文献类型：期刊文献

英文题名：A generalizable framework for molten salt liquidus temperature prediction

作者：Zhang, Zian[1];Cheng, Hua[1];Zhang, Yanmei[3];Sun, Ze[2];Qiu, Qiuling[3];Luo, Yuxin[3];Wang, Huifeng[1];Yu, Hongbo[2]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, 130 Meilong Rd, Shanghai 200237, Peoples R China;[2]East China Univ Sci & Technol, Sch Resources & Environm Engn, 130 Meilong Rd, Shanghai 200237, Peoples R China;[3]Shanghai Elect Grp Co Ltd, 960 Zhongxing Rd, Shanghai 200070, Peoples R China

年份：2026

卷号：179

外文期刊名：JOURNAL OF ENERGY STORAGE

收录：;EI(收录号：20263021181444);Scopus(收录号：2-s2.0-105045573172);WOS:【SCI-EXPANDED(收录号:WOS:001836650700001)】；

基金：This work was supported by the Shanghai Explorer Program (Grant No. 24TS1412100) from Shanghai Electric Group Co., Ltd.

语种：英文

外文关键词：Multicomponent molten salts; Deep learning; Tokenization; Differentiable physical decoder

摘要：Deep learning has shown great potential in predicting chemical and material properties, but there are still bottlenecks in predicting the liquidus temperatures of multicomponent molten salt systems. The model lacks finegrained characterization of the complex interactions between individual components and ions, and has insufficient generalization ability for high-dimensional sampled data, which limits the R & D efficiency of new molten salts. This paper proposes a Tokenization-Encoder-Decoder prediction framework (TED) aiming to capture component-level and interaction-level features of multicomponent molten salt systems. This method constructs salt tokens based on the physical and chemical properties of single salts to characterize the individual characteristics of components; defines pair tokens by interaction properties to explicitly model the interactions between component ions. In the encoding stage, we designed a FAST (fraction-aware set transformer) dual-stream encoder to process salt tokens and pair tokens respectively. The composition-aware attention mechanism embeds the molar ratio as prior knowledge, and finally generates a multi-granularity mixed salt representation. In the decoding stage, we use a differentiable physical decoder. Different from traditional PINNs that use physical loss as a soft constraint, we embed the Redlich-Kister-Muggianu activity model and the liquid-solid phase equilibrium equation as differentiable layers into the inference network to achieve end-to-end connection between the thermodynamic parameter network and the phase equilibrium solver. Experiments show that on the dataset composed of nitrates and chlorides, the model exhibits competitive transfer ability from binary to multicomponent systems. The mean absolute percentage errors of ternary salt generalization under the few-shot (3-shot) and zero-shot settings are only 4.90% and 6.84% respectively, and the average error of quaternary salt generalization compared with experimental data is 13.4 degrees C.