240v3 [cs. LG]
[45] Ganea O E, Pattanaik L, Coley C W, Barzilay R, Jensen K F, Green W H, Jaakkola T S 2021 arXiv: 2106.07802v1 [physics. chem-ph]
[46] Zhang H, Georgescu A B, Yerramilli S, Karpovich C, Apley D W, Olivetti E A, Rondinelli J M, Chen W 2025 arXiv: 2412.17283v2 [cond-mat. mtrl-sci]
[47] Murphy L R, Meek T L, Allred A L, Allen L C 2000 J. Phys. Chem. A 104 5867
Google Scholar
[48] Sessa F, Rahm M 2022 J. Phys. Chem. A 126 5472
Google Scholar
[49] Arjovsky M, Chintala S, Bottou L 2017 arXiv: 1701.07875v3 [stat. ML]
[50] Maaten L van der, Hinton G 2008 J. Mach. Learn. Res. 9 2579
[51] Green M L, Choi C L, Hattrick-Simpers J R, Joshi A M, Takeuchi I, Barron S C, Campo E, Chiang T, Empedocles S, Gregoire J M, Kusne A G, Martin J, Mehta A, Persson K, Trautt Z, Van Duren J, Zakutayev A 2017 Appl. Phys. Rev. 4 011105
Google Scholar
[52] Pickard C J, Needs R J 2011 J. Phys. Condens. Matter 23 053201
Google Scholar
[53] Song Y, Sohl-Dickstein J, Kingma D P, Kumar A, Ermon S, Poole B 2021 arXiv: 2011.13456v2 [cs. LG]
[54] Schmidt J, Wang H C, Cerqueira T F T, Botti S, Marques M A L 2022 Sci. Data 9 64
Google Scholar
[55] Jiao R, Huang W, Lin P, Han J, Chen P, Lu Y, Liu Y 2023 arXiv: 2309.04475v2 [cond-mat. mtrl-sci]
[56] Sultanov A, Crivello J C, Rebafka T, Sokolovska N 2023 J. Chem. Inf. Model. 63 6986
Google Scholar
[57] Xie T, Grossman J C 2018 Phys. Rev. Lett. 120 145301
Google Scholar
[58] Ward L, Agrawal A, Choudhary A, Wolverton C 2016 npj Comput. Mater. 2 1
Google Scholar
[59] Han X Q, Wang X D, Xu M Y, Feng Z, Yao B W, Guo P J, Gao Z F, Lu Z Y 2025 Chin. Phys. Lett. 42 027403
Google Scholar
[60] Long T, Zhang Y, Zhang H 2024 arXiv: 2409.19124v1 [cond-mat. mtrl-sci]
[61] Chen L, Zhang W, Nie Z, Li S, Pan F 2021 J. Mater. Inform. 1 N/A
[62] Chen Y, Wang X, Deng X, Liu Y, Chen X, Zhang Y, Wang L, Xiao H 2024 arXiv: 2408.07608v1 [cond-mat. mtrl-sci]
[63] New A, Pekala M, Pogue E A, Le N Q, Domenico J, Piatko C D, Stiles C D 2023 arXiv: 2309.12323v1 [cond-mat. mtrl-sci]
[64] Banik S, Dhabal D, Chan H, Manna S, Cherukara M, Molinero V, Sankaranarayanan S K R S 2023 npj Comput. Mater. 9 1
Google Scholar
[65] Cao Z, Luo X, Lv J, Wang L 2024 arXiv: 2403.15734v2 [cond-mat. mtrl-sci]
[66] Yang S, Cho K, Merchant A, Abbeel P, Schuurmans D, Mordatch I, Cubuk E D 2024 arXiv: 2311.09235v2 [cs. LG]
[67] Jiao R, Huang W, Liu Y, Zhao D, Liu Y 2024 arXiv: 2402.03992v2 [cs. LG]
[68] Tsai W F, Fang C, Yao H, Hu J 2015 New J. Phys. 17 055016
Google Scholar
[69] Ho J, Salimans T 2022 arXiv: 2207.12598v1 [cs. LG]
[70] Zhang L, Rao A, Agrawala M 2023 arXiv: 2302.05543v3 [cs. CV]
[71] Shuaibi M, Kolluru A, Das A, Grover A, Sriram A, Ulissi Z, Zitnick C L 2021 arXiv: 2106.09575v1 [cs. LG]
[72] Guo Z, Zhang C, Yu W, Herr J, Wiest O, Jiang M, Chawla N V 2021 arXiv: 2102.07916v1 [cs. LG]
[73] Ryan K, Lengyel J, Shatruk M 2018 J. Am. Chem. Soc. 140 10158
Google Scholar
[74] Chenebuah E T, Nganbe M, Tchagang A B 2024 npj Comput. Mater. 10 198
Google Scholar
[75] Fung V, Zhang J, Juarez E, Sumpter B G 2021 npj Comput. Mater. 7 1
Google Scholar
[76] Varol Altay E, Alatas B 2020 Artif. Intell. Rev. 53 1373
Google Scholar
[77] Butler K T, Choudhary K, Csanyi G, Ganose A M, Kalinin S V, Morgan D 2024 npj Comput. Mater. 10 1
Google Scholar
[78] Wu Y L, Li X Y, Guo R, Xu R Q, Ju M G, Wang J L 2025 Natl. Sci. Rev 12 nwaf081
Google Scholar
[79] Song Z L, Lu S H, Ju M G, Zhou Q H, Wang J L 2025 Nat. Commun. 16 6530
Google Scholar
[80] Seko A, Hayashi H, Tanaka I 2018 J. Chem. Phys. 148 241719
Google Scholar
下载:
计量
- 文章访问数: 639
- PDF下载量: 34
- 被引次数: 0
