Research on An Intelligent Recommendation Model for Crystal Synthesis Procedures Based on Small-Sample Data

doi:10.11871/jfdc.issn.2096-742X.2026.01.018

Abstract

Abstract:

[Background] Crystal synthesis is vital for developing new materials but challenging due to complex, variable conditions and limited experimental data. [Problem] Intelligently generating and evaluating crystal synthesis processes from limited “structure-process” datasets is urgently needed. [Methods] We constructed a feasibility evaluation model based on the structure-understanding model (CrysBert) optimized for small datasets, and a generative synthesis model utilizing the large structure-generation model (CrysGPT). Integrating both models enabled automatic process generation and screening. [Results] Trained on 162 small-sample data points, the discriminative model achieved an accuracy of 0.90, outperforming traditional methods. The generative model efficiently produced candidate processes, achieving a feasibility rate of 60.7% after discriminative screening, approaching the expert benchmark (62.3%). [Conclusions] This study demonstrates that integrating CrysBert and CrysGPT provides a promising new pathway for intelligent crystal synthesis design.

Key words: CrysBert, CrysGPT, crystal synthesis process intelligent recommendation, small-sample data, crystal synthesis process design

ZHU Dong,YANG Xiaoyu,TANG Shujie,ZHU Fengfeng,KONG Xiao,GUO Yanfeng,LI Bing,QIN Zhipeng. Research on An Intelligent Recommendation Model for Crystal Synthesis Procedures Based on Small-Sample Data[J]. Frontiers of Data and Computing, 2026, 8(1): 219-231, https://cstr.cn/32002.14.jfdc.CN10-1649/TP.2026.01.018.

Figures/Tables 10

Fig.1

Table 1

Process data display"

材料名称	生长方法	元素比例	生长温度曲线和加热后操作	单晶生长结果
Ce₃Pt₃Bi₄	干烧	Ce $∶$ Pt $∶$ Bi=3 $∶$ 3 $∶$ 4	从30 ℃以15小时升温至1,100 ℃，保持20小时，再用100小时降温至800 ℃后离心	获得极小单晶颗粒

Table 1

Fig.2

Table 2

Fig.3

Fig.4

Table 3

Examples of process data"

材料名称	生长方法	元素比例	生长温度曲线和加热后操作	单晶生长结果
Ce₃Pt₃Bi₄	干烧	Ce $∶$ Pt $∶$ Bi=3 $∶$ 3 $∶$ 4	从30 ℃以15小时升温至1,100 ℃，保持20小时…	获得极小单晶颗粒
Ca2Ge	干烧	Ca $∶$ Ge=13 $∶$ 7	从30 ℃用11小时升温到1,050 ℃，维持20小时…	未成功
$?$	$?$	$?$	$?$	$?$
PrAlGe	助熔剂	Pr $∶$ Ga $∶$ Ge $∶$ In=1 $∶$ 1 $∶$ 1 $∶$ 20	从30 ℃用20小时升温1,150 ℃，维持15小时…	生长出单晶

Table 3

Table 4

Fig.5

Fig.6

References 27

[1]	QI R, ZHU B, HAN Z, et al. High-Throughput Screening of Stable Single-Atom Catalysts in CO2 Reduction Reactions[J]. ACS Catalysis, 2022, 12(14): 8269-8278. doi: 10.1021/acscatal.2c02149
[2]	ZHANG B, ABBAS A, ROMAGNOLI J A. Monitoring crystal growth based on image texture analysis using wavelet transformation[J]. IFAC Proceedings Volumes, 2012, 45(15): 33-38.
[3]	ASKARI S, BASHARDOUST SIAHMARD A, HALLADJ R, et al. Different techniques and their effective parameters in nano SAPO-34 synthesis: A review[J]. Powder Technology, 2016, 301: 268-287. doi: 10.1016/j.powtec.2016.06.018
[4]	RAWLINGS J B, MILLER S M, WITKOWSKI W R. Model identification and control of solution crystallization processes: a review[J]. Industrial & Engineering Chemistry Research, 1993, 32(7): 1275-1296. doi: 10.1021/ie00019a002
[5]	ROSSO V W, YIN Z, ABOURAHMA H, et al. High-Throughput Crystallization Screening Technique with Transmission PXRD Analysis[J]. Organic Process Research & Development, 2023, 27(8): 1437-1444.
[6]	BAIRD S G, LIU M, SAYEED H M, et al. Data-Driven Materials Discovery and Synthesis using Machine Learning Methods[M]. Comprehensive Inorganic Chemistry III (Third Edition), 2023: 3-23.
[7]	谢建新, 宿彦京, 薛德祯, 等. 机器学习在材料研发中的应用[J]. 金属学报, 57(11): 1343-1361.
[8]	HIMANEN L, GEURTS A, FOSTER A S, et al. Data-Driven Materials Science: Status, Challenges, and Perspectives[J]. Advanced Science, 2019, 6(21): 1900808. doi: 10.1002/advs.v6.21
[9]	MERCHANT A, BATZNER S, SCHOENHOLZ S S, et al. Scaling deep learning for materials discovery[J]. Nature, 2023, 624(7990): 80-85. doi: 10.1038/s41586-023-06735-9
[10]	ZHU D, XIN Z, ZHENG S, et al. Addressing the Accuracy-Cost Trade-off in Material Property Prediction Using a Teacher-Student Strategy[J]. Journal of Chemical Theory and Computation, 2024, 20(13): 5743-5750. doi: 10.1021/acs.jctc.4c00625
[11]	ZHU D, WANG C, ZOU P, et al. Deep-Learning Aided Atomic-Scale Phase Segmentation toward Diagnosing Complex Oxide Cathodes for Lithium-Ion Batteries[J]. Nano Letters, 2023, 23(17): 8272-8279. doi: 10.1021/acs.nanolett.3c02441
[12]	SONG Z, LU S, JU M, et al. Is Large Language Model All You Need to Predict the Synthesizability and Precursors of Crystal Structures?[EB/OL]. arXiv, 2024.
[13]	MYERSON A S, ERDEMIR D, LEE A Y. Handbook of Industrial Crystallization[M]. 3rd ed. Cambridge: Cambridge University Press, 2019: 249-266.
[14]	JHA D, CHOUDHARY K, TAVAZZA F, et al. Enhancing materials property prediction by leveraging computational and experimental data using deep transfer learning[J]. Nature Communications, 2019, 10(1): 5316. doi: 10.1038/s41467-019-13297-w pmid: 31757948
[15]	FISHER O J, WATSON N J, ESCRIG J E, et al. Considerations, challenges and opportunities when developing data-driven models for process manufacturing systems[J]. Computers & Chemical Engineering, 2020, 140: 106881. doi: 10.1016/j.compchemeng.2020.106881
[16]	LUO Y, BAG S, ZAREMBA O, et al. MOF Synthesis Prediction Enabled by Automatic Data Mining and Machine Learning[J]. Angewandte Chemie International Edition, 2022, 61(19): e202200242.
[17]	GRATTAFIORI A, DUBEY A, JAUHRI A, et al. The Llama 3 Herd of Models[EB/OL]. arXiv, 2024.
[18]	LU W, LUU R K, BUEHLER M J. Fine-tuning large language models for domain adaptation: Exploration of training strategies, scaling, model merging and synergistic capabilities[EB/OL]. arXiv, 2024.
[19]	MULLIN J W. 6-Crystal growth[M]// MULLINJ W. Crystallization (Fourth Edition). Oxford: Butterworth-Heinemann, 2001: 216-288.
[20]	DEVLIN J, CHANG M W, LEE K, et al. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding[EB/OL]. arXiv, 2019.
[21]	SUTSKEVER I, VINYALS O, LE Q V. Sequence to Sequence Learning with Neural Networks[C]// Advances in Neural Information Processing Systems: Vol. 27. Curran Associates, Inc., 2014.
[22]	WEI J, BOSMA M, ZHAO V Y, et al. Finetuned Language Models Are Zero-Shot Learners[EB/OL]. arXiv, 2022.
[23]	OUYANG L, WU J, JIANG X, et al. Training language models to follow instructions with human feedback[EB/OL]. arXiv, 2022.
[24]	SAHOO P, SINGH A K, SAHA S, et al. A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications[EB/OL]. arXiv, 2024.
[25]	LI X L, LIANG P. Prefix-Tuning: Optimizing Continuous Prompts for Generation[EB/OL]. arXiv, 2021.
[26]	DETTMERS T, PAGNONI A, HOLTZMAN A, et al. QLoRA: Efficient Finetuning of Quantized LLMs[EB/OL]. arXiv, 2023.
[27]	BREIMAN L. Random Forests[J]. Machine Learning, 2001, 45(1): 5-32. doi: 10.1023/A:1010933404324

工艺种类	预处理方法	向量长度描述
生长方法	采用one-hot	5维（干烧法、助熔剂法、气相传输法、复合法、固相反应法）
反应物	采用one-hot 以及数值型	Onehot 4×25维（4种元素）数值4维（元素比例）
反应温度	采用数值型	最多7个阶段，每阶段记录2个数值（温度和时长），共14维
加热后处理	采用one-hot	4维（离心、退火、淬火以及无操作）

模型	输入	判别准确率
SynBert	晶体结构+数据预处理工艺	0.90
CrysGPT微调判别	CIF文本+工艺文本	0.88
工艺神经网络	特征化晶体结构+数据预处理工艺	0.78
随机森林	特征化晶体结构+数据预处理工艺	0.75